The Tyranny of Simple: Seven Centuries of Wrong
AlphaFold has more trainable parameters than there are synaptic connections in a fruit fly’s brain. It also solved a 50-year-old protein structure problem in a single year. Every simpler approach that came before it failed.
That is not a coincidence. It may be a warning.
A paper just published in PNAS — “Is Ockham’s Razor Losing Its Edge?” by Dubova and colleagues — makes a careful, well-sourced case that science’s oldest methodological shortcut is now working against us in some of the places that matter most. The argument isn’t that simple models are bad. It’s that treating parsimony as a universal principle, rather than a context-dependent tool, leads to measurable scientific error.
The heuristic and where it breaks
William of Ockham’s principle has survived 700 years because it works in physics. Fewer assumptions usually mean fewer places to go wrong. In clean experimental systems, this holds.
Biology is not a clean experimental system.
The PNAS paper distinguishes two forms of parsimony: parsimony by constraint (models that make narrow, specific predictions) and parsimony by components (models with fewer variables). These often conflict. And in living systems — where causal chains are long, nonlinear, and context-sensitive — both forms can produce models that are technically elegant and factually wrong.
The authors describe a striking example from neuroimaging. Simple models applied to live brain scans consistently inferred oscillatory patterns — rhythmic back-and-forth activity — that simply weren’t there. The brain wasn’t oscillating. The model was. The simplifying assumptions imposed a false pattern onto data that told a different story.
The double descent problem
The conventional statistical wisdom: more parameters relative to data means more overfitting. Fit a complex model with too few data points and you’ve memorized noise.
Recent work has overturned this. Researchers found that prediction error follows a U-shaped curve as model complexity increases — it rises, peaks, then falls again as parameters grow large. This “double descent” means highly overparameterized models, trained with methods like gradient descent, can achieve both low bias and low variance. The theoretical case for reflexive parsimony — keep it simple so you don’t overfit — no longer holds unconditionally.
Dubova et al. document a study on moral judgment that makes the implication concrete. Researchers built a large machine-learning model trained on 40 million moral decisions. The model was opaque and complex. Once it identified structure in the data, that structure was distilled into a simpler psychological theory — one that would have been invisible to a researcher who started with a parsimonious model from the beginning. The complexity came first. The clarity followed.
What this means for drug development
Most disease models in clinical development start simple by design. Single target, single mechanism, linear causal chain. This is not just a scientific choice — it’s a regulatory and communicative one. Simple models are easier to test, explain, and fund.
Biological disease processes are not simple. Autoimmune disease, cancer, neurodegeneration — these involve gene networks, environmental context, cell states, and feedback loops that no parsimonious model captures well. When a drug fails in Phase III, the question worth asking isn’t always “did the molecule work?” It’s often: “was the disease model right in the first place?”
The PNAS paper suggests a different workflow: start with the most complex model the data supports, find structure, then distill it. Use ensemble approaches rather than forcing a single explanatory account. Accept that the right model for discovery may not be the right model for communication — and treat those as separate problems.
This matches what is already happening at the edges of the field. AlphaFold changed structural biology not by simplifying the problem but by refusing to. Climate science now uses multi-model ensembles rather than a single parsimonious forecast. Quantum biology — a topic I wrote about last week — finds that quantum effects in enzyme function and energy transfer cannot be adequately described by classical approximations, no matter how clean those approximations look on paper.
The line worth drawing
Occam’s razor is not wrong. It is misused.
Applied to physics, instrument calibration, or clinical communication, it works well. Applied as a prior assumption about how biological systems are organized, it produces models that are internally consistent and externally misleading.
Francis Crick said it plainly: “Biologists must constantly keep in mind that what they see was not designed, but rather evolved.” Evolution does not favor elegance. It favors survival. Surviving systems are redundant, layered, and contingent in ways no parsimonious account captures.
Seven hundred years is a long run for a heuristic. The question isn’t whether to retire it. The question is whether we’ve confused a working shortcut for a scientific law — and how many wrong models we’ve built in the meantime.
Eswar Krishnan, MD
Drug Development Executive
This should not be considered to be specific medical or financial advice.
#DrugDiscovery #Biotech #ClinicalDevelopment #QuantumBiology
Approve this draft and I’ll save it to output/blog_occams_razor_26_04_11_v1.0.0.md. Or flag any sections to revise first.
Sources: