A generative model can produce ten thousand molecules in an afternoon. That is not the bottleneck and never was. The bottleneck is deciding which five to make, and a list of structures without an assessment moves that decision nowhere: the chemist still has to evaluate every suggestion by hand, which is slower than designing them in the first place. A usable generative run therefore inverts the emphasis. The generation is the cheap part; the constraints before it and the scoring after it are the product.
What makes generated output unusable
Four recurring problems, in the order they waste time:
1. Not synthesisable. Strained systems, impossible substitution patterns, reagents that do not exist. A chemist spots these in seconds and loses trust in the whole list. 2. No property assessment. Structures with no prediction of the properties the project cares about. Ranking is left entirely to the reader. 3. Off-scaffold. Suggestions that leave the chemical series the project is committed to, ignoring the patent position, the synthetic platform and everything already learned. 4. Variations without substance. A hundred molecules that differ by a methyl group, which is not design but enumeration.
Constraints belong before generation, not after
Filtering a large generated set down afterwards wastes the generation and usually leaves the worst kind of survivors, those that pass the filters by accident. Stating the constraints up front changes what gets produced:
- Scaffold or core to retain, and the positions open to change.
- Reaction and building block scope: the chemistry your lab can actually run, and ideally the building blocks you can buy or already have.
- Property windows per objective, with the direction and the acceptable range, not just "better".
- Hard exclusions: substructures you will not make, known liabilities, intellectual property constraints.
- Novelty requirement: a minimum distance from existing compounds, so the output is not a rediscovery of your own file.
A generative run with these constraints produces fewer molecules, and a much higher proportion of them are worth reading.
What a usable output row contains
For each suggestion, the information a chemist needs to make a decision in one line:
| Field | Why |
|---|---|
| Structure | |
| Predicted values for each objective, with intervals | A value without an interval cannot be weighed against another |
| Applicability domain flag per prediction | Says which numbers are extrapolations |
| Synthesisability assessment | A proposed route to purchasable building blocks, with the number of steps |
| Novelty and nearest known compound | Shows whether this is new and what it resembles |
| Liability flags | Structural alerts, reactive groups, known problem substructures |
| Reason for the suggestion | Which objective it is supposed to improve relative to the starting point |
The last row is the one that turns a list into a conversation. "Proposed to raise solubility while keeping predicted potency inside the interval of the parent" is a hypothesis a chemist can accept or reject on chemical grounds.
Multi-objective is the normal case
Single-objective optimisation produces the known failure: a potency-maximising run that walks straight into insoluble, metabolically unstable territory. Real projects balance three to six objectives that trade off against each other.
Two practical rules. First, prefer a Pareto view to a single weighted score: show the set of suggestions where no other suggestion is better in every objective, and let the project choose the trade-off. A weighted sum hides the trade-off in coefficients nobody can justify. Second, state which objectives are hard constraints and which are to be maximised. "Solubility above a threshold" and "potency as high as possible" are different kinds of requirement and should not be averaged together.
Where generative design fits in the DMTA loop
It is one step inside a cycle, and its value depends on the cycle closing. Generation proposes, the lab measures, the measurements update the models, the next generation is better informed. Without the feedback, a generative model keeps proposing from the same prior, and the second cycle is no better than the first.
That is also the honest answer to what generative design cannot do: it cannot tell you whether your objective is the right objective. If the predicted property does not correlate with the project's actual endpoint, the optimisation will confidently walk in the wrong direction, and faster than a chemist would.
Limits of the statement
- Every predicted property in the output carries the uncertainty of its model, and generated molecules are by construction new, so they are more often outside the models' comfort zone than a normal series is.
- A synthesisability assessment is a proposal, not a validated route. The step that looks fine on paper may be the one that fails.
- Novelty against public databases is not freedom to operate. That assessment belongs with a patent professional.
- Generative design is a candidate generator. It does not replace the medicinal chemistry judgement that decides what is worth testing, and it works best when it is read as a set of hypotheses rather than a ranking to execute.
Frequently asked questions
What is wrong with a list of generated molecules?
It leaves the whole evaluation to the reader. Without property predictions, intervals, domain flags, a synthesis proposal and a stated reason per suggestion, a chemist has to assess each structure by hand, which is slower than designing candidates directly.
Should constraints be applied before or after generation?
Before. Post-filtering wastes the generation and tends to leave suggestions that pass the filters by chance. Stating scaffold, reaction scope, property windows, exclusions and novelty requirement up front changes what is produced.
How should multiple objectives be handled?
As a Pareto set rather than one weighted score, with hard constraints separated from objectives to maximise. That keeps the trade-off visible and lets the project make it instead of hiding it in weighting coefficients.
How is synthesisability assessed?
By proposing a route back to purchasable building blocks and reporting the number of steps and the confidence in each. It is an indication that a route plausibly exists, not a validated procedure.
Does generative design need experimental feedback?
Yes, to improve. Without measurements feeding back into the models, every cycle proposes from the same prior, so the second round is no better informed than the first.