An AI biotech can generate thousands of molecular ideas before lunch. That says little about whether any are worth making.
The more revealing question is what happens after an idea meets an experiment. Can the team identify why a prediction failed, improve the method, and use the lesson on another program? Can it show that this process makes a research decision better?
Building around AI means designing a company that can answer those questions repeatedly. The models are one part of that organization. Its data practices, experimental capabilities, and review habits matter just as much.
Where a Shared Platform Helps
A drug program needs many recurring capabilities: preparing molecular structures, checking assay data, comparing properties, planning experiments, and keeping documents consistent. Reusing these capabilities can save a team from rebuilding routine work for every target.
But reuse has limits. A workflow that handles molecular identity reliably may transfer well. A model trained on one assay may perform poorly on another. A target’s role in disease requires its own evidence, even if the software that organizes that evidence is shared.
Molecular identity · data checks · analysis code · provenance · document templates
Establish the target rationale, assay performance, candidate activity, and relevant exposure.
Establish the cell model's relevance, endpoint reliability, candidate response, and mechanism.
An improvement to shared infrastructure may help both programs. A positive result in one does not validate the other.
This is why a platform should make its boundaries easy to inspect. Which steps use the same tested code? Which require new training data? Which conclusions still depend on a scientist’s assessment of the biology?
What Investors Need to Assess
The same questions help potential partners and research teams. Ask for a specific example of a decision that changed: a compound not synthesized, an assay repeated, a target deprioritized, or a new chemical series selected.
Then ask what comparison supports the claimed improvement. Saving time against a fully manual workflow is useful, but it may say little about performance against an experienced team using ordinary software. A comparison should name its baseline, its resource limits, and the outcome measured.
Generating more proposals and choosing better proposals are separate achievements. Counting both under a single productivity number obscures what the platform actually contributes.
A Benchmark Is a Starting Point
A model can look strong when the test set contains molecules very similar to its training examples. That may be appropriate for predicting close analogues. It is less informative when a team wants to explore unfamiliar chemistry.
The MoleculeNet benchmark includes different ways of separating training and test data, including grouping molecules by their structural frameworks. The broader lesson is practical: the evaluation needs to resemble the decision the model will face.
A time-based evaluation can also help. Train using information available before a defined date, then evaluate on later observations. Whatever split is used, duplicates, related measurements, and repeated optimization against the test set can make the result overconfident.
The AgentCures Model
At AgentCures, AI agents, shared scientific tools, and version-controlled research form the basis of our approach. An analysis should stay connected to its inputs, methods, and interpretation so the team can review it and revisit the decision later.
The useful unit of work is a research question with a checkable result. For example: among a fixed set of compounds, which should be tested next to learn about activity and selectivity? Recording the shortlist before results arrive makes the prediction easier to assess honestly.
The diagrams in this article describe evaluation principles. They are not performance results for AgentCures or a claim that shared software has already improved a clinical outcome.
What AI-First Does Not Mean
It does not remove the need for assay development, medicinal chemistry, clinical expertise, or manufacturing. Those activities supply much of the evidence that gives a model’s output meaning.
It also does not make programs independent. Several candidates can share the same mistaken biological assumption or assay artifact. A portfolio with many project names may still have one underlying source of failure.
A useful review therefore examines shared dependencies as well as individual programs. If a method changes, which earlier conclusions should be checked again?
Measuring the Contribution of AI
A small set of well-defined measures can be more informative than a large dashboard. Track the time from an accepted result to a reviewed next step. Record how many proposed compounds can actually be made and how many satisfy the experiment’s prespecified criteria. Include unsuccessful proposals and repeated measurements in the account.
Interpret these measures together. A higher hit rate may reflect an easier test set. Faster execution may reflect narrower scope. A method that finds fewer hits could still be valuable if it explores a previously inaccessible chemical series.
Building for the Future
A durable platform leaves behind useful research assets: reliable data, tested methods, explicit limitations, and decisions that another person can reconstruct.
That is a demanding standard. It gives a team a concrete way to improve, and gives collaborators a clearer basis for deciding what to trust. The strongest demonstration is a useful prediction made before the experiment, followed by an honest account of what happened.