Generative AI for Drug Design: From New Molecules to Testable Ideas

Suppose a research team has a molecule with promising activity against a protein. The next design brief is specific: keep the features that seem important for that activity, explore changes that might improve solubility, and propose structures the team could realistically make.

Generative AI can help turn that brief into molecular proposals. The interesting work starts when the team asks which proposals deserve an experiment, and why.

A generated molecule is a design idea with questions attached. Its value depends on whether the structure makes chemical sense, whether it can be made, and what evidence would support developing it further.

What does a molecular generator produce?

A generative model proposes structures. A property model estimates something about a structure, such as behavior in an assay. Design systems often bring these capabilities together so that predictions can guide the next proposals.

The representation matters. A molecule can be described as text, as a graph of atoms and bonds, or with information about its three-dimensional arrangement. A generator works with a particular representation and must produce outputs that can be interpreted correctly.

One common text notation is SMILES. It encodes molecular connectivity using characters, with rules for branches, rings, and other features. A different string does not necessarily describe a different molecule. The simple example below shows why researchers need to compare structures, rather than just count distinct pieces of text.

Open Babel’s guide to molecular representations explains these distinctions and the use of standardized representations. They matter when removing duplicate proposals, searching for earlier work, and connecting a prediction to the structure that produced it.

How does a model learn to suggest structures?

Generative models learn patterns from molecular data. Depending on the method, a system may build a structure in stages, modify a starting structure, or search within a learned representation of molecules. Desired properties can help guide that search.

An influential 2018 study by Gómez-Bombarelli and colleagues demonstrated an approach that encoded molecules into continuous numerical representations and decoded them back into structures. That made it possible to explore changes within the learned space while estimating properties. It is one example of a generative approach, rather than a description of every molecular design model.

Whatever the architecture, the design brief needs to be clear. Is the team exploring a new chemical family? Making small changes to a known starting point? Preserving a particular part of a molecule? Those goals can lead to different searches and different ways of judging the output.

The model also needs information relevant to the task. A collection of molecular structures is different from a dataset linking structures to reliable measurements against the target of interest.

Why the highest score may not be the best choice

Imagine four invented candidates compared on predicted target activity and solubility. For this example, the team prefers higher values on both properties. The profiles are assumptions for teaching, not predictions for real molecules.

Candidates A, B, and C offer different trade-offs. None is better than the other two on both shown properties. Choosing among them requires a clearer view of the program’s needs and the uncertainty in the predictions.

Candidate D is lower on both properties than the others in this example. That makes it less attractive on the information shown. It could still matter if an unshown property, such as a more practical synthesis, changed the comparison.

Combining everything into one score can hide these decisions. A weighted score embeds choices about which properties matter and how much one should compensate for another. Some requirements may be firm constraints: a team may need to resolve a serious liability before any improvement elsewhere becomes useful.

From a proposed structure to a useful experiment

A medicinal chemist needs more than a high ranking. The team should be able to inspect the structure, understand the proposed route to making it, and connect each prediction to a defined question.

Check what was actually generated

Parsing a structure and checking basic bonding rules are useful first steps. They do not establish that the molecule is stable, easy to handle, or worth investigating. A proposal also needs the right chemical identity and enough detail to distinguish the structure being evaluated.

The team should remove duplicates and identify close relatives of known compounds. A large batch of slightly different-looking records may contain less chemical diversity than it appears to.

Examine whether it can be made

A proposed synthesis route is a plan. Its practicality depends on available starting materials, the required transformations, and the work needed to obtain and characterize the intended compound.

A computational estimate of synthetic accessibility can help prioritize that review. It is not equivalent to successfully making the molecule. A complicated proposal might still be worth pursuing, but the reason should be visible alongside the cost and uncertainty of testing it.

Ask what the experiment would settle

An assay, meaning a test of a chemical or biological effect, should address a question that matters for the next decision. Is the predicted target activity present? Does the compound dissolve under the test conditions? Does it behave similarly against related proteins?

These are different experiments. A result that supports one property leaves the other questions open. Keeping those distinctions visible helps a team decide what to measure next.

What has reached laboratory testing?

A 2019 study in Nature Biotechnology used a generative system called GENTRL to design inhibitors of DDR1, a protein kinase. The authors reported four compounds active in biochemical assays and two validated in cell-based assays. One lead was selected for a pharmacokinetic study in mice, which investigated how it behaved in the body.

Those results connect computational proposals with experimental evidence. They also illustrate why the endpoint matters. Activity in an assay and pharmacokinetic measurements are different findings; neither is evidence of treatment benefit in people.

When reading a generation result, look for what was made, what was measured, and how those measurements relate to the original design goal. The number of proposed structures alone cannot answer those questions.

What do molecular generation benchmarks measure?

The GuacaMol research paper introduced standardized tasks for evaluating molecular design methods. Useful distinctions include:

• Validity: Can the output be interpreted as a molecular structure under the checking rules?
• Uniqueness: How many generated structures are distinct within the batch?
• Novelty: Is a structure absent from the reference training set used for that comparison?
• Goal-directed performance: How well does the method optimize the specified computational objective?

These measures answer different questions. A novel structure can be difficult to make. A valid structure can have poor properties. A strong computational score can still disagree with an experiment.

Benchmark novelty is also narrower than being unknown to science. A molecule absent from one training set may already appear in another database or a publication. The scope of the search needs to be stated when a team describes something as new.

For a practical evaluation, compare a generative method with a suitable existing approach under similar conditions. Give both methods a comparable budget, apply the same review criteria, and examine the quality of the resulting candidates.

Why a molecule’s design history matters

A candidate may pass through many revisions before it reaches an experiment. Keeping that history helps the team understand which features were deliberately preserved, which predictions influenced a change, and which results challenged the idea.

The record should connect the structure with the design brief, model and tool versions, proposed synthesis, property estimates, and review decisions. It should also preserve reasons for rejecting a candidate. Otherwise, a later search may simply propose the same unsuitable idea again.

This is useful when measurements arrive. The team can compare the result with the prediction it actually used, rather than a newer model or a changed structure that happens to have a similar name.

Molecular generation at AgentCures

At AgentCures, molecular generation is connected to candidate filtering, property assessment, and scientific review. We compare proposals on several properties and keep the design history and rationale in version control so scientists can inspect the work.

The aim is to bring a manageable set of ideas and their supporting analyses to the team. New measurements can then inform the next proposals and help reveal which predictions were useful.

Our guides to ADMET , research agents , and AI in drug discovery explain the surrounding questions about exposure, tool use, and experimental evidence.

When a larger search helps

Generating many structures can broaden a search. It can also produce repeated ideas, uncertain predictions, and more material than a team can sensibly review.

The useful output is a set of testable options that helps researchers make their next decision. A smaller, diverse shortlist with a clear scientific rationale can be more valuable than a much longer list whose scores are hard to interpret.