A lot of the conversation in drug discovery AI still assumes the best model is the one with the broadest prior knowledge. The biggest pretraining corpus. The richest multimodal input. The largest molecular foundation model. The most chemically fluent latent space.

That sounds sensible.

But in a live project, it is often the wrong way to think.

The best deployment model is not always the one that knows the most on day one. It is often the one that learns fastest by day ten.

Drug discovery is not a static knowledge test

Benchmark culture still treats molecular ML as though it were an exam. Train on a giant corpus, hold out a test set, measure how well the model generalises, reward the strongest prior.

Live small-molecule projects do not feel like exams. They feel like fast, messy, biased, local feedback loops.

A project starts with global prior knowledge. But very quickly it generates its own reality: its own assay conditions, chemotype, liability pattern, synthesis constraints, structure–activity quirks, and surprises.

Once that starts happening, the question changes.

It is no longer: How much chemistry did the model absorb before the project began?

It becomes: How quickly can the model adapt once the project starts speaking?

That is a different standard. And it is badly under-rewarded.

Rapid adaptation is often more valuable than broad knowledge

This is not an argument against large pre-trained models. Strong priors matter. If you have zero local data, general chemical knowledge and cross-project memory start you from a better place.

But that is only the beginning.

In many real settings, local project data rapidly become more important than broad priors. Project data are not just “more labels”. They are labels from the exact world that now matters — the exact assay readout, the exact compound series, the exact route constraints, the exact potency cliffs, the exact permeability failure modes.

That local information is dense, specific, and relevant. Once you have even 5 or 10 or 20 points, a model that absorbs those points aggressively can outperform a more knowledgeable but less adaptable global system.

The field still underestimates this asymmetry.

Project data are sparse, biased, and assay-specific — and that is fine

One reason people romanticise generality is that project data are messy. Sparse, biased, highly local, gathered in bursts, contaminated by strategy and synthesis constraints.

In benchmark culture, that sounds like a defect. In deployment, it is just reality.

A live lead optimisation campaign is not trying to estimate universal truth over all of chemistry. It is trying to make the next good decision under local conditions.

A permeability model learned from ten compounds in the exact series, run in the exact assay system, may be dramatically more useful than a broader model with better average performance elsewhere. A potency model that has seen the first few vectors explored in a congeneric series may now be operating in the regime that matters most.

Sample efficiency matters. The local world takes over fast.

The field over-rewards what the model knows before work begins

Most model evaluation implicitly rewards one thing: how much useful information the model carries in from elsewhere.

That is one kind of intelligence. But in live drug discovery, another kind is often more valuable: how efficiently the model converts small amounts of local feedback into better next decisions.

It is closer to learning rate than memory. Closer to adaptation than pretraining.

Many evaluations barely look at it. We celebrate models for broad zero-shot competence while under-measuring whether they become materially better after a handful of local datapoints. A model that starts slightly worse but updates extremely well may dominate the real project.

Local context rapidly dominates global priors

Many ML people only half-believe this until they sit inside a real medicinal chemistry programme.

The project acquires personality very quickly. A lipophilicity move that usually helps potency may destroy permeability in this series. A heterocycle class that looks attractive globally may be a dead end in this assay system. A scaffold that looks tractable in the abstract may be operationally terrible for this team at this moment. A broad synthesis planner may prefer elegant routes the project will never run.

Local information compounds in value because it is not merely narrowing uncertainty — it is correcting the model’s worldview to match the actual game being played.

Good scientists do this on projects. The best deployed models should do it too. They should become project-native quickly.

The right benchmark is not just zero-shot performance

If the field took deployment seriously, one of the most standard benchmark plots would be:

Performance after 0, 5, 10, 20, and 50 local datapoints.

That curve would tell you more about practical value than a single held-out test number. It answers the real question: How quickly does this model become useful once the project starts generating feedback?

That benchmark should sit next to every claim of generalisation. Not just: how good is the prior? But: how steep is the local learning curve? When does the model start beating simple project baselines? How many assays before it becomes decision-relevant?

”Learns fast” should be a first-class metric

Potency

For potency prediction inside a live series, the useful model is often the one that quickly sharpens its ranking after early assay data arrive. Not the one with the most universal view of structure–activity relationships — the one that best absorbs the first local SAR inflections.

ADME

Nobody serious is trying to eliminate ADME measurement entirely. The real question is whether the model becomes useful after a few local readouts. After 5 or 10 compounds, can it start ranking which analogues are likely to preserve permeability? Can it help avoid repeating the same stability mistake?

Synthesis planning

The best planner is not always the one trained on the largest historical corpus. It may be the one that most quickly adapts to the chemistry preferences, reagent availability, and tactical style of the current team.

Simple local models remain competitive for a reason

People sometimes talk as though it is embarrassing that ECFP4 plus XGBoost, nearest-neighbour methods, matched-pair rules, or project-local delta models remain hard to beat.

It is exactly what you would expect in a domain where decisions are local, data are sparse, assays are specific, chemotypes are narrow, and the highest-value signals are concentrated in the current neighbourhood.

A simple model that absorbs local information efficiently can be strong not because it is philosophically satisfying, but because it is aligned to the real structure of the task.

Some sophisticated models disappoint in deployment for the inverse reason: they are optimised to know a lot broadly but not to update sharply under project-local feedback.

Pretraining is a starting position, not a finished product

Pretraining gives you a starting position. A good one matters. But discovery projects are not won from the starting position. They are won through the subsequent moves.

A model that begins from an excellent prior but adapts sluggishly can still lose to one that begins from a weaker prior but updates aggressively from local data.

Who would you rather have on a project? A scientist who has read everything, but keeps applying generic truths after the project has clearly started behaving differently? Or a scientist with strong fundamentals who notices the local pattern, updates, and starts making better calls after the first few compounds?

The second scientist is usually more valuable. The same holds for models.

A better evaluation stack

The field should talk about two distinct axes:

Prior quality: How useful is the model before any local project data?

Adaptation quality: How quickly does the model improve once local project data arrive?

For early target enablement or cold-start discovery, prior quality may dominate. For lead optimisation, ADME refinement, and active project support, adaptation quality often becomes the main event.

Once you frame things this way, benchmark confusion disappears. You stop asking for one magical number. You start asking the model to declare what kind of usefulness it offers.

The right benchmark question

Not just: How much does the model know?

But: How quickly does the model learn once the project begins?

In drug discovery, the second question is the one that matters.