The High Cost of Euclidean Thinking

There is a particular kind of silence that fills a boardroom about fourteen months into a Phase 3 trial. The enrollment forecast has been revised three times. The “Site Rescue” task force has stopped asking why sites are behind and started negotiating how much the timeline must shift. In this room, everyone is an expert. The sites are world-class, the CRO is reputable, and the clinical leads are tireless.

Yet the trial is in trouble—structurally, irreversibly.

The traditional response is to double down on “linear” solutions: hire more CRAs, add more sites, or increase the recruitment budget. This approach assumes that a clinical trial is a Euclidean machine—a system where inputs and outputs share a proportional, straight-line relationship. In the Euclidean worldview, if you are 10% behind, you simply need 10% more effort to catch up.

But clinical trials are not machines. They are non-linear dynamical systems. They are iterative feedback loops where the output of Month 1 (enrollment, safety signals, site engagement) becomes the input for Month 2. In mathematics, when you iterate a non-linear function, you don’t get a straight line; you get a fractal.

If your trial is too slow, it isn’t because you lack “effort.” It’s because your operational parameters have drifted into a region of mathematical chaos. To understand how to navigate this, we must look at the geometry of fate.

I. Gaston Julia and the Ghost in the Execution

The History of the Leather Mask

In 1918, amidst the ruins of post-WWI France, a mathematician named Gaston Julia published a 199-page masterpiece on the iteration of rational functions. Julia was a war hero who had lost his nose in the trenches; he wore a leather mask for the rest of his life. Working without computers, Julia (and his contemporary Pierre Fatou) visualized a world where simple equations could produce boundaries of infinite complexity.

For a fixed protocol configuration (a constant c), the Julia set is the boundary of the set of starting points z0 that do not escape to infinity under the iteration:

The Clinical Parallel: Sensitivity to Initial Conditions

In this equation, z0represents your starting conditions. This is the specific sequence of your first ten sites, your initial patient cohort, and your first regulatory green light.

1. Connected Julia Sets (Robustness): When your protocol (c) is stable, the Julia set is a “connected” shape. This means the outcome is robust. Whether Site A or Site B activates first, the system stays within the “basin” of success.

2. Cantor Dust (The Fragmented Trial): If the protocol is slightly off, the Julia set shatters into “dust.” Success becomes hyper-sensitive to z_0. This is the Butterfly Effect of operations. If your high-enrolling site opens on a Tuesday, you succeed. If they open on a Wednesday, the whole program “diverges to infinity” (fails).

Most “failed” trials were actually “Cantor Dust” executions. The protocol was so fragile that only a perfect, impossible sequence of events could have led to a positive readout. We blame “bad luck,” but the math tells us the failure was baked into the starting geometry. Here is a graphic from a recent blog post

II. Pierre Fatou and the Gravity of Failure

The Discovery of Attractors

While Julia looked at the boundary (chaos), his contemporary Pierre Fatou—working in isolation at the Paris Observatory—explored the interior. Fatou was arguably the first to understand Basins of Attraction.

The Math: The Fatou Set (F(f))

The Fatou set is the region where behavior is stable. Within it lie attractors. An “attractor” is a state toward which the system naturally evolves. If a point z is in a basin B(L), then:

The Clinical Parallel: The Three Basins of Fate

A clinical trial is essentially a point moving through a high-dimensional space, pulled by three primary “Fatou Basins”:

1. The Success Basin: Operational and biological variables converge toward a clean, statistically significant readout.

2. The Futility Basin: The noise of the data eventually overwhelms the signal, regardless of how many patients you add.

3. The Toxicity Basin: The system “escapes” to a state of unacceptable risk.

The Insight: Most trials fail because they are launched in the “Julia Set”—the razor-thin boundary between basins. In the Julia set, the trial is unstable. A single outlier patient can knock the entire program out of the Success Basin and into the Futility Basin. Fatou’s theorem in mathematics teaches us that stability is a property of the initial coordinates, not just of the effort applied during the journey.

III. The Mandelbrot Set: The Map of the Possible

The History: IBM and the Thumbprint of God

Sixty years after Julia and Fatou, Benoit Mandelbrot used early computer graphics to plot all the possible Julia sets on a single graph. He discovered the most complex object in mathematics: the Mandelbrot Set.

The Math: Mapping Parameter c

The Mandelbrot set is the set of all protocol configurations (c) for which the Julia set is connected. It is defined by iterating from z_0 = 0:

The Clinical Parallel: Protocol Topology

Every point c on the map represents a different combination of inclusion criteria, primary endpoints, and dosing regimens.

• The Black Heart: If your protocol (c) is in the middle of the set, you are safe. You have built a “connected” path to success.

• The Seahorse Valleys: If you are on the “fringes”—the “Seahorse Valleys” of the Mandelbrot set—your trial is technically viable but infinitely complex. A tiny change in the competitive landscape or a minor regulatory tweak moves you out of the set, and your trial “diverges.”

The failure of modern drug development is that we spend all our time managing z_n (the execution) and zero time calculating where our c (the protocol) sits on the map of stability.

IV. Fractal Time and the Coastline Paradox

The History: Richardson’s Ruler

In the 1950s, Lewis Fry Richardson noticed that the length of the border between Spain and Portugal changed depending on who was measuring it. If you use a 1-km ruler, the border is short. If you use a 1-meter ruler, the border “grows” because you are now measuring every rock and inlet.

The Math: The Fractal Dimension (D)

The measured length L follows a power law:

As the ruler ‘epsilon’ gets smaller, the length L increases.

The Clinical Parallel: Why the Last 10% Takes 50% of the Time

At the start of a trial, we measure progress with a “large ruler” (e.g., “Open 50 sites”). It looks simple. But as we approach Database Lock, we switch to a “microscopic ruler” (e.g., “Reconcile the SAE on Patient 104-02”).

As the “ruler” gets smaller, the amount of work increases exponentially. This is why the final 10% takes 50% of the time: you have increased the resolution of the project, and in doing so, you have discovered the infinite complexity of the fractal edge. This is Fractal Time. You aren’t “almost done”; you are simply looking closer at an infinite coastline. To a Euclidean manager, this appears to be a delay. To a Fractal manager, this is just the nature of closing a complex system.

V. Other Applications: The Fractal Body and Pink Noise

The reason non-linear dynamics are so relevant to medicine is that the human body itself is a fractal masterpiece. 1. Fractal Physiology: Our lungs and vasculature are branching fractals.

This allows massive surface area (for oxygen exchange or drug absorption) to fit into a finite 3D volume. When we dose a drug, we aren’t dosing a “cylinder”; we are dosing a branching network. If the drug’s diffusion constant doesn’t match the fractal dimension of the tissue, the Phase 3 trial will fail despite Phase 1 success.

2. Oncology and Chaos: Tumor growth is non-linear. The “Fractal Dimension” of tumor margins is now used as a biomarker. Smooth margins (low D) imply stability; highly fragmented margins (high D) imply aggressive, chaotic invasion.

3. Financial “Fat Tails”: Mandelbrot proved markets follow Power Laws, not Bell Curves. Biotech portfolios suffer from “fat tail” risks where “impossible” crashes happen far more often than linear models predict. Trials don’t just “underperform”; they “jump” from one Fatou basin to another.

Conclusion: The “So What” – The Operational Topologist

The lesson of Fatou, Julia, and Mandelbrot is that boundaries are more important than averages.

In clinical development, we are obsessed with averages: average enrollment, average p-values, average site performance. But averages only matter in linear systems. In a non-linear system, the only thing that matters is where the boundary lies.

1. Calculate Parameter Distance: Before spending $200M, ask: “How close is our protocol (c) to the Mandelbrot boundary?” If a 10% change in screen failure rate pushes you into the “Chaos Zone,” your protocol is structurally flawed.

2. Map the Basin: Stop asking if sites are “doing a good job.” Ask if the system has fallen into a “Futility Basin.” If the feedback loops are generating more queries than they are resolving, you are in a divergent orbit.

3. Force a Lower Resolution: To escape the “Coastline Paradox” at the end of a trial, you must stop shrinking your ruler. You must define a “closure resolution” (e.g., frozen data snapshots) early, or you will be measuring the coastline forever.

4. Identify the Branch: In a fractal system, a problem at the “Patient Level” is usually just a high-resolution version of a “Leadership Level” failure. Zoom out. Fix the branch, not the leaf.

The future of drug development belongs to the Operational Topologist. They will understand that a trial doesn’t fail because people didn’t work hard enough. It fails because it was placed in a region of the map where failure is the only mathematical attractor.

Know your boundary, or the boundary will find you.

“Beautiful and chaotic things exist on the same boundary. The job of the executive is to keep the program in the heart.”

Nobody in the room is incompetent. The sites are not badly run. The CRO is not negligent. The data management team is not behind. And yet the trial is in trouble — structurally, irreversibly, in a way that more site visits and more patient recruitment support will not fix.

The reason is geometric. The trial entered its execution phase with operational assumptions that were sitting near a mathematical boundary — a threshold below which normal variation in execution is absorbed, and above which small additional degradations produce a catastrophic, irreversible cascade. Nobody identified that boundary before the trial started. Nobody monitored the distance to it during execution. By the time the meeting happens, the program crossed it six months ago.

This essay is about that boundary. What it is, how to find it before it finds you, and what operational management looks like when you take it seriously.

Part One: Three ideas in plain English

Fractals: why the same problem exists at every level

A fractal is a system where zooming in reveals the same pattern you saw when zoomed out. The canonical image is a tree: the trunk splits into major branches, each branch splits into smaller branches, each smaller branch splits into twigs, each twig ends in something that looks like a tiny tree. The structure repeats perfectly at every scale.

A clinical trial is also a fractal. The operational hierarchy — Program, Region, Country, Site, Patient — has the same structure at every level. And more to the point, the same set of operational problems appears at every level: activation delays, resourcing gaps, vendor dependencies, compliance risk, and communication lag. The pattern repeats.

This has one specific and uncomfortable implication. In a fractal system, a failure at an upper level does not produce one problem. It produces N problems simultaneously, where N is the number of nodes below it. A protocol amendment at the program level does not change one thing. It touches every region, every country, every site, and every patient currently in the trial — simultaneously. The amendments arrive at different teams at different times and appear unrelated. They are not. They share a common origin.

Why this matters operationally

When multiple sites show the same failure pattern simultaneously, the correct question is not ‘what is wrong with these sites?’ It is ‘what upstream node failure is generating this pattern across the network?’ The answer changes both the intervention and the urgency.

Figure 1 · Fractal Structure of a Clinical Trial

The fractal structure of a clinical trial. Program → Region → Country → Site → Patient. The same operational challenges repeat at every level. A disruption at the program level creates N simultaneous downstream problems — one per node below it. Self-similarity means a regional-level signal is a leading indicator of site-level failures that have not yet surfaced.

The Mandelbrot set: the map of operational stability

The Mandelbrot set is, in essence, a map of all possible starting configurations — showing which ones produce stable, bounded outcomes and which ones spiral out of control. The boundary between the two zones is the mathematically interesting part: it is infinitely complex, and a single step on either side determines everything.

You do not need mathematics to use this idea. Every clinical trial has a composite set of operational assumptions: site activation pace, screen failure projections, CRO capacity estimates, country regulatory timelines, dropout assumptions, IP shelf-life relative to the enrollment window. Think of this composite as a single point on a map. The Mandelbrot set tells you whether that point is in the stable zone (where normal execution variance is absorbed and the trial delivers), near the boundary (where modest degradation in any single parameter flips the trial to unrecoverable), or in the escaped zone (where no operational excellence can fix the structural incompatibility between the protocol and the execution plan).

The practical observation is this: most programs that enter execution in trouble are not in the escaped zone. Their operational plan was not fundamentally broken. It was near the boundary — sitting in a region where a screen failure rate drifting from 2:1 to 3:1, combined with a single key country running two regulatory cycles behind, combined with a modest CRO staffing reduction, was sufficient to push the composite past the threshold. None of those three events triggered an alarm individually. Together, they crossed the boundary.

The question of feasibility never asks

Standard feasibility asks: are our site and enrollment assumptions reasonable? The right question is: how much degradation across our assumptions is needed to make the trial unrecoverable — and are we comfortable with that margin? A parameter can pass every reasonableness check and still be sitting at the operational boundary.

Figure 2 · The Mandelbrot Set – operational parameter space

The Mandelbrot set as an operational stability map. Each pixel = one combination of operational parameters. Dark interior: orbit stays bounded — trial is robust. Colored boundary zone: orbit escapes slowly — trial is near its operational fragmentation threshold. Most operational failures originate here. Marker A (stable), B (boundary), C (escaped) illustrate three program positions.

Julia sets: how sensitive is your trial to starting conditions?

For each specific operational configuration — each position on the Mandelbrot map — there is a corresponding Julia set: a picture of which starting conditions (which sites activate first, which patients enrol early, which country comes online when) lead to trial success and which lead to failure.

When the operational composite is well within the stable zone, the Julia set is a single connected region. A wide range of starting conditions all lead to the same place: trial success. Whether Site 01 activates first or Site 58, whether Germany comes online in month 3 or month 7, the trial delivers. The outcome is robust to execution sequence.

When the composite is near the Mandelbrot boundary, the Julia set fractures into a lacy, thread-thin structure riddled with gaps. Now it matters intensely which sites activate first. If the high-enrolling sites happen to activate late and the low-enrolling sites fill the early cohort, the program enters a recovery trajectory it cannot complete within the timeline — even with full rescue operations. The same protocol, the same sites, a different activation sequence: different outcome. This is not bad luck. It is the structural property of a fragmented Julia set.

When the composite is outside the boundary, the Julia set fragments into disconnected dust. There is no connected basin of successful starting conditions. Some windows may briefly enrol adequately. The overall system cannot sustain itself.

The clinical translation

When a Phase 2 program produces the result ‘it worked in one subgroup analysis but not the overall population,’ the team typically looks for an operational explanation. Often, there is no satisfying one. The more accurate explanation is geometric: the program was running in the fragmented zone of its Julia set. Different patient cohorts were enrolled in different sequences, ending up in different disconnected regions of the outcome space. This is not noise. It is structure.

Figure 3 · Julia sets –– 3 operational fates

Julia sets — three operational fates. Left (connected): robust site network, wide basin of successful starting conditions. Center (fragmented): outcome highly sensitive to site activation sequence — same protocol, different timeline depending on which sites activate first. Right (Cantor dust): structural operational failure — no execution path leads to success.

Part Two: What to actually do

The framework generates four specific operational responses, each timed to a distinct phase of program management.

Stage 1 — Diagnose: map the parameter composite before the trial starts

The diagnosis happens at operational planning, before the first CTA submission. For each high-branching operational parameter — the ones whose failure cascades to multiple levels simultaneously — calculate two numbers: the current planned value, and the value at which the failure becomes irreversible. The distance between them is what matters.

This is not a standard risk register. Risk registers list risks and their probabilities. A parameter distance analysis maps how far each assumption sits from its fragmentation threshold. A parameter can have a low probability of failure and still be near the boundary — if the threshold is close to the planned value, normal variation is enough to cross it.

The parameter distance table

For each high-branching parameter:

• What is the planned value?

• What is the failure threshold — the value at which the cascade becomes irreversible?

• What is the distance between them?

• What is the typical variance for this parameter in comparable programs?

• Is the variance larger or smaller than the distance?

Parameters where the typical variance is roughly equal to the distance — or greater — are Amber-zone parameters. They need structural mitigation before the trial begins. Parameters with large distance-to-variance ratios are in the stable interior. Standard monitoring is sufficient.

Figure 4 · Paramater distance

The parameter distance table — the practical instrument. Six named Phase 3 parameters with planned value, failure threshold, distance, and variance. Screen failure rate and country activation are Amber-zone: typical variance is within the range of the distance to failure. These require structural mitigation at the design stage, not monitoring improvements.

Stage 2 — Prevent: build structural margin into boundary-zone parameters

Prevention happens at protocol design and vendor contracting. For each Amber-zone parameter, the response is structural — not a monitoring improvement, but a design change that increases the distance to the fragmentation threshold.

Screen failure rate

If screen failure is an Amber parameter, the response is over-siting: activate 25–30% more sites than the base-case model requires, specifically targeting sites with documented high-volume screening capacity in this indication. The cost is higher activation spend. The cost of not doing it is a mid-trial boundary crossing with no recovery path.

Country activation

If country activation is Amber, the response is country redundancy: identify two additional countries that can be added within 90 days if any primary country misses its critical activation milestone. These contingency countries should be at IND-ready status before primary country submissions begin. This is deliberate Julia set engineering — expanding the connected basin of viable execution outcomes so that a wider range of country activation sequences all lead to on-time enrollment.

Vendor capacity

If CRO monitoring capacity is Amber, the contractual response is a minimum staffing clause with a financial penalty for falling below the committed CRA-to-patient ratio, plus a pre-negotiated right to bring in a backup CRO within 60 days if the clause is triggered. This keeps the parameter in the stable interior even if the primary vendor experiences staffing disruption.

Stage 3 — Mitigate: classify before you intervene

When operational problems surface in a running trial, the first question is classification, not intervention. Is this a leaf failure or a branch failure?

• A leaf failure is localized to a specific node — one site, one country, one vendor interaction. It has limited downstream spread. Local intervention works.

• A branch failure has its origin at a high-branching node, and produces the same class of problem at multiple leaf nodes simultaneously. Local intervention does not work because the cascade from the upstream node continues arriving at new leaves faster than you can remediate existing ones.

The diagnostic test is structural: do the failing nodes share a common upstream cause? Seven geographically dispersed sites, all reporting high screen failure rates in the same two-month window, are not seven simultaneous leaf failures. It is one branch failure — the inclusion criteria are eliminating the available patient population, manifesting at seven leaves at once.

Once a branch failure is classified, the intervention moves up the hierarchy: protocol amendment, inclusion-criterion modification, CRO contract action, or country portfolio change. This requires more organizational authority and faster escalation than a leaf intervention. The fractal framework argues for keeping that escalation path short and pre-rehearsed, because branch failures require decisions at exactly the moment when organizational pressure is highest to continue with incremental leaf-level rescue.

Pre-committed branch-level triggers

For each Amber-zone parameter identified at Stage 1, specify in advance: the trigger threshold, the response, and the response time. For example, if screen failure exceeds 2.5:1 for two consecutive months, activate contingency sites and convene an I/E review within 30 days. Pre-commitment removes the decision from the moment of maximum pressure to continue.

Stage 4 — Monitor: track distance, not just performance

Standard operational monitoring tracks performance against target: enrollment curve, data completeness, protocol deviation rates, and site performance rankings. These are all leaf-level outputs. They show what has already happened at the leaves.

Parameter distance monitoring tracks the operational composite against its fragmentation thresholds — at the branch level, at sufficient frequency to detect boundary drift before it manifests at the leaves.

The living parameter map

The parameter distance table from Stage 1 becomes a living document, updated monthly for Amber-zone parameters and quarterly for stable-interior parameters. For each parameter, track: current value, trend direction, distance to threshold, and rate of boundary approach. A parameter performing within target but drifting toward its threshold is a leading indicator — it requires attention before it becomes a problem.

Branch-level leading indicators

Because the trial hierarchy is self-similar, a regional-level anomaly is a leading indicator of site-level problems that have yet to surface. A regional screen failure rate running 20% above the trial average is not yet a crisis — individual sites in that region may each be within normal variance. But the regional branch signal indicates that leaf-level failures are imminent. The structure of the system commits them before they appear.

Acting on the regional signal rather than waiting for individual site flags is the difference between early-stage intervention and late-stage rescue. In a program near the Amber boundary, it is sometimes the difference between a recoverable adjustment and an unrecoverable cascade.

The underlying principle

All four stages of this framework rest on one observation that fractal mathematics makes precise:

In complex, branching, self-similar systems — which is what a Phase 3 clinical trial is — the information that predicts failure is structurally available before the failure occurs. It sits at the branch level, weeks to months before it is visible at the leaf level.

The operational task is never ‘what went wrong?’ It is always ‘what branch failure is generating the leaf signals we are currently seeing, and how long ago did it occur?’

The failure rate in Phases 2 and 3 is not primarily a failure of science or execution. It is a failure of parameter awareness — of not knowing, at the moment when it would be actionable, how close the program’s operational assumptions are to the threshold where normal variation is sufficient to produce an irreversible cascade.

The Mandelbrot boundary does not care whether the assumptions were reasonable. It only cares whether the composite lies to its right. Knowing where the boundary is — before the trial starts — is the whole job.

The companion LinkedIn post (Biotech Triallist newsletter) covers the practical tools without the framework — parameter distance table, three-zone classification, and the leaf-vs-branch diagnostic. The figures referenced throughout this piece are available as downloadable PNG files.

Discovery scientists are the border collies. Cytokine pathways, inflammatory cascades, and cell signalling networks are the sheep.

And right now, most of us are stampeding them.

A paper just published in Science Advances by engineers at Georgia Tech — studying actual sheepdogs in an actual competition — has formalised this mathematically. The biological pathways we are trying to control are not waiting to be blocked. They are switching, oscillating, and rerouting continuously. The drug development industry has been designing interventions as if the pathways stand still. They do not.

The biology is not a machine that integrates your signal and responds proportionally. It is an indecisive flock — switching between states, attending to one signal at a time. The scientist who understands that has an enormous advantage over the one who does not.

Read more

We're going to dive into two pivotal papers that illuminate this duality. The first, published in Arthritis & Rheumatology, highlights how defective efferocytosis is a root cause of inflammation in a range of rheumatic diseases.¹ The second, appearing in Science Immunology, uncovers a novel, pro-inflammatory form of efferocytosis, aptly named "efferoptosis," which can wreak havoc in acute inflammatory states like sepsis.¹ Together, these studies paint a nuanced picture, suggesting that the precise modulation of efferocytosis could be a universal therapeutic target.

Paper 1: When the Cleanup Crew Fails – Defective Efferocytosis in Autoimmune Diseases

The Arthritis & Rheumatology paper, "Efferocytosis and its role in rheumatic diseases," lays out the fundamental importance of efferocytosis in maintaining our health.¹ It's a three-phase process:

1. The "Smell Phase": Dying cells release "find me" signals like sphingosine-1-phosphate (S1P) and nucleotides, attracting phagocytes (our cleanup cells) to the scene.¹

2. The "Eating Phase": Apoptotic cells expose "eat me" signals, primarily phosphatidylserine (PS), on their surface. Phagocytes recognize these signals directly or via bridging molecules like MFGE8 and Gas6, then engulf the dying cells.¹

3. The "Digestion Phase": Inside the phagocyte, the ingested cellular material is broken down. This isn't just disposal; it's a metabolic reprogramming that actively produces anti-inflammatory mediators like IL-10 and TGF-β, promoting tissue repair and immune tolerance.¹

This efficient, silent clearance is crucial. It prevents dead cells from undergoing "secondary necrosis," a messy process that spills their contents into the body.¹ These spilled contents, known as damage-associated molecular patterns (DAMPs) – like nucleic acids and histones – are highly inflammatory. They activate immune receptors, triggering a cascade of pro-inflammatory cytokines such as type I interferons (IFN), TNF, and IL-6, which fuel chronic inflammation and autoimmunity.¹

The Autoimmune Connection: A Cascade of Consequences

When efferocytosis falters, this delicate balance is shattered. The Arthritis & Rheumatology paper details how defective efferocytosis is a central mechanism in many autoimmune diseases:

• Systemic Lupus Erythematosus (SLE): This is a prime example. Genetic mutations in complement components (like C1Q) or DNA-degrading enzymes (DNase 1, DNASE1L3) impair the clearance of apoptotic cells.¹ This leads to an accumulation of dead cell debris in tissues, which then exposes self-antigens and DAMPs. These DAMPs activate pathways like cGAS-STING, driving the production of type I IFN and other inflammatory cytokines, perpetuating the disease.¹

• Rheumatoid Arthritis (RA): In RA, there's a significant reduction in specialized efferocytic macrophages in the joint lining.¹ This impaired clearance contributes to persistent inflammation, enhanced bone destruction, and reduced tissue repair.¹

• Sjögren's Syndrome (SS): Increased apoptosis of glandular cells, coupled with defective efferocytosis, leads to the accumulation of dead cells in salivary and lacrimal glands. This exposes self-antigens and DAMPs, amplifying autoimmune responses.¹

• ANCA-associated Vasculitis (AAV): Autoantigens like PR3 and MPO-ANCA directly interfere with efferocytosis pathways, promoting inflammation and hindering the clearance of neutrophils.¹

• Systemic Sclerosis (SSc): Impaired efferocytosis contributes to the widespread fibrosis seen in SSc, fostering autoantibody production and chronic inflammation that activates fibroblasts and enhances collagen deposition.¹

• Antiphospholipid Syndrome (APS): Antiphospholipid antibodies (aPL) interfere with the normal clearance of apoptotic cells, triggering pro-inflammatory cytokine release and driving disease progression.¹

• Gout and Osteoarthritis (OA): Even in conditions like gout and OA, impaired efferocytosis of inflammatory cells or joint tissue debris contributes to persistent inflammation and tissue damage.¹

Therapeutic Promise: Restoring the Balance

The insights from this paper highlight clear therapeutic strategies:

• Apoptotic Cell (AC) Infusion: Administering ACs or their metabolites can "overwhelm" defective clearance mechanisms, promoting an anti-inflammatory response. A clinical trial for refractory RA is already proposed.¹

• Bridging Molecules: Molecules like Gas6 and MFGE8, which help phagocytes recognize and bind to apoptotic cells, can be administered to boost clearance.¹

• DNase Supplementation: For genetic defects in DNA degradation, providing exogenous DNase could prevent the accumulation of inflammatory DNA.¹

• Anti-CD47 Antibodies: Blocking the "don't eat me" signal (CD47) can enhance phagocytic clearance.¹

• PPAR/LXR Agonists: These can improve cholesterol management within phagocytes, preventing inflammation triggered by lipid overload.¹

• Mesenchymal Stromal Cell (MSC) Infusion: MSCs can generate apoptotic debris, which promotes efferocytosis and shifts phagocytes towards an anti-inflammatory state.¹

Paper 2: When the Cleanup Crew Turns Rogue – Efferoptosis in Acute Inflammation

The second paper, "TNF switches homeostatic efferocytosis to lytic caspase-8-dependent pyroptosis and IL-1β maturation," published in Science Immunology, reveals a darker side of efferocytosis.¹ It introduces "efferoptosis," a novel form of inflammatory cell death.

Traditionally, efferocytosis is anti-inflammatory. But in acute, dysregulated inflammatory environments, such as sepsis or systemic inflammatory response syndrome (SIRS), the presence of high levels of Tumor Necrosis Factor (TNF) acts as a "master switch".¹ When phagocytes, particularly macrophages, engulf dead or dying neutrophils in the presence of TNF, they don't just clear them silently. Instead, they undergo a lytic, pro-inflammatory form of cell death: efferoptosis.¹

The Molecular Mayhem of Efferoptosis

Efferoptosis is distinct from other forms of inflammatory cell death:

• Caspase-8 Dependent, NLRP3 Independent: Unlike canonical pyroptosis, which relies on NLRP3 and caspase-1, efferoptosis is driven by caspase-8. This activated caspase-8 directly cleaves gasdermin-D (GSDMD), a key protein that forms pores in the cell membrane, leading to cell lysis.¹

• Direct IL-1β Maturation: Crucially, caspase-8 also directly cleaves pro-IL-1β, leading to its maturation and release, bypassing the usual inflammasome activation pathway.¹

• The TRIFosome: This process involves a complex called the "TRIFosome," formed by the TLR4 adaptor TRIF, ZBP1, and RIPK1. This complex activates caspase-8.¹

• Signaling Rewiring: Normally, efferocytosis inhibits pro-inflammatory NF-κB signaling. However, in efferoptosis, TNF-activated efferocytosis inhibits TAK1/NF-κB, leading to the downregulation of prosurvival factors like cFLIP. Simultaneously, PLCy/MAPK signaling is sustained, which upregulates pro-IL-1β, ensuring a substrate for caspase-8.¹

Pathological Impact: Sepsis and Beyond

Efferoptosis significantly contributes to the pathology of sepsis and SIRS. In mouse models, inhibiting efferocytosis (e.g., via a TIM3 antibody) protected mice from TNF-induced SIRS, reducing macrophage death and improving survival.¹ This suggests that in these acute inflammatory conditions, the negative impacts of efferoptosis outweigh the beneficial functions of homeostatic efferocytosis.¹

Speculating on Myocardial Infarction

While the Science Immunology paper focuses on sepsis, the mechanisms of efferoptosis have profound implications for other acute inflammatory events, such as myocardial infarction (MI). MI involves massive cell death in the heart, leading to a robust inflammatory response.

Consider these connections:

• Extensive Cell Death: MI results in a large number of dying cardiomyocytes. These apoptotic cells, if not cleared efficiently, can release DAMPs, triggering inflammation.¹

• Cholesterol Overload: The Arthritis & Rheumatology paper highlights that in atherosclerosis (a major cause of MI), cholesterol accumulation from uncleared apoptotic cells can trigger macrophage apoptosis and NLRP3 inflammasome activation.¹ This adds another layer of inflammatory cell death.

• TNF and Efferoptosis: Post-MI, there's a significant inflammatory response, often including elevated TNF levels. It's highly plausible that macrophages engulfing dying heart cells and recruited neutrophils in the damaged heart, under the influence of TNF, could undergo efferoptosis. This would contribute to the inflammatory burden and adverse cardiac remodeling, similar to its role in SIRS.¹ The direct cleavage of IL-1β by caspase-8 in efferoptosis could be a significant driver of sterile inflammation in the infarcted heart.

Therefore, targeting efferoptosis, perhaps through caspase-8 inhibition or specific PS receptor modulation (like TIM3 inhibition), could represent a novel therapeutic strategy to reduce post-MI inflammatory injury, distinct from targeting canonical inflammasomes.

The Converging Insights: A Unified View of Inflammation

These two papers, from different journals and focusing on seemingly distinct disease categories, reveal a profound commonality: efferocytosis is a double-edged sword. It is absolutely essential for maintaining immune tolerance and resolving inflammation.¹ But it can become a potent source of inflammation if it is either:

1. Defective: Leading to the accumulation of uncleared apoptotic cells and the release of DAMPs, driving chronic autoimmune diseases.¹

2. Aberrantly Activated (Efferoptosis): Where, under acute inflammatory conditions like high TNF, the very act of efferocytosis triggers a pro-inflammatory, lytic cell death in the phagocyte itself, fueling acute inflammatory crises.¹

The common theme is the critical need for precise modulation of efferocytosis. We need to enhance it when it's failing (as in autoimmune diseases) and prevent its detrimental pro-inflammatory switch when it's being hijacked (as in acute inflammation like sepsis and potentially MI).

The therapeutic landscape is exciting. Strategies that boost efferocytosis (like AC infusions or DNase supplementation) can restore immune balance in chronic conditions.¹ Meanwhile, interventions that prevent efferoptosis (such as TIM3 inhibition or targeting caspase-8) could mitigate acute inflammatory damage.¹ The challenge lies in developing therapies that can differentiate between these contexts or be delivered in a highly targeted manner to specific tissues, as the role of efferocytosis can be tissue-specific.¹

This converging understanding offers a new paradigm. Instead of merely suppressing inflammation, we can aim to restore the fundamental efferocytic balance, offering more profound and sustained therapeutic effects. The future of immune therapies may well lie in mastering the art of the cleanup crew – ensuring they always work for us, never against us.

Here’s why sleep loss deserves the same attention as diabetes—and what it means for drug development.

The Metabolic Chaos of Sleep Loss

Like diabetes, sleep deprivation disrupts how your body handles glucose, ramps up inflammation, and throws your metabolism into disarray. A 2015 Science Signaling review (DOI: 10.1126/sciadv.1504018) shows that just 4–6 hours of sleep deprivation in mice spikes adenosine levels in the hippocampus, impairing glucose uptake and mimicking insulin resistance seen in diabetes. This isn’t just a brain issue—peripheral tissues like adipocytes also show reduced insulin sensitivity, leading to fat accumulation and oxidative stress. For drug developers, this points to a clear target: metabolic pathways disrupted by sleep loss.

Inflammation and Oxidative Stress: A Shared Culprit

Sleep loss doesn’t just tire you out; it ignites a firestorm of inflammation. The Science Signaling study found that sleep deprivation boosts oxidative stress proteins and lipid droplet accumulation in glial cells, which then rely on β-oxidation to cope. This mirrors the chronic inflammation in diabetes, where oxidative stress damages tissues and drives complications like cardiovascular disease. In the context of hypertrophic cardiomyopathy (HCM), a condition linked to sudden death in young athletes, sleep deprivation could amplify inflammation, worsening cardiac fibrosis. Therapies targeting oxidative stress—like antioxidants or anti-inflammatory biologics—could address both sleep loss and diabetes-related damage.

Synaptic and Systemic Fallout

The brain takes a hit from sleep loss, much like it does in metabolic disorders. The study highlights how sleep deprivation downregulates neural plasticity by increasing adenosine and A1R activity, leading to memory deficits. This is eerily similar to diabetes-induced cognitive decline, where poor glucose metabolism harms neurons. Systemically, sleep loss redirects energy away from non-essential processes, like synapse formation, to cope with metabolic stress—paralleling diabetes’ energy misallocation. Drug developers could explore adenosine receptor antagonists or neuroprotective agents to mitigate these effects, potentially benefiting both conditions.

HCM and Sleep: A Deadly Combo

For those with HCM, sleep deprivation could be a silent killer. The Science Translational Medicine study (DOI: 10.1126/scitranslmed.aad2516) showed that inflammation drives HCM’s progression, with regulatory T cells (Tregs) struggling to control it. Sleep loss exacerbates this by upregulating inflammatory pathways, potentially increasing the risk of sudden cardiac death in young athletes. Imagine a college basketball player, burning the candle at both ends, unaware that their heart is under extra strain. Therapies like IL-2 agonists (e.g., sifalimumab) or PD-1 agonists, which boost Treg function, could tackle inflammation in both HCM and sleep-deprived patients, offering a dual-purpose pipeline.

Drug Development Opportunities

The parallels between sleep loss and diabetes open exciting avenues for drug development:

• Adenosine receptor modulators: Blocking A1R could restore neural plasticity and glucose metabolism, addressing cognitive and metabolic deficits.

• Anti-inflammatory biologics: IL-2 or PD-1 agonists, inspired by HCM research, could reduce systemic inflammation, benefiting both sleep-deprived and diabetic patients.

• Antioxidants: Targeting oxidative stress could protect tissues from the damage caused by sleep loss and diabetes.

• Insulin sensitizers: Drugs like metformin might be repurposed to improve glucose uptake in sleep-deprived individuals.

Clinical trials could focus on biomarkers like adenosine levels, inflammatory cytokines, or LGE-CMR for fibrosis, accelerating the path to market for these therapies.

Why It Matters for You

Sleep loss isn’t just about feeling groggy—it’s a metabolic crisis that could shorten your life, especially if you’re at risk for conditions like HCM. For drug developers, it’s a call to action: prioritize sleep as a therapeutic target with the same urgency as diabetes. The science is clear, and the stakes are high.

Want to stay ahead on groundbreaking therapies? Subscribe to www.drugdevelop.com for the latest in cardioimmunology, metabolic disease, and more. Let’s wake up to the power of sleep—and save lives in the process.

Citations:

1. [Authors]. (2015). Sleep loss is a metabolic disorder. Science Signaling, 8, adp9358. DOI: 10.1126/sciadv.1504018

2. Wang, Y.-J., et al. (2015). Regulatory T cells attenuate chronic inflammation and cardiac fibrosis in hypertrophic cardiomyopathy. Science Translational Medicine, 7, eaad2516. DOI: 10.1126/scitranslmed.aad2516

The story begins with an elegant scientific plan. It ends with a molecular lesson about the humbling complexity of human biology. And it should change how we think about target selection, translational strategy, and the seductive confidence that comes from a clean in vitro result.

What follows is that story — and three mathematical frameworks that let us reason about it more rigorously.

The Architecture of Confidence

In drug discovery, we build cathedrals on paper. We map pathways, identify nodes, select targets, and construct rationales of extraordinary logical coherence. The best programs feel almost inevitable in their design. Every arrow points in the right direction. Every experiment confirms the hypothesis.

This confidence has occasionally been justified. Gleevec — imatinib — is the canonical example: a single kinase target, BCR-ABL, a single disease defined by a single chromosomal translocation, and a clinical result so dramatic it reshaped oncology. The principle seemed sound: find the right target, block it precisely, and the disease collapses.

But CML is, in retrospect, the exception that proved a much more uncomfortable rule. BCR-ABL is a genuine single point of failure in a cancer that has, for molecular reasons, stripped away most of its redundancy. Most diseases are not like this. Most biology is not like this.

Biological systems did not evolve to be druggable. They evolved to survive. Redundancy, pathway crosstalk, and adaptive rewiring are not bugs — they are the fundamental architecture of resilience.

The assumption that a single, well-chosen target is sufficient to collapse a complex inflammatory or metabolic disease is what I call the Target Fallacy. It is not a foolish assumption — it is a reasonable prior, given some remarkable successes. But it is a prior we have updated too slowly, and the Novartis NEK7 story is the latest evidence of why.

The Plan Was Beautiful

The NLRP3 inflammasome has been one of the most intensively pursued drug targets in inflammation for over a decade. Its activation drives IL-1β and IL-18 release, and dysregulated NLRP3 activity has been implicated in gout, CAPS, atherosclerosis, Alzheimer’s disease, and type 2 diabetes complications.

The Novartis team focused on NEK7, a mitotic kinase that moonlights as an essential scaffold for NLRP3 inflammasome assembly. Critically, NEK7’s role in inflammasome function is independent of its kinase activity — it acts as a structural bridge. The team built a molecular glue degrader, NK7-902, that recruits the CRBN E3 ubiquitin ligase machinery to NEK7, tagging it for proteasomal destruction. NK7-902 was potent, selective, and mechanistically clean, achieving greater than 95% degradation of NEK7 protein.

The preclinical data in mice was compelling. Oral administration strongly degraded NEK7 in splenic tissue and produced meaningful inhibition of IL-1β release in both an acute peritonitis model and a CAPS disease model.

Then came the human and NHP data. In human peripheral blood cells, even with virtually complete NEK7 degradation, IL-1β suppression was partial and highly variable across donors. In non-human primates, sustained 95%+ NEK7 degradation produced only transient and partial inhibition of IL-1β release.

The target was eliminated. The biology barely noticed.

Mathematical Framework I — Pathway Redundancy

The NEK7 result is not mysterious once we model the inflammasome as a network with multiple weighted activation routes, rather than a single linear pathway.

The Core Equation

Define the total NLRP3 activation A as the sum of contributions from all activation routes:

A = Σᵢ wᵢ · aᵢ

where wᵢ is the fractional weight of route i and aᵢ is its activation level (0 to 1). By definition, Σᵢ wᵢ = 1.

When we block route j (e.g., by degrading NEK7), the remaining activation is:

A_blocked = Σᵢ≠ⱼ wᵢ · aᵢ = 1 − wⱼ · aⱼ

This is the key insight. For NEK7 degradation to suppress NLRP3 activity by ≥80%, we need:

wⱼ ≥ 0.80

In mouse cells, NEK7 carries approximately 90% of activation weight — so eliminating it nearly collapses the system. In human and NHP cells, NEK7 appears to carry only ~35% of the weight, with substantial traffic flowing through PTEN-L, mtDNA sensing, and potassium efflux routes. Blocking one road does not close the network.

Figure 1. Pathway Redundancy Model. Panel A: Residual NLRP3 activation as a function of NEK7 blockade under mouse vs. human route-weight assumptions. At 95% NEK7 degradation (dashed red line), mouse activation falls below the therapeutic threshold while human/NHP activation remains well above it. Panel B: Estimated fractional contributions of each activation route by species. The NEK7 dominance in mice versus the distributed architecture in humans explains the translational failure.

This framework makes the failure predictable, not mysterious. Any team that had estimated human route weights before committing to lead optimization would have recognized that NEK7 degradation, however complete, was unlikely to produce therapeutic-level pathway suppression in primates.

The practical implication: before selecting a target in a redundant pathway, estimate the target’s fractional contribution in human tissue. This is now possible with modern tools — donor-matched primary cell experiments, siRNA knockdown across a diverse human panel, and pathway flux analysis in organoids. It is not free. But it is far cheaper than a Phase II failure.

Mathematical Framework II — Bayesian Translational Confidence

The NEK7 story is also a story about how evidence should change our beliefs — and about how the current drug development system is structured to update beliefs too slowly, too late.

The Framework

Model our belief in the mechanism as a probability θ — the probability that the mechanism works in humans. We can represent our state of knowledge at each stage as a Beta distribution, Beta(α, β), which is a natural choice for probabilities and updates cleanly as new data arrives.

Prior belief: θ ~ Beta(α₀, β₀)

Each experimental stage provides new evidence. After observing s successes and f failures in a relevant species, we update:

Posterior: θ | data ~ Beta(α₀ + s, β₀ + f)

The mean of this distribution is α/(α+β), which is our best estimate of P(mechanism works in humans) given all evidence so far.

Applying It to the NEK7 Program

A reasonable prior for a target-based program in inflammation is Beta(3, 5), giving a mean P ≈ 0.375 — consistent with historical Phase II success rates of 30-40% in this space.

Positive mouse efficacy in 4 of 5 models raises the posterior to Beta(7, 6), mean ≈ 0.54. This is the stage where many teams accelerate into IND-enabling studies. The evidence looks convincing.

But then add the NHP pharmacodynamic data: partial and transient IL-1β suppression despite complete target degradation. This is a near-failure — one success out of effectively zero in the most translatable species. The posterior falls to Beta(7, 7), mean ≈ 0.50. Confidence has eroded.

Add the human primary cell variability — inconsistent suppression across donors — and the posterior drops further to Beta(7, 9), mean ≈ 0.44. The belief distribution has shifted materially left. We are now in genuine uncertainty about whether this mechanism is pharmacologically sufficient in humans.

Each experiment is a likelihood update. The question is not whether your compound works in the system you tested. It is whether the full evidence profile justifies conviction in the system that matters — the human patient.

Figure 2. Bayesian Updating of Translational Confidence. Panel A: Beta distributions representing belief in mechanism validity at each evidence stage. Mouse efficacy raises confidence; failed NHP PD and inconsistent human cell data erode it. Panel B: Mean P(mechanism valid) ± 80% credible interval at each stage. The trajectory reveals how mouse-only data produces false conviction that later evidence must correct — often at great expense.

The critical observation from this framework is not the final number. It is the shape of the update trajectory. Programs that incorporate human-relevant data early get the bad news when it is cheap. Programs that defer human data accumulate false confidence, consume resources, and then face a large, late, costly correction.

There is a structural drug-development argument here: the expected cost of generating human primary-cell data in lead optimization is roughly $2-5M for a well-run campaign. The expected cost of a Phase II failure — after the Bayesian update arrives two years later, in the clinic — is $50-200M. The mathematics of belief updating strongly favors early investment in human-relevant evidence.

Mathematical Framework III — Expected Value of Development

The Bayesian framework tells us how evidence should change our beliefs. The expected value framework asks the harder question: given a realistic probability of mechanism validity, which preclinical strategy maximizes the expected return on drug development investment?

The Model

Define the expected value of a program as:

EV = P(mechanism valid) × P(clinical success | mechanism valid) × 
V(approval) − C(preclinical) − P(enter Phase I) × 
[C(Phase I) + P(Phase II) × (C(Phase II) + P(Phase III) × C(Phase III))]

where V(approval) is the net present value if the drug reaches market, and each C(stage) represents the expected cost of that stage.

The key variable is P(mechanism valid) — which is precisely what the choice of preclinical strategy determines. A mouse-centric program produces a higher prior estimate of this probability (because mouse models are confirmatory), but a lower true probability. An early human translation program produces a more accurate — and typically lower — estimate, but one that the team can actually rely on.

Why the Arithmetic Favors Early Human Translation

Assume a mechanism with a true P(valid in humans) = 0.20 — typical for a novel inflammatory target. A standard mouse-centric program might estimate this at 0.30-0.40 based on confirmatory rodent data. An early human program would correctly estimate it at or close to 0.20, potentially identifying the translational gap before entering clinical development.

The standard program spends less in preclinical work ($15M vs. $28M with human tissue and NHP PD) but then runs Phase I, Phase II, and potentially Phase III on the basis of a false prior. The early human program front-loads the cost of truth.

The expected value comparison, modeled across a range of mechanism validity probabilities, consistently favors early human translation once the mechanism validity falls below approximately 30-35% — which covers the majority of programs in inflammation, CNS, and metabolic disease.

Figure 3. Expected Value of Development by Preclinical Strategy. Panel A: Program EV across a range of pre-human mechanism validity probabilities. The early human strategy (blue) generates higher EV across most of the realistic probability range, with the advantage increasing as mechanism uncertainty rises. Panel B: Expected cost by development stage at P(mechanism) = 20%. Higher preclinical investment in the human-first strategy is more than offset by reduced downstream expected failures.

The cost of false confidence is not paid at the preclinical stage. It is paid in Phase II, when the evidence finally arrives — too late, too expensive, and with patients enrolled.

What Needs to Change

These three frameworks point to the same practical agenda.

First, measure route weights before committing to a target. If a target’s fractional contribution to pathway activation in human primary cells is below 50%, the program faces a high burden to demonstrate therapeutic sufficiency. This is not a reason to abandon the target — it is a reason to investigate polypharmacology, combination strategies, or direct pathway blockade downstream of the redundant nodes.

Second, treat Bayesian updates as decision triggers, not post-hoc observations. Programs should define, in advance, what posterior probability threshold justifies progression to IND-enabling studies. NHP pharmacodynamic failure should move a program into formal go/no-go review, not result in an explanation and a pivot to the next mouse model.

Third, use the EV framework to justify early investment in human translation. The argument that human tissue experiments are “too expensive” for early discovery fails basic expected value arithmetic. The question is never whether early human data costs money. The question is how that cost compares to the expected cost of the alternative.

Fourth, be structurally honest about the limitations of mouse models. A positive result in a murine inflammasome model is evidence that the mechanism is pharmacologically tractable in that species. There is no evidence that the mechanism is relevant in humans. These are different claims, and the organizational habit of treating the former as evidence for the latter has cost the industry billions of dollars and years of patients’ lives.

The Novartis Team Did the Right Thing

The Novartis scientists did not make a mistake. They asked a precise scientific question, built precise tools to answer it, and reported the answer transparently — including the parts that were uncomfortable. That is what science is supposed to do.

The problem is not the science. The problem is an ecosystem that consistently asks the wrong question first: “Can we hit the target?” rather than “Is this target sufficient in humans?” and pays for that inversion at a late and expensive stage.

The NEK7 data suggest that direct NLRP3 inhibitors — which act downstream of the redundant activation routes — may have advantages in human systems. Programs using compounds such as selnoflast and inzomelid, currently in clinical development, will be informative on this point.

But the broader lesson is about first principles. Biological systems are networks, not flowcharts. They evolved to survive perturbation. The three frameworks described here — pathway redundancy, Bayesian belief updating, and expected program value — are not academic exercises. They are practical tools for structuring discovery decisions, ensuring we ask the right questions in the right order.

The most important question in drug discovery is not whether your compound works. It is whether your biology is asking for what your compound offers.

#DrugDiscovery#NLRP3#TranslationalMedicine#InflammasomeBiology#MolecularGlue#PreclinicalModels#TargetFallacy

The story begins with an elegant scientific plan. It ends with a biological lesson about redundancy, evolution, and the uncomfortable gap between species. And it should change how we think about target selection, translational strategy, and the seductive confidence that comes from a clean result in a mouse.

The Architecture of Confidence

In drug discovery, we build cathedrals on paper. We map pathways, identify nodes, select targets, and construct rationales of extraordinary logical coherence. The best programs feel almost inevitable in their design. Every arrow points in the right direction. Every experiment confirms the hypothesis.

This confidence has occasionally been justified. Gleevec — imatinib — is the canonical example: a single kinase target, BCR-ABL, a single disease defined by a single chromosomal translocation, and a clinical result so dramatic it reshaped oncology. Find the right target, block it precisely, and the disease collapses. The principle seemed sound.

But CML is, in retrospect, the exception that established a flawed template. BCR-ABL is a genuine single point of failure in a cancer that has, for molecular reasons, stripped away most of its biological redundancy. Most diseases are not like this. Most biology is not like this.

Biological systems did not evolve to be druggable. They evolved to survive. Redundancy, pathway crosstalk, and adaptive rewiring are not bugs — they are the fundamental architecture of resilience.

The assumption that a single, well-chosen target is sufficient to collapse a complex inflammatory or metabolic disease is what I call the Target Fallacy. It is not a foolish assumption — it is a reasonable one, given some remarkable successes. But it is an assumption we have updated too slowly, and at extraordinary cost.

The Plan Was Beautiful

The NLRP3 inflammasome has been one of the most intensively pursued drug targets in inflammation for over a decade. Its activation drives the release of interleukin-1β and IL-18 — inflammatory signals implicated in gout, rheumatoid arthritis, cardiovascular disease, Alzheimer’s disease, and a spectrum of autoinflammatory conditions. Getting control of NLRP3 activation has been a central ambition of the inflammatory disease field.

The Novartis team focused on NEK7, a mitotic kinase that serves as an essential structural scaffold for NLRP3 inflammasome assembly. The key insight was that NEK7’s role in the inflammasome is not catalytic — it does not drive a chemical reaction — but architectural. It holds the complex together. This means that rather than trying to block an enzymatic activity, you could simply remove the protein entirely.

The drug they built to do this — NK7-902 — was a molecular glue degrader. It works by tricking the cell’s own disposal machinery into destroying NEK7, tagging it for proteasomal degradation via the CRBN E3 ubiquitin ligase system. NK7-902 was potent, selective, and mechanistically elegant. In biochemical assays, it worked exactly as designed: it degraded more than 95% of NEK7 protein.

On paper, eliminating the scaffold should collapse the inflammasome. The logic is sound. The chemistry is elegant. The execution was flawless. And in mice, it worked.

Figure 1 — Why the same drug succeeds in mice and fails in primates. Panel A: in mouse cells, NEK7 is the dominant gateway for NLRP3 activation, carrying roughly 90% of activation traffic. Blocking it effectively shuts down the system. Panel B: in human and NHP cells, activation traffic distributes across multiple parallel routes, most of which bypass NEK7 entirely. Eliminating one node reroutes — it does not stop — the pathway. Note: route weight estimates are conceptual and illustrative of the biological principle.

The Biology Disagreed

The preclinical data in mice was encouraging. Oral NK7-902 strongly degraded NEK7 in splenic tissue and produced meaningful inhibition of IL-1β release in both an acute peritonitis model and a CAPS disease model. The team had a molecule that worked, a target that was tractable, and a disease rationale that held in rodents. Any discovery team would have been cautiously optimistic.

Then came the human and non-human primate data, and the story changed entirely.

In human peripheral blood cells, even with virtually complete NEK7 degradation, IL-1β suppression was partial and highly variable — differing substantially across individual donors and experimental conditions. That variability is itself a signal. When a mechanistically clean intervention produces inconsistent functional effects across human donors, the system is communicating something important about its internal architecture.

In non-human primates, the dissociation was even more striking. Oral dosing at 2 mg/kg produced sustained NEK7 degradation exceeding 95% in blood — the compound had done exactly what it was designed to do. IL-1β inhibition was transient and partial. The target was eliminated. The biology barely noticed.

The target was eliminated. The biology barely noticed. This is a sentence that should be framed and hung in every drug discovery department in the world.

What this tells us is not that NEK7 is unimportant in inflammasome biology. In mouse cells, it clearly is important — essential, even. What it tells us is that the inflammasome pathway in primates has evolved redundant activation routes that do not depend on NEK7. When one road is blocked, the system takes a different highway. Evolution, unlike our target-validation assays, tested this over millions of years.

Figure 2 — The efficacy disconnect and what the evidence should tell us. Left: NK7-902 achieves near-identical NEK7 degradation across all three species. IL-1β suppression tracks with species — not with compound performance. Right: how confidence in the mechanism should update as each piece of evidence arrives. Mouse efficacy raises it; NHP and human data erode it. Programs that collect human-relevant data early get the correction cheaply. Programs that defer it pay compound interest.

Why This Keeps Happening

The NEK7 story is not unique. It sits in a long and expensive tradition. The p38 MAPK inhibitors showed exceptional preclinical promise in rheumatoid arthritis for nearly a decade before a string of clinical failures revealed that the human RA synovium is far more pathway-redundant than any rodent model had suggested. The CANTOS trial demonstrated that blocking IL-1β with canakinumab reduces cardiovascular events — the biology was real — but the increase in fatal infections and the commercial calculus made it unworkable. The list of programs that succeeded in mice and failed in humans is long enough to be its own clinical speciality.

There is a structural reason this pattern persists. Mouse models are inexpensive, fast, and genetically manipulable. They generate clean, interpretable data. They are scientifically satisfying in a way that messy human primary cell experiments rarely are. The incentive architecture of early drug discovery — publication pressure, milestone timelines, the cost differential between rodent and human studies — consistently pushes teams toward the models most likely to produce affirmative results, not the models most likely to predict human outcomes.

We select for confidence. Biology rewards robustness.

The two are not the same thing, and the gap between them is where drug development money goes to disappear.

The Species Gap Is a Biological Fact, Not a Technical Problem

A common response to translational failures is to reach for better tools: humanized mice, organoids, organs-on-chips. These are genuinely valuable, and investment in them is warranted. But there is a more fundamental issue that better experimental systems alone cannot resolve.

Human and mouse immune systems diverged roughly 65 to 80 million years ago and have been under substantially different selective pressures ever since. The human NLRP3 pathway appears to have developed redundant NEK7-independent activation routes that are simply not present in the mouse — not because our models are imperfect, but because that is what evolution produced. This is not a gap in our understanding that better technology will close. It is a biological reality that demands a different strategic response.

The strategic response is to ask the critical question earlier. Not “does our target matter in this pathway?” but “is this pathway, in human tissue, sufficiently non-redundant that eliminating our target will produce the functional outcome we need?” These are different questions. The second one requires human data. And it needs to be asked before lead optimisation is complete — not at Phase II.

The question is not whether your target matters. The question is whether, in human biology, your target is a single point of failure — or merely one of several roads to the same destination.

The Cost of Getting This Wrong

The financial argument for earlier human translational work is less intuitive than it should be. Adding NHP pharmacodynamic studies and human tissue experiments in early discovery can double or triple the cost of lead optimisation. Program teams face pressure to move fast, hit milestones, and advance compounds. Spending more time and money preclinically feels like slowing down.

But the arithmetic is straightforward. A Phase II failure in an inflammatory indication costs between $50 million and $200 million, depending on trial size and duration. A Phase III failure costs an order of magnitude more. The industry-wide failure rate for drugs entering Phase II in inflammation and immunology is approximately 70%. If even a fraction of those failures trace back to a translational assumption that human primary cell data would have challenged — and the evidence suggests many do — the financial case for front-loading that investment is overwhelming.

There is also a patient cost that rarely enters these calculations directly. Every program that reaches clinical trials based on mouse efficacy data represents a commitment of patient time, risk, and hope. Phase II failures are not just expensive — they are a burden on people who enrolled because they or their physicians believed the mechanism had a reasonable chance of working. Better preclinical translation does not just protect pipeline value. It protects patients.

Figure 3 — The economics of early human translation. Left: expected program value as a function of true mechanism probability in humans. The human-first strategy consistently outperforms once mechanism uncertainty rises above roughly 25%, which covers the majority of programs in inflammation, CNS, and metabolic disease. Right: the strategic decision timeline — the window to ask the human translational question is lead optimisation, not Phase II. Waiting until clinical studies to discover pathway redundancy is the most expensive form of discovery research.

What Needs to Change

The path forward is not a single intervention. It is a reorientation of how discovery programs are structured, and what counts as evidence of confidence versus evidence of translatability.

Human primary cell data should function as a gate, not a late confirmatory box to tick. When a compound shows robust efficacy in mouse models, the immediate question should not be which additional mouse models to run, but what happens in human primary cells and tissue. Variability across donors — as seen in the Novartis data — should be interpreted as information about the mechanism, not experimental noise to be averaged away.

NHP pharmacodynamic studies, where feasible, should be incorporated before lead optimisation is complete. This is a significant resource commitment, and not every program will justify it. But for mechanisms where the translational risk is structurally high — innate immune pathways, CNS targets, metabolic nodes — the investment is warranted. A partial NHP pharmacodynamic result, collected early, is worth far more than a complete one collected at IND.

Phenotypic approaches deserve serious investment alongside target-based programs. Phenotypic screens measure functional outcomes in human or humanised systems rather than target engagement alone. A compound that suppresses IL-1β release in human peripheral blood across a diverse donor panel — through whatever mechanism — has demonstrated human-relevant functional activity that no target-engagement assay can match. This is not an argument against target-based discovery. It is an argument for complementing it.

Finally, we need institutional honesty about what mouse efficacy data actually proves. A positive result in a murine model is evidence that the mechanism is pharmacologically tractable in that species. It is not evidence that the mechanism is sufficient in humans. Treating the former as evidence for the latter — routinely, under timeline and resource pressure — is the Target Fallacy in its most common and costly form.

What the NEK7 Story Really Tells Us

The Novartis scientists did not make a mistake. They asked a precise scientific question, built precise tools to answer it, and reported the answer honestly — including the parts that were uncomfortable. The paper is a model of scientific integrity. That the result was not what they hoped for is not a failure of the team. It is a failure of the assumptions that the field, as a whole, brought to the program.

The problem is not the science. The problem is an ecosystem that consistently asks the wrong question first — can we hit the target? — rather than the more important one: is this target sufficient in humans? And pays for that inversion at a late, expensive, and patient-facing stage.

The NEK7 data also points forward constructively. Direct NLRP3 inhibitors — which block the protein itself rather than an upstream activator — may have advantages in human systems precisely because they sit downstream of the redundant activation routes that bypass NEK7. Compounds like selnoflast and inzomelid, currently in clinical development, will be informative on exactly this point. The biology is telling us where to look next.

But the broader lesson is about first principles. Biology is a network, evolved over hundreds of millions of years to survive perturbation. Blocking one node rarely collapses it. Before we invest in any mechanism at scale, we owe it to patients — and to the basic integrity of the enterprise — to find out whether we are blocking a single point of failure, or merely rerouting traffic.

The most important question in drug discovery is not whether your compound works. It is whether your biology is asking for what your compound offers.

The Target Fallacy will not be the last story of this kind. But if we take it seriously — if we use it to restructure how and when we ask translational questions — it might be one of the last times we are surprised by it.

#DrugDiscovery #NLRP3 #TranslationalMedicine #TargetFallacy #PreclinicalModels #InflammasomeBiology #NEK7 #Novartis

TL;DR : My preliminary pharmacometric analysis of TOGETHER-PsO suggests weight loss-driven drug exposure recovery may be more important than direct synergy with putative anti-inflammatory effects of Zepbound . This is should be considered a curiosity-driven hypothesis, not a definitive finding as I do not have ‘insider access’.

Eli Lilly’s TOGETHER-PsO results were not unexpected: ixekizumab + tirzepatide achieved 40.6% PASI 100 clearance versus 29% with ixekizumab monotherapy. Many assumed the mechanism: “GLP-1 reduces inflammation, ixekizumab modulates immunity, they synergize beautifully.” I completed a causal mediation analysis* (10,000 Monte Carlo simulations) to decompose this efficacy gain (see Figure 1, the cover picture for the mechanistic framework). The finding: ~70–80% of the benefit appears to derive from weight loss, reversing obesity’s suppression of ixekizumab drug exposure. Direct GLP-1 anti-inflammatory effects? Likely secondary in this context.

The “Drug Sink” Insight

Here’s the mechanism I explored: Obesity increases how much “space” antibodies distribute into (higher Vd) and how fast they get cleared (altered CL). Result? The estimated ixekizumab exposure decreases by ~40% at a BMI of 40 compared with normal BMI (Figure 2). When Zepbound causes weight loss (14.1% mean in TOGETHER), BMI falls from 39.5 to ~32. The model estimated that this recovers ~30% of the lost drug exposure. Plus, adipose tissue stops pumping out inflammatory cytokines (IL-6, TNF-α, IL-23). That’s powerful synergy—but not the kind we usually discuss. This reframes obesity from “comorbidity” to “reversible pharmacokinetic defect.”

The Evidence (With Caveats)

The model calibration was excellent: The model predicted monotherapy at 29.1% versus the observed 29.0%—a good sign that the mechanistic framework is sound. The observed efficacy gain decomposes into mediated (weight loss-driven) and direct (GLP-1-driven) pathways (Figure 3), with sensitivity analyses robust to ±25% parameter variation (Figure 4).

Literature support: Hjort’s 2024 meta-analysis across 15+ trials showed each 5 BMI-unit increase associates with ~15–20% relative odds reduction in PASI 100—consistent with a PK-driven mechanism.

But here’s my key caveat: This analysis is based on published aggregate data from TOGETHER PSO only. No patient-level data. No serum ixekizumab concentrations. No adipose tissue biomarkers. The 7.6 percentage point gap between my model (estimated 33% PASI 100) and observed (40.6%) suggests unmeasured mechanisms—possibly a larger direct GLP-1/GIP effect than our conservative model estimated, or behavioral/temporal factors. This is should be considered a curiosity-driven hypothesis, and I hope Lilly will publish a full analysis confirming or refuting this.

Why It Matters

If my estimated weight loss → PK recovery pathway is primary (as suggested by cover page):

1. High-BMI patients failing biologics don’t need escalation—they need mechanism-informed weight loss intervention

2. Dosing timing is critical. Zepbound’s rapid weight loss (weeks 4–16) overlaps ixe’s PD buildup (weeks 12–24)—that’s the synergistic window

3. Obesity management becomes therapeutic, not cosmetic

The Challenge

What would prove/disprove this hypothesis? TOGETHER subgroup analysis by weight loss magnitude. My model predicts that the <5% weight-loss cohort within this trial shows minimal PASI benefit vs. monotherapy.

• If true → weight loss dominates (supporting Figure 1 framework).

• If false → direct GLP-1 effect is bigger than we estimated.

That data likely exists. #EliLilly, release it. Let science win.

What’s next?

I’ve built a rigorous framework with limited public data. I expect to be wrong in the details. Feel free to push back with evidence or counter-analysis.

1. What data do you have?

2. What’s your mechanistic intuition? Genuine pushback makes science better.

Source data for PK-efficacy assumptions : EPAR/SmPCs, publicly available Pop-PK data from regulatory submissions.

#EliLilly #NovoNordisk #AstraZeneca #Roche #Pfizer #Obesity #Diabetes #Psoriasis #Pharmacometrics #GLP1 #Dermatology #Thisis psoriasis # #UCB #EADV #Thisis PsA #ACR #Rheum #GRAPPA

TL;DR: AI drug discovery excels at prediction but stumbles at explanation- and in 2026, regulators are making that gap legally and ethically untenable. Developers now face a simple choice: explain your black boxes or don’t deploy them. Even if you are allowed to, should you?

Here’s the question keeping drug developers awake at night: Would you market a medicine that demonstrably helps patients—but whose mechanism of action you cannot fully explain?

That’s not hypothetical anymore. It’s 2026, and that question is no longer academic.

For the first time, pharmaceutical companies are running clinical trials on drugs that were designed by AI systems so complex, so layered with machine-learning inference, that even their inventors cannot trace the complete logic chain back to first principles. We can show you: This molecule works. Here’s the data. But we cannot always show you: Here’s exactly why.

This isn’t a failure of the technology. It’s a success that’s outpaced our ability to understand it. And that success is now colliding with something more fundamental: the ethics of deploying medicines we don’t fully comprehend.

Welcome to the Black Box Reckoning of 2026. There are three crises that matter right now, and they’re all variations on the same theme: AI’s predictive power has outrun human understanding—and the system is finally pushing back.

The “Phenotypic Trap”: When Cells Lie—And Patients Pay

Let’s start with the clearest cautionary tale: REC-994, Recursion’s superoxide scavenger for cerebral cavernous malformations (CCM), a rare disease affecting roughly 25,000 Americans. Most have no cure. Many face progressive neurological decline.

In February 2025, Recursion proudly announced that REC-994 had met its primary endpoint of safety and tolerability in a Phase 2 trial called SYCAMORE—and, more importantly, showed what the company called “promising trends” in lesion volume reduction. Fifty percent of patients on the 400 mg dose showed a reduction in total lesion volume compared to 28% on placebo after 12 months. (Recursion Pharmaceuticals, February 2025)

For patients with a devastating rare disease, this was hope. The AI-enabled phenotypic screening platform had done what it was built to do: scan millions of cellular images, find patterns that distinguish diseased from healthy states, and identify molecules that “rescued” those states. The AI worked. The science checked out. The patients might have a chance.

Except they didn’t.

By May 2025, when Recursion shared results from the long-term extension phase, the hope evaporated. Patients who had crossed over from placebo to the active drug showed no improvement. Worse, the patients who stayed on 400 mg the whole time didn’t sustain their initial benefit—their results became “indistinguishable from natural history,” in the clinical euphemism that means basically, the drug stopped working, and we’re back where we started. (GEN, May 2025)

This is the Phenotypic Trap in its most honest form: Recursion’s computer vision system optimized for a specific cellular readout—visual morphology that suggested disease reversal. The AI was right about what it was measuring. But the measurement itself didn’t predict what mattered: whether patients got better.

The epistemological failure here is subtle but critical. The AI learned to predict one thing (cellular appearance) while the researchers assumed it would predict another (human disease). No one set out to deceive. But the system was so sophisticated at pattern recognition that it found correlations that looked like causation—until humans tried the drug and reality intervened.

Recursion has since halted or deprioritized several pipeline programs. For the CCM patients in SYCAMORE, the news was simply: your trial drug doesn’t work. Time to wait for something else.

This is why the question matters: We trusted the AI because it made sense on a cellular level. But the cell isn’t the patient.

But Here’s the Other Side

It’s worth noting that REC-994’s failure doesn’t invalidate the phenotypic screening approach entirely. Recursion’s other AI-discovered program, REC-4881 (a MEK inhibitor for familial adenomatous polyposis), is showing what appear to be genuine clinical signals. December 2025 data showed a 43% median reduction in polyp burden after 12 weeks. (Recursion Pharmaceuticals, December 2025) The platform works sometimes—and when it does, it works well.

Additionally, traditional drug discovery has always involved failures in translation. Compounds that show promise in cell models routinely fail in humans; this is not unique to AI-driven approaches. The difference is partly one of scale: AI can generate and test more hypotheses faster, which paradoxically means both more failures and more successes, compressed into a shorter timeframe.

There’s also the question of mechanism itself. Many drugs used clinically for decades—aspirin, for example—had unknown or incompletely understood mechanisms of action when they were first approved. The assumption that we must understand why a drug works before deploying it is historically somewhat anachronistic. What matters most is that it works, and that it’s safe.

The FDA’s Credibility Demand: Protecting Patients by Demanding Answers

Here’s where the conversation gets uncomfortable for the tech side.

On January 6, 2025, the FDA released draft guidance titled “Considerations for the Use of Artificial Intelligence to Support Regulatory Decision-Making for Drug and Biological Products.” (FDA, January 2025) This was the first comprehensive regulatory framework addressing AI in drug development, and it said something that made the “move fast and break things” crowd grip their seats:

The agency is not accepting “the AI said so” as a valid reason to approve a medicine.

Instead, the FDA’s guidance outlined a “risk-based credibility framework” that requires sponsors to:

• Define their AI model’s “context of use” precisely

• Demonstrate that the model outputs actually predict clinical relevance

• Provide explainability artifacts showing how the AI arrived at a specific prediction

• Validate the model on independent data

That phrase—“Explainability Artifacts”—is doing a lot of work. (Federal Register, January 2025) It’s essentially asking: Can you show us your reasoning? Not in pseudocode. Not in a PowerPoint. In evidence that traces from the AI’s input to its clinical prediction.

Then, just last month (January 14, 2026), the FDA and EMA jointly published “Ten Guiding Principles of Good AI Practice in Drug Development.” (PharmaSource, January 2026) While these are currently non-prescriptive and voluntary, they’re the blueprint for future mandatory guidance. The principles emphasize “interpretability, explainability, and predictive performance” and demand “clear, essential information” about AI capabilities and limitations.

The message is unmistakable: Transparency is now table stakes. And it’s not a technicality—it’s a patient safety issue.

Deep learning models that even their creators can’t fully explain—like those used by Insilico Medicine or the newly merged Recursion-Exscientia entity—are now tagged as higher-risk by regulators. (Pinsent Masons, December 2025) That means slower approval timelines and more expensive validation requirements. But that slowdown is intentional. It’s there to protect patients from another REC-994 scenario: a drug that looks good in cells but fails in humans because we didn’t understand what we were actually treating.

The industry is already reacting. According to PharmaSource (December 2025), pharmaceutical companies are now “pulling back on fully autonomous AI design and re-inserting ‘Human-in-the-Loop’ steps just to ensure they have a human scientist who can sign off on liability.”

That sounds like bureaucracy. But it’s actually a return to something older: scientific accountability. Someone has to stake their reputation on the claim that they understand why this drug works. The human-in-the-loop isn’t a bottleneck. It’s a guardrail.

But Here’s the Other Side

Explainability requirements, while well-intentioned, face a fundamental challenge: some machine learning systems may be mathematically impossible to fully explain, even in principle. Deep learning models with billions of parameters operate in ways that resist human-comprehensible interpretation. Demanding explainability artifacts could create a false sense of certainty—a detailed explanation that sounds authoritative without actually being more predictive.

Additionally, pharmaceutical companies are already investing heavily in explainability research. Major biotech firms are hiring machine learning engineers specifically to develop interpretability tools (SHAP values, attention mechanisms, saliency maps). The burden isn’t unreasonable, and the FDA’s phased guidance approach (draft in January 2025, final guidance Q2 2026) gives sponsors time to adapt.

There’s also a practical precedent: regulatory agencies have long accepted empirical efficacy without complete mechanistic understanding. A drug that works is a drug that works, whether or not the mechanism is fully transparent. The risk here may be over-regulation—imposing standards of explainability that are scientifically unrealistic for complex AI systems, thereby slowing down the very innovation the agencies claim to support.

The EU AI Act “Liability Cliff”: August 2, 2026—Who Answers for an AI Mistake?

Now for the existential question: Who is responsible if an AI-designed drug harms patients in ways no one predicted?

On August 2, 2026—less than six months from today—the European Union’s AI Act enforcement provisions for high-risk AI systems become fully applicable. (EU AI Act Service Desk) This is not a guideline. This is law with teeth.

Here’s the problem: When a drug is designed by human chemists and a patient is harmed, the liability chain is clear. The pharmaceutical company knew what they were doing; they either understood the risk or failed to check. But when a drug is designed by an AI system that no one fully understands, the accountability becomes murky:

• Is the pharmaceutical company liable for deploying a black box?

• Is the software vendor liable for building an unexplainable system?

• Is the CRO liable for providing biased training data?

• Is the data broker liable for the “unbiased” patient cohort that wasn’t actually representative?

The EU AI Act answers simply: The pharmaceutical company is liable. Period. (Gardner Law)

Under the Act, high-risk AI systems (which include those supporting healthcare decisions) must be registered in an EU database, undergo conformity assessments, and be monitored for performance drift and bias. Companies must maintain detailed documentation showing how their AI system works and why it makes the decisions it makes.

If a company fails to meet these requirements by August 2, 2026, it faces fines up to €15 million or 3% of global turnover. If the violation involves a prohibited AI practice, it’s up to €35 million or 7%. (Software Improvement Group, January 2026)

But the real cost isn’t the fine. It’s the accountability itself.

Pharma companies are already shifting strategy. According to legal analysis from Pinsent Masons (December 2025), many firms are “de-risking their AI pipelines” by reintroducing governance steps that slow down discovery but create clear audit trails and human accountability. Human scientists are being asked to sign off on AI decisions—not because anyone thinks the human is smarter, but because someone needs to be responsible if things go wrong.

This creates what insiders call “governance debt”: the speed advantage that AI promised starts to vanish. But what’s lost in speed is gained in something older and more important: answerability to patients.

But Here’s the Other Side

The EU AI Act’s August 2, 2026 enforcement date, while tight, does provide a clear deadline that allows companies to plan and prepare. Some industry observers argue that regulatory clarity—even if strict—is preferable to ambiguity. Companies know what they need to do, and they have time to implement it.

Moreover, the Act’s framework allows for proportional risk assessment. Not all AI systems are classified as high-risk; systems used purely in research or early discovery phases are exempt. Companies can also reduce the AI risk classification of a drug by running it through traditional clinical validation pathways, which many are already doing. The impact may be less severe than initially feared.

Additionally, regulatory sandboxes—at least one per EU member state, operational by August 2, 2026—are designed specifically to help companies navigate novel AI-enabled drug development in a supervised environment. These sandboxes create space for innovation while maintaining oversight. And the Act does allow for graduated compliance: pharmaceutical companies deploying AI systems already in use before August 2026 have an extended transition period (until August 2, 2027) if they can demonstrate they’re taking steps toward compliance.

Finally, liability clarity may actually benefit the industry. Right now, responsibility for AI failures is murky; the Act clarifies that the pharmaceutical company is the responsible party. This removes ambiguity and could reduce expensive legal disputes downstream.

The Tension That Defines 2026

So here’s the Black Box dilemma, stripped bare:

LayerThe TensionThe 2026 RealityScientificCan we trust a drug if we don’t know its mechanism of action? Phase 3 results will answer this. REC-994 suggests: maybe not.Regulators want transparency; AI is inherently opaque.FDA and EMA are pushing toward “White Box” models (interpretable AI). Companies are hiring explainability engineers.LegalWho pays for an “AI mistake”?EU AI Act (Aug 2, 2026) forces companies to own the risk, even if they outsourced the AI.

The tension isn’t between Silicon Valley and Big Pharma anymore. It’s between predictive power and human understanding—and we’re in a standoff.

Recursion’s merger with Exscientia (closed November 2024) (Recursion Pharmaceuticals, December 2024) was supposed to create an unstoppable “full-stack” AI drug discovery powerhouse. And maybe it will. But REC-994’s stumble showed that stacking biology engines and chemistry engines doesn’t guarantee that the output will be biologically intelligible—or clinically relevant.

The old guard—the medicinal chemists who’ve been doing this for 30 years—are quietly vindicated. They never trusted black boxes. The young biotech founders are scrambling to make their black boxes look less black. And the regulators have drawn a line: Predict what you want, but explain what you find.

By August 2, 2026, that line becomes law in Europe.

The Epistemological Reckoning: Why Understanding Matters

But here’s the deeper concern, beneath all the regulatory and legal wrangling: What do we lose when drug developers deploy medicines they don’t understand?

Medicine isn’t just about outcomes. It’s about knowledge. When a physician prescribes a medicine, they carry a responsibility to understand—or at least to attempt understanding—why it works. That understanding is what allows them to:

• Adjust dosing when a patient isn’t responding

• Predict side effects based on mechanism

• Design better combination therapies

• Build on the scientific foundation for the next breakthrough

When the AI finds a drug that works but the mechanism is opaque, we’ve compressed the discovery timeline at the cost of scientific depth. We’ve optimized for outcomes at the expense of understanding.

REC-994 is the clearest example. The molecule did something in cells. But without knowing what or why, researchers couldn’t anticipate that it would fail in living humans. They were flying blind—not because they lacked data, but because the AI’s reasoning process was too complex to audit.

This is the risk of the black box: It can lead you to the right answer for the wrong reasons. And when it leads you to the wrong answer, you have no insight into why it failed.

The regulatory push toward explainability—the FDA’s guidance, the EMA’s principles, the EU AI Act’s accountability measures—isn’t just bureaucratic caution. It’s an attempt to preserve something essential: the connection between discovery and understanding.

A medicine that works is a victory. A medicine that works and that we understand is science. There’s a difference. And in 2026, that difference is becoming a dividing line.

What Comes Next

The Black Box Wars aren’t over. They’re escalating.

By August 2, 2026, pharmaceutical companies operating in Europe will face real legal consequences for deploying high-risk AI systems they can’t explain. The FDA’s final guidance on AI credibility is coming in Q2 2026—expected to be more prescriptive than the draft. And patients are becoming aware that the medicines they trust might have been designed by systems no one fully understands.

The industry will adapt. Some companies will invest heavily in explainability research, hiring teams to make their AI systems more interpretable. Others will pull back from fully autonomous discovery, reintroducing the human chemist as a check on the algorithm. A few will double down on black boxes and argue (perhaps successfully) that results matter more than reasoning.

But the core question remains, the one posed in the title: Would you market a drug you can’t explain?

For most of 2026, the answer from regulators, patients, and increasingly from developers themselves is: Only if you’re willing to stake your company’s reputation and liability on it. And only with complete documentation of your reasoning process.

That doesn’t mean the answer is always “no.” It means the answer is conditional—contingent on transparency, accountability, and proof that you’ve done your homework. The industry isn’t being asked to abandon AI drug discovery. It’s being asked to grow up about it. To acknowledge that speed and predictive power are valuable, but not if they come at the cost of scientific accountability.

That’s not a loss. That’s maturity.

I have been monitoring the CAR-T-for-autoimmunity field with growing concern. Not because the science is bad — some of the early lupus data from Erlangen is genuinely remarkable — but because I have seen this movie before. Multiple times. And it doesn’t end well for the patients who need these therapies most.

Let me explain.

The Numbers Don’t Support the Enthusiasm

As of February 2026, fewer than 400 patients worldwide have received CAR-T therapy for autoimmune conditions (my optimistic estimate). The published, peer-reviewed literature as of 2024 comprised approximately 80 patients across 24 studies (Kattamuri et al., Rheumatol Int, 2025). Eighty. That is not a typo.

Using the Rule of Three — a standard biostatistical method published in JAMA (Hanley & Lippman-Hand, 1983) — if zero serious adverse events are observed in 400 patients, the upper bound of the 95% confidence interval on the true adverse event rate is 0.75%. That means we cannot rule out that up to 3 in every 400 patients could experience a serious, potentially life-threatening complication that simply hasn’t shown up yet.

For context: oncology CAR-T required 34,400 treated patients before the FDA identified 22 cases of secondary T-cell lymphoma — a signal at 0.064% that triggered a boxed warning in January 2024 (Marks & Verdun, NEJM 2024). The autoimmune safety database is 86 times too small for equivalent pharmacovigilance. At current enrollment rates, matching that level of safety surveillance would take 15–20 years. Perhaps longer. Perhaps never.

We are flying blind.

Too Many Products, No Market Consolidation — We’ve Seen This Before

Currently, there are more than 100 registered clinical trials of CAR-T for autoimmune diseases (Frontiers systematic review, 2025). Over a dozen companies are developing competing constructs targeting CD19, BCMA, or dual targets, each with slightly different manufacturing processes, conditioning regimens, and vector designs. The fragmentation is staggering.

History tells us what happens next.

In the early 1980s, there were dozens of personal computer manufacturers — Commodore, Atari, Tandy, Osborne, Kaypro, and IBM clones by the hundreds. The market consolidated brutally. Microsoft and Intel survived. Most did not. The technology was transformative, but the companies were largely destroyed.

Smartphones followed the same arc. Remember Palm, BlackBerry, Windows Phone, Symbian, WebOS? By 2013, Apple and Android accounted for more than 95% of the market. The underlying technology thrived. The business landscape was a graveyard.

Henry Ford didn’t invent the automobile. He survived the consolidation of more than 250 American car companies between 1900 and 1920. By 1930, three companies controlled 75% of the U.S. market.

The pattern is iron-clad: transformative technology → irrational proliferation → brutal consolidation → a few winners, many casualties. CAR-T for autoimmunity is deep in phase two. The question is not whether consolidation will happen, but how many patients will be mid-treatment when it does.

The “Dumb Capital” Problem

Between 2021 and 2025, cell therapy startups collectively raised more than $15 billion through venture capital and public offerings. Much of this money was not invested by groups with deep immunology expertise or realistic timelines. It was speculative capital chasing the next platform story — the same capital that inflated and then abandoned gene therapy, microbiome therapeutics, and digital health before that.

When this bubble pops — and the Rule of Three analysis above suggests the safety reckoning is a matter of when, not if— the collateral damage will extend far beyond the companies that fail. The entire cell therapy ecosystem will suffer. Investors will retreat. Regulatory scrutiny will intensify. And the patients with refractory lupus, systemic sclerosis, or myositis who desperately need these treatments will find themselves stranded.

We observed precisely this in the field of gene therapy following the death of Jesse Gelsinger in 1999. It took nearly 15 years for clinical gene therapy to recover. Fifteen years of patients waiting because one preventable catastrophe poisoned the well.

“But What About the Alternatives?”

Some colleagues point to emerging competitors—bispecific antibodies, CD19-targeting monoclonals, CAR-NK cells, and in vivo CAR approaches—as evidence that the field will be fine even if autologous CAR-T cells stumble. They argue these alternatives are cheaper, more scalable, and avoid the manufacturing bottleneck.

They are not wrong about the short-term advantages. Bispecific T-cell engagers don’t require leukapheresis, clean-room manufacturing, or a two-week wait. They cost a fraction of the cost of CAR-T. For a budget-constrained rheumatology market, that is a compelling pitch.

But here is what the “alternatives will save us” argument misses: if a safety signal emerges in CAR-T — particularly one involving T-cell malignancy or delayed autoimmune flare — it will not be contained to CAR-T alone. The regulatory and public perception fallout will splash across every therapy that involves engineered immune cell activation. The FDA will not distinguish between CAR-T and CAR-NK when a CNN headline reads “Cancer Therapy Turns Deadly in Lupus Patients.” Fair or not, that is how it works.

My prediction from two years ago still holds: without rigorous, adequately powered safety monitoring — a proper pharmacovigilance infrastructure built before the crisis, not after — this field is heading for a correction that will hurt everyone. Patients. Investors. Developers. The science itself.

So, What Should We Do?

I have specific, actionable recommendations. A unified safety registry. Bayesian adaptive monitoring frameworks. Minimum reporting standards. A realistic conversation about what 400 patients can and cannot tell us. And a framework for how investors should evaluate safety risk—not just efficacy upside—in their due diligence.

But that is Part 2.

Coming soon.

Eswar Krishnan, MD | www.drugdevelop.com

Data sources: Hanley & Lippman-Hand, JAMA 1983; Kattamuri et al., Rheumatol Int 2025; Mueller et al., ASH 2024; Marks & Verdun, NEJM Jan 2024.

#CARTtherapy #CellTherapy #Autoimmune #Pharmacovigilance #DrugSafety #Biostatistics #ClinicalTrials #Rheumatology #Lupus #DrugDevelopment #BiotechInvesting #VentureCapital

There’s a little theatre that plays out on Substack, LinkedIn and other social media. A tidy post appears—hint: correct use of oxford commas—and almost on cue someone comments: “This sounds like AI.” The tone implies a sting operation: the grammar is too clean, the paragraphs too disciplined, the culprit surely silicon.

Since when did coherent prose become suspicious behaviour?

Somewhere along the way, “authentic” got quietly redefined as “unfiltered.”

Typos became personality. Rambling and raging turned into sincerity.

We began treating rough edges as proof of life.

Is the only honest song the one performed out of tune?

It’s charming in its way, like old cassette hiss (remember the SONY Walkman?).

But must every post sound like a live rehearsal?

A rude question: Are we defending authenticity—or defending the old inefficiencies that made us feel special?For years, first drafts were expensive in time and ego. That effort conferred a little halo: “I suffered for these sentences; kindly clap.” Now, with a single nudge, a language model can take your thought from muddy to presentable. The halo slips, and we call the polish “fake.”

But authenticity was never about the presence or absence of mistakes. It was about the presence of self. A post feels human when there’s a recognisable mind behind it: a choice made, a risk taken, a detail only you could supply. Whether the comma arrives by your hand or with a bit of algorithmic help is frankly none of the reader’s concern. They want clarity; they care who’s speaking; they don’t mind who wrote the script.

Do you think President Obama wrote all the soaring speeches he gave?

“But AI posts feel robotic,” people say. Sometimes they do. Then again, many human posts feel robotic too—parades of well-meaning platitudes.

The real offence isn’t the polish; it’s emptiness.

When a post has no friction, no specificity, no argument, it is humbug. Whether typed by a monk with a quill or a model with an NVDA Hopper GPU.

This is hardly the first time presentation tools made us anxious about “realness.” We fretted at print, then telephones, then radio, then the web, and yet somehow the human voice kept smuggling itself through.

A small confession: we are sentimental about creative suffering. We love the myth of the writer battling the page like a wrestler in a dusty akhada. It flatters the rest of us when our own drafts look scruffy; at least we are “real.”

So what is the bottomline? AI is here to stay. DEAL WITH IT !.

If you’re still uneasy, try this framing: AI doesn’t diminish writing; it exposes it.

AI writing does all of us a great service by leveling the playing field. By removing the role of writing aesthetics, it allows concepts to be judged for their truth.

Food to be judged by smell, taste, and nutritional value- not by visual presentation

It brings the idea to the front and asks, politely but firmly, “Is there anything here?” If the answer is no, the gloss won’t rescue it. If the answer is yes, the gloss helps it land. That is uncomfortable for anyone who relies on visible effort as a substitute for visible thought. It is also, modestly put, progress.

If you need a rule of thumb, make it this: Soul belongs to the writer; shine can come from anywhere. If you bring only shine, the post glides across the feed and vanishes. If you bring only soul and refuse any shine, the post may trip over its own shoelaces before it reaches a reader. But if you bring a real idea and allow a little polish to help it arrive on time, full marks. The reader is served, the conversation moves, and the platform becomes slightly less of a fish market.

A final, mildly naughty question to leave hanging in the air: if you can have an authentic thought and a clear sentence, why insist they travel in separate compartments? The muse does not award extra credit for split ends. She cares that you brought something only you could bring.

So yes—boo the emptiness. Applaud the clarity. And if the clarity arrives with a touch of robot polish, please adjust. The human inside the words is still the one wearing the shirt.

Cardiovascular diseases challenge global health. Traditional risk factors are key, but cardioimmunology shows the immune system's complex role in these diseases, especially in atherosclerosis, now seen as a chronic inflammatory condition. Findings like structured immune cells, termed tertiary lymphoid organs (TLOs), in atherosclerotic plaques, indicate the immune system's active role in disease progression and potential coronary events triggers.

Tertiary Lymphoid Organs (TLOs) in Plaques

The discovery of TLOs, specifically referred to as Plaque Tertiary Lymphoid Organs (PTLOs) in the context of atherosclerosis, offers clear evidence of local adaptive immune responses occurring directly within the arterial wall. These structures are not encapsulated but show some level of organization, containing collections of lymphoid cells. Their formation is typically linked to chronic inflammation. The identification of PTLOs in human carotid atherosclerotic plaques using methods like spatially resolved single-cell transcriptome mapping, H&E staining, and immunofluorescence confirms their presence and classification as TLOs within these lesions.

Cellular Composition of PTLOs

PTLOs are marked by a notable accumulation of lymphoid cells, with prominent populations of T cells and B cells. Compared to other areas within the plaque, PTLOs tend to have a higher number of cells from the B lineage, including plasma cells. Analysis of human coronary atherosclerotic plaques shows a substantial proportion of alpha-beta T cells. Many of these T cells found in plaques are not naive but are antigen-experienced memory T cells, primed to respond quickly upon re-exposure to an antigen.

The Critical Role of T Cells

T cells found in plaques show signs of clonal expansion, suggesting they have been activated by specific antigens present in the plaque environment. Single-cell RNA sequencing (scRNA-seq) analysis of human atherosclerotic plaques has revealed various T cell subsets. Importantly, studies have found that plaque T cells, especially CD8 T cells, can be specific for antigens from common viruses like Influenza (Flu), Cytomegalovirus (CMV), and Epstein-Barr virus (EBV). This finding implies that ongoing or repeated viral infections might help maintain persistent inflammation by activating these resident T cells.

Furthermore, the concept of molecular mimicry comes into play, where viral structures are similar enough to self-proteins that an immune response initially targeting the virus mistakenly attacks the body's own tissues. Experimental data support this, showing that T cells isolated from plaques can react to both viral and self-derived targets, suggesting that long-term viral exposure could initiate or sustain an autoimmune element in atherosclerosis.

Beyond promoting inflammation, plaque T cells also influence the physical characteristics of the plaque. In more advanced atherosclerotic lesions, there is an increased presence of T cells that produce Amphiregulin (AREG), a cytokine known for its pro-fibrotic effects. AREG from plaque T cells can affect smooth muscle cells (SMCs) and macrophages, potentially contributing to the development of fibrosis and structural changes seen in later stages of plaques. Certain T cell types, like CD8 Tem2 cells, display characteristics linked to inflammation and cell killing in advanced disease.

B Cells and Local Antibody Production

B cells are essential components of plaque TLOs and the atherosclerotic lesion as a whole. While their role is complex, their structured arrangement in PTLOs is particularly significant. PTLOs can contain B cells that express markers of germinal centers, indicating local processes of maturation, proliferation, and differentiation into plasma cells capable of producing antibodies within the arterial wall, mirroring processes observed in some autoimmune diseases. Specific chemokine signaling, such as the CXCL13-CXCR5 pathway and others like CXCL12, CCL19, and CCL21, are associated with the organization of B cells and the formation of TLOs.

The local generation of antibodies, particularly IgG antibodies, by plasma cells found in or moving into PTLOs, seems to have important consequences. These locally produced IgG antibodies can interact with Fcγ receptors (FcγRs) present on immune cells like macrophages. Specifically, binding to FCGR3A (CD16), found on macrophages, can mediate the effects of these antibodies. Studies have shown a correlation between the expression of IGHG4 in plasma cells and FCGR3A in macrophages within atherosclerotic plaques. High levels of macrophage FCGR3A have been observed in symptomatic carotid plaques, suggesting a possible link to plaque instability. IgG antibodies can form immune complexes with oxidized LDL (oxLDL), which can then trigger inflammatory responses in macrophages. Therefore, the differentiation of B cells and the local production of IgG within PTLOs likely play a role in plaque instability by boosting macrophage activation.

Well-structured artery tertiary lymphoid organs (ATLOs) arise adjacent to advanced atherosclerotic plaques in the abdominal aorta of aged Apoe-/-mice.

1. Cellularity, structures, and the territoriality of ATLO neogenesis indicate comprehensive T-and B-cell responses toward unknown arterial wall-derived autoantigens

2. Advanced ATLO stages are characterized by separate T-cell areas, activated B-cell follicles, and plasma cell niches in the periphery.

3. Autoantigen presentation is indicated by the presence of follicular dendritic cells (FDCs) in activated germinal centers; the abundance of cDCs and monocyte-derived DCs (mDCs) in T-cell areas;

4. B-cell affinity maturation is indicated by multiple centroblasts and their progeny in B-cell follicles; a balance between pro-and anti-inflammatory and T cells is indicated by multiple effector T cells and regulatory T cells (Tregs);

5. newly formed conduits may maintain chemokine gradients and possibly guide autoantigen diffusion from the diseased arterial wall toward ATLO antigen-presenting cells. HEV indicates high endothelial venule; and SMC, smooth muscle cell.

   ---

TLOs, Lymphangiogenesis, and Plaque Instability

The formation and function of TLOs are also connected to lymphangiogenesis, the growth of new lymphatic vessels. Research in transplant arteriosclerosis, a related vascular condition, has demonstrated a link between the growth of lymphatic vessels and the formation of TLOs, with early interventions targeting lymphangiogenesis showing potential long-term benefits. In atherosclerosis, genes related to lymphangiogenic chemokines are associated with the development of PTLOs, pointing to an interaction between local lymphatic networks, the movement of immune cells, and the organization of TLOs. The density of CD31+ microvessels is higher in areas with PTLOs compared to areas without. PTLOs themselves have been shown to be associated with symptomatic carotid atherosclerosis.

Implications for Therapeutics

The presence of TLOs in atherosclerotic plaques fundamentally alters our understanding, highlighting atherosclerosis as a disease partly driven by organized, local adaptive immune responses. These structures act as central points for interactions and activation among immune cells, and potentially for the local production of autoantibodies. Understanding the detailed roles of T cells in inflammation and fibrosis, and the complex activities of B cells, particularly the pro-inflammatory capacity of locally produced IgG antibodies binding to FcγRs on macrophages, is essential for grasping plaque progression and the associated risk of clinical events.

The in-depth knowledge of PTLOs and the specific immune cell populations within them provides promising avenues for identifying new targets for treatment. The key challenge is to develop therapies that can specifically target these local immune structures and pathways without broadly suppressing the immune system. The rapid progress in cardioimmunology, supported by technologies like single-cell and spatial transcriptomics, continues to reveal the extensive involvement of the immune system in atherosclerosis, paving the way for more targeted and effective strategies to prevent and treat this widespread disease.

#Atherosclerosis, #Cardioimmunology, #Immunology, #TertiaryLymphoidOrgans, #VascularBiology, #Inflammation, #TCells, #BCells, #Macrophages, #SingleCellRNAseq, #SpatialTranscriptomics, #CardiovascularDisease

This is the inconvenient truth of residual cardiovascular risk—the persistent hazard that remains despite optimal lipid-lowering therapy. In CANTOS, even with well-controlled LDL cholesterol, cumulative major cardiovascular event rates exceeded 20% at five years.

One in five patients. On guideline-directed therapy.

The culprit? Inflammation. The therapeutic response? Target IL-1β, measure CRP, declare victory when hsCRP drops. That was the playbook from CANTOS through COLCOT.

IVORY just tore it up.

The CRP-Centric Era: What It Got Right and Wrong

The intellectual groundwork spans decades. Russell Ross’s “response to injury” hypothesis. Peter Libby’s meticulous characterization of inflammatory cell infiltration in atherosclerotic lesions. Paul Ridker’s epidemiological observations linking high-sensitivity C-reactive protein to cardiovascular events independent of LDL.

CANTOS delivered the proof. Canakinumab—a monoclonal antibody targeting interleukin-1β—reduced major adverse cardiovascular events by 15% in post-MI patients with hsCRP ≥2 mg/L. No change in lipids. The inflammatory hypothesis was validated.

The CRP responder analysis in CANTOS was striking: patients achieving on-treatment hsCRP <2 mg/L had a 25% reduction in MACE. Non-responders: 5%. CRP became both biomarker and therapeutic target.

But this created a blind spot. The field conflated inflammation reduction with CRP reduction. If hsCRP didn’t drop, was the intervention working?

IVORY answers: yes.

IVORY: Efficacy Without CRP Movement

The IVORY trial, published this month in Nature Medicine, takes a fundamentally different approach. Rather than blunting the IL-1β/IL-6/CRP axis, investigators amplified regulatory T cells—the adaptive immune system’s endogenous peacekeepers.

Low-dose interleukin-2 selectively expands Treg populations. At high doses, IL-2 activates cytotoxic T cells—hence its use in metastatic melanoma. But Tregs are enriched for the high-affinity trimeric IL-2 receptor (CD25/IL-2Rα), making them exquisitely sensitive to low concentrations. This pharmacological window enables precision immunomodulation.

The trial enrolled 60 ACS patients with hsCRP >2 mg/L, randomized to low-dose IL-2 (1.5 × 10⁶ IU) or placebo for 8 weeks. Arterial inflammation was quantified by ¹⁸F-FDG PET-CT.

Here’s what happened:

Arterial inflammation dropped 7.7% versus placebo (p=0.015).

In actively inflamed segments: 8.3% reduction (p=0.009).

In the most diseased segments: 9.3% reduction (p=0.010).

And hsCRP?

No difference between groups.

This is the paradigm shift: low-dose IL-2 reduced arterial inflammation through a CRP-independent mechanism. The inflammasome pathway and Treg-mediated immunoregulation are orthogonal.

The Immunological Mechanism: Tregs Reshape the Inflammatory Milieu

Regulatory T cells exert immunosuppression through multiple pathways: direct cell-cell contact via CTLA-4 and LAG-3, IL-2 consumption that starves effector T cells, secretion of IL-10 and TGF-β, and promotion of tolerogenic dendritic cell phenotypes.

In ACS patients, Tregs are depleted and dysfunctional. Mor and colleagues demonstrated reduced CD4+CD25+ Treg numbers in acute coronary syndromes back in 2006. Flego’s group later showed impaired CREB activation blocks Treg differentiation in NSTEMI patients.

IVORY reversed this deficit.

Treg cells increased 57% after the induction phase. During maintenance dosing, levels remained 34% above placebo. The adaptive immune system was reset.

But the immunological story extends beyond Treg expansion:

• TH1 cells decreased — these IFN-γ producers drive plaque inflammation

• T follicular helper cells decreased — reducing germinal center B cell activation

• CD8+ cytotoxic T cells decreased — the effector arm was suppressed

The net effect: a comprehensive shift from pro-atherogenic to regulatory immune phenotype.

Importantly, in preclinical models, Treg expansion doesn’t just reduce inflammation—it promotes plaque stability. Meng et al. showed Tregs prevent plaque disruption in ApoE-knockout mice. Weirather’s group demonstrated Tregs improve post-MI healing by modulating monocyte/macrophage differentiation.

Regulation. Not suppression.

Effect Size in Context: Matching Statins without Touching Lipids

How meaningful is a 7-9% reduction in arterial inflammation?

High-intensity statins reduce arterial inflammation by ~10% versus low-intensity therapy. PCSK9 inhibition with alirocumab: ~8%. Low-dose IL-2 achieved comparable anti-inflammatory efficacy—on top of optimal medical therapy including high-dose statins in 98% of participants.

This is the key point. IVORY didn’t replace standard of care. It added orthogonal benefit. The anti-inflammatory effect was independent of lipid lowering, independent of CRP reduction, and additive to statins.

The dose-response relationship was also instructive: higher baseline inflammation predicted greater treatment effect. The sickest patients benefited most.

Safety: Why Immunomodulation Beats Immunosuppression

Anti-inflammatory therapy walks a tightrope. CANTOS demonstrated that potent IL-1β inhibition increases fatal infections. Colchicine carries gastrointestinal toxicity that limits adherence. And the recent CLEAR trial showed colchicine was ineffective in acute MI altogether.

Low-dose IL-2 threaded the needle.

Infection rates: 9% IL-2 versus 13% placebo. No difference.

The only significant adverse effect: injection site reactions in 91% of IL-2 recipients—mild, transient, resolved within 48 hours.

The cumulative IL-2 dose in IVORY was ~1.3% of oncology dosing. Treg expansion preserves host defense while attenuating pathological inflammation. This is pharmacological precision.

The 2-year follow-up from IVORY-FINALE strengthens the signal:

Zero MACE in the IL-2 arm.

Three patients in placebo experienced major adverse cardiovascular events.

Small numbers. Unambiguous direction.

Drug Development: The Adaptive Immunity Frontier Opens

The innate immune system—IL-1β, IL-6, NLRP3 inflammasome—has dominated cardiovascular drug development. CANTOS targeted IL-1β. Ziltivekimab targets IL-6 (ZEUS-CV ongoing). Colchicine acts through multiple innate pathways.

IVORY cracks open the adaptive immunity door.

Several development paths emerge:

Optimized IL-2 formulations. Engineering IL-2 muteins with enhanced Treg selectivity could improve the therapeutic index. Neoleukin-2/15 and similar next-generation constructs are in early development.

Combination strategies. Low-dose IL-2 did not reduce hsCRP, suggesting orthogonal mechanisms to IL-6 pathway inhibition. Combining Treg expansion with ziltivekimab could provide synergistic benefit—hitting both innate and adaptive arms.

Oral Treg-expanding agents. Injectable biologics face adherence challenges. Small molecules that selectively activate Treg-specific pathways represent the holy grail.

Antigen-specific tolerization. If atherosclerosis-associated autoantigens (oxidized LDL, ApoB100 peptides, heat shock proteins) can be definitively identified, tolerogenic vaccines could induce plaque-specific Tregs. Early preclinical work from Nilsson, Mallat, and others suggests feasibility.

The Bottom Line

CRP told us inflammation mattered.

Tregs show us how to control it.

IVORY demonstrates that harnessing adaptive immunity—specifically CD4+CD25+FoxP3+ regulatory T cells—safely reduces arterial inflammation in the highest-risk ACS patients. The mechanism is CRP-independent. The effect size matches statins. The safety profile outperforms canakinumab and colchicine.

This is not about replacing lipid management. Residual risk is multifactorial. But for ACS patients who remain inflamed despite guideline-directed therapy, the Treg pathway offers a genuinely new therapeutic axis.

The paradigm is shifting.

Forget CRP. Think Tregs.

The IVORY trial (NCT04241601) was published in Nature Medicine on January 8, 2026. The IVORY-FINALE study (NCT06427694) continues follow-up for cardiovascular outcomes through 5 years.

We need to talk.

It’s not that I don’t appreciate everything you’ve done. The late nights formatting tables. The delicate diplomacy of managing eleven co-authors who all believe their edits are non-negotiable. The way you transformed raw clinical data into documents that actually made regulators nod instead of frown.

You’ve been good to this industry. You’ve been good to me.

But something has changed. And I think you feel it too.

The Numbers Tell a Paradox

Let’s start with the facts that seem contradictory. The medical writing market reached approximately $4.3 billion in 2024 and is projected to grow to $10-12 billion by 2032, expanding at roughly 10-11% annually. At first glance, this seems impossible in an era when AI can draft documents in minutes.

But look closer. The FDA cleared 50 new drugs in 2024 and anticipates up to 70 approvals in 2025. Clinical trials are becoming demonstrably more complex—a Boston Consulting Group analysis of over 16,000 trials confirmed this trend. Cell and gene therapies demand extensive chemistry, manufacturing, and controls sections. Personalized medicine requires adaptive trial protocols that stratify by biomarker. The documentation burden is exploding.

Meanwhile, 83% of life sciences companies reported difficulty filling medical writing roles in 2024. Forecasts predict a 35% talent deficit by 2030.

The algorithm isn’t eliminating jobs. It’s filling a capacity gap that the industry cannot close with humans alone.

The Old Rituals

Remember when writing a clinical study report meant weeks of hunting for source documents across shared drives with folder names like “FINAL_v3_REVISED_USE THIS ONE”? Remember when the statistical outputs would arrive at 4:47 PM on a Friday, and you’d smile through the pain because that’s just how it was?

Remember the track changes? The glorious, maddening track changes. A single document passing through regulatory, medical, legal, and that one executive who hadn’t read anything but suddenly had thoughts about your executive summary. Each round a negotiation. Each comment a tiny battle.

These were your rituals. They were exhausting and inefficient and sometimes absurd. But they were yours.

In June 2024, Certara launched CoAuthor, a regulatory writing platform combining generative AI with document templates. The following month, Cognizant partnered with Yseop to accelerate scientific documentation using AI. The industry is moving fast.

The FDA Is Now Using AI to Read Your Work

Here’s something that should get your attention. On June 2, 2025, the FDA launched Elsa, a large language model built on Anthropic’s Claude, deployed agency-wide to help scientific reviewers work more efficiently.

Commissioner Marty Makary announced that tasks previously taking reviewers two to three days now take six minutes. The agency is using Elsa to accelerate clinical protocol reviews, summarize adverse events for safety assessments, and identify high-priority inspection targets. Makary’s stated vision includes “rapid or instant reviews” of drug applications.

But early reports reveal limitations. CNN and STAT News reported that Elsa has been producing fabricated citations and misrepresenting research—the same hallucination problems that plague all LLMs. FDA staff told reporters the tool cannot yet assist with formal review work because it lacks access to many relevant documents, including industry submissions.

What This Means for Medical Writers

Think about what’s happening. The regulator reviewing your submission is now using AI to read and summarize your work. If your documents are optimized for human readers but not for AI parsing, you may face unintended friction. If the FDA’s AI misinterprets your carefully crafted safety narrative, who catches the error?

This cuts both ways. On the one hand, AI-readable documents may be reviewed faster. On the other hand, if both sides are using AI—you to write, FDA to review—the human expertise that catches subtle errors becomes more valuable, not less.

The writers who thrive will be those who understand how these systems work on both ends of the regulatory submission.

What the Algorithm Cannot Do

A 2025 study by Boston Consulting Group, published in Clinical Trials, evaluated GPT-4’s ability to write sections of clinical trial protocols. The findings reveal the precise contours of AI’s limitations.

For content relevance and appropriate use of medical terminology, GPT-4 scored above 80% and 99%, respectively. The output looks professional. The jargon is correct.

But for clinical thinking and logic—whether the AI’s recommendations actually followed regulatory guidance—the off-the-shelf model scored approximately 40%. The AI confidently suggested excluding HIV patients from tuberculosis trials, directly contradicting FDA guidance that explicitly requires their inclusion.

When enhanced with retrieval-augmented generation (providing the AI access to current regulatory documents), that clinical logic score rose to approximately 80%. Still not perfect. Still requiring human oversight.

The AI cannot read the room during a sponsor meeting. It doesn’t know why the medical monitor just sighed.

It has never had to explain to a client, gently but firmly, that removing 40 pages will also remove the actual evidence.

The Uncomfortable Part

I won’t pretend this transition is painless. Companies that once needed ten writers may soon need three, armed with AI tools that can produce first drafts in seconds rather than days. BCG reports that AI-assisted writing reduces end-to-end document creation time by 25-50%, depending on the document type.

Real people with specialized skills and years of training will face real disruption.

This is not something to celebrate. It’s something to acknowledge honestly, even as we talk about evolution and opportunity.

My Prediction

Here’s what I believe will happen.

The FDA’s use of Elsa will accelerate, despite its current limitations. Other regulators—EMA, PMDA, NMPA—will follow within 18-24 months. Documents optimized for AI parsing will have an advantage.

This doesn’t hurt medical writers. It changes what they do.

When both submission and review are AI-assisted, the humans on each side become quality controllers, strategic advisors, and exception handlers. The medical writer ensures the AI-generated draft doesn’t contain the tuberculosis error. The FDA reviewer becomes the person who catches what Elsa missed.

The premium shifts from drafting speed to judgment, regulatory strategy, and the ability to verify AI outputs against evolving guidance.

The Reinvention

The medical writer of tomorrow is not a document generator. The algorithm handles that now. The medical writer of tomorrow is a document architect. A curator. A skeptic who reviews AI output with the same rigor once applied to junior writers.

You become the one who knows why certain information matters to a reviewer in Brussels versus one in Silver Spring. The one who understands that regulatory writing is not just about clarity but about strategy. The one who catches the hallucination before it becomes a 483 observation—or before the FDA’s AI misreads your intent.

The talent shortage tells us something important: the industry still needs human expertise. It just needs that expertise applied differently.

The Case for Optimism

Here’s the math that matters.

The volume of regulatory documentation is growing faster than AI can reduce costs. The talent gap is widening. The FDA itself is deploying AI because it cannot review submissions fast enough with humans alone.

Documents will be produced faster, which means treatments may reach patients sooner. Writers freed from formatting drudgery can focus on higher-value work. Small biotechs without massive budgets can compete with pharma giants on documentation quality. The projected 35% talent deficit by 2030 might be addressed through human-AI collaboration rather than through pure headcount growth.

The craft doesn’t disappear. It concentrates. It becomes more strategic, more specialized, more human in the ways that matter.

A Closing Thought

Dear Medical Writing, you are not dying. You are molting.

The exoskeleton of manual assembly and endless revision cycles is cracking. What emerges will be leaner, faster, and, frankly, less tedious.

The algorithm can do part of what you do. But 40% accuracy on clinical logic isn’t going to pass muster with the FDA—even if the FDA is also using AI that makes mistakes.

The rest, the hard part, the part that requires judgment and nuance and the occasional well-timed pushback, that remains yours.

So don’t mourn the old ways too long. They were never the point.

The point was always the work: translating science into decisions that help patients. That work continues. You just have a new collaborator now.

One that doesn’t need coffee breaks. But also one that recommends excluding HIV patients from tuberculosis trials.

With respect and only a little anxiety,

Eswar Krishnan, A Fellow Traveler in Drug Development

Most biosimilar companies are playing a losing game.

They spend 8-10 years and $100-200M developing a product that immediately enters a race to the bottom. They compete on 5% margin differences. Meanwhile, originators maintain their infusion center relationships and patient loyalty programs. The biosimilar manufacturers fight over scraps.

Read more

As multi‑specific antibodies move from concept to clinic, the ecosystem around trial design, PD readouts, and safety surveillance is being rewritten. A useful case s…

Read more

The pharmaceutical industry runs on a crushing paradox. We possess petabytes of biological data—genomic sequences, real world evidence, clinical telemetry—yet we starve for actionable insight. For the Chief Medical Officer or Chief Scientific Officer, the bottleneck is no longer the generation of data. It is the latency of decision.

Read more

My experience suggests that the true value of a pIND lies in mitig…

Read more

CAR-T Illustration. Courtesy New Scientist

A recent head-to-head study assessed the comparative effectiveness of different B-cell depletion strategies in patients with autoimmune diseases, specifically comparing protein-based therapies like obinutuzumab (OBI), blinatumomab (BLI), and rituximab (RTX) with cell-based CD19 chimeric antigen receptor (CAR) T-cell therapy. Researchers analyzed lymph node biopsies to assess the extent of B-cell depletion and changes in follicular architecture. They documented that CAR-T was most effective in depleting B cells and disrupting follicular architecture, and concluded that these data indicate potential superiority of CAR-T cell therapies.

Share