Generative Design

Generative Molecular Design

The Design half of design-make-test-analyse. REINVENT 4 turns a scoring function into molecules — de novo, or by decorating a scaffold, linking two fragments, or transforming a known active. It is the most capable open-source generative framework in medicinal chemistry, and it will optimise exactly what you ask for, including the parts you did not mean.

Tool: REINVENT 4 v4.8, Apache-2.0. Not on PyPI — git clone and pip install -e .. Priors ship with the repository. An NVIDIA GPU is effectively required for reinforcement learning. Checked against: v4.8, June 2026.

Read references/reinvent-configuration.md before writing a run file, references/scoring-functions.md before defining an objective, and references/evaluating-generated-molecules.md before believing the output — that one is judgement, not syntax.

The three scripts

The scoring function is the experiment

The model does not know what you meant. Everything it produces is a literal consequence of the objective, and unbounded rewards get exploited without exception:

So every numeric component gets a window, not a direction. Most properties have a desirable range — molecular weight should be 250–500, not "as low as possible" — which is what double_sigmoid is for. A reverse_sigmoid on molecular weight optimises toward methane.

python skills/generative-design/scripts/scoring_profile.py profile --objective cns

[stage.scoring]
type = "geometric_mean"

[[stage.scoring.component]]
[stage.scoring.component.MolecularWeight]
[[stage.scoring.component.MolecularWeight.endpoint]]
name = "MolecularWeight"
weight = 0.15
transform.type = "double_sigmoid"
transform.high = 380.0
transform.low = 200.0

Geometric mean, not arithmetic, wherever a component is a requirement rather than a preference: a single zero zeroes the total, which is exactly what you want from custom_alerts. Without that alerts component the agent rediscovers reactive and PAINS-like chemistry, because those substructures score well on everything else.

Four generators, and matching the input to the prior

python skills/generative-design/scripts/reinvent_config.py generators

Attachment points are mandatory and the script refuses without them:

$ reinvent_config.py staged --generator libinvent --scaffold "c1ccccc1"
error: the scaffold has no [*] attachment point. LibInvent decorates at [*];
without one there is nothing to decorate.

sampling ignores the scoring function entirely — it draws from the prior. If a run seems to ignore its objective, check run_type before anything else. Optimisation is staged_learning.

Mol2Mol's choice of prior is the novelty dial: mol2mol_similarity gives conservative analogues, mol2mol_scaffold_generic hops aggressively. There is no separate setting.

Mode collapse looks exactly like success

The default failure. The agent finds one scaffold that scores well and decorates it forever; the score curve rises smoothly and the run looks like it worked.

python skills/generative-design/scripts/parse_run.py summary --csv run_1.csv

# 60 rows -> 60 distinct (0 duplicates)
# MODE COLLAPSE: only 2 ring systems across 60 molecules (3.3%). The agent has
# found one core it can decorate. Add a diversity filter, lower sigma, or shorten the run.

The score column cannot show this; the scaffold count can. Always set a diversity filter — without one, having found something that scores well, decorating it is the cheapest way to keep scoring well.

sigma is the other dial: too high and the agent collapses onto degenerate high scorers, too low and it barely moves from the prior. If output stops looking like chemistry, lower sigma first.

Four ways generative runs mislead

1. Nothing in the default objective knows what can be made. The harder you optimise, the further into unsynthesisable space the agent goes. Add an SA-score component, and check the output with retrosynthesis.
2. Validating with the scoring function is circular. If a docking score was in the objective, the output scores well on it by construction. Assess with something the optimiser never saw.
3. Novelty is not automatically good. A molecule absent from ChEMBL may be absent because nobody wanted it.
4. Uniqueness is far below the generated count. "50 000 generated" is usually a few thousand distinct.

When to stop and look

Fifty structures, by eye, will reveal a reward hack that no metric does. Absurdity is not a computable property, and this step is not optional.

Composing with the rest of the bundle

• chembl → before: actives to seed Mol2Mol or to fit a potency model for the objective.
• binding-site-analysis → before: is there a pocket worth designing into?
• medchem → after: alerts and PAINS on the output, as a check on the alerts component.
• retrosynthesis → after: can any of these actually be made?
• admet-prediction → after: developability across the survivors.
• autodock-vina / free-energy-perturbation → after: an orthogonal score the agent never saw.

Reporting results honestly

Give the number generated, distinct, and distinct-by-scaffold — not just the first. Quote the scoring function in full, because the output is a function of it, and say whether a diversity filter was used. Confirm that any post-hoc assessment used a method absent from the objective. And report how many molecules a human actually looked at.