Each candidate is a small ligand (a core atom plus 3 mutable substituents) proposed by a generative policy, in the spirit of REINVENT-style reinforcement learning: a policy network samples molecular edits, receives a reward, and is nudged toward higher-reward edits over many generations.
shape_fit = exp(-(d - d0)^2 / 2σ²) — docking pose vs pocket geometry
ΔG_dock ≈ -(w1·shape_fit + w2·H-bonds - w3·clash) — predicted binding energy
ADMET_pass = MW∈[160,500] ∧ logP∈[-0.5,5] ∧ HBD≤5 — Lipinski-like filter
reward = α·norm(-ΔG_dock) + β·ADMET_pass — used for selection
- Molecular generation — every generation, the elite fraction is kept and the rest of the population is re-derived from it with random substituent swaps and position jitter (mutation rate), the same explore/exploit trade-off an RL policy makes when sampling new molecules.
- Docking prediction —
shape_fit scores how well a candidate's distance and orientation match the pocket's ideal binding geometry (d0, set by pocket tightness); clashing too close is penalised, sitting far away scores near zero.
- ADMET prediction — a Lipinski-like Rule-of-5 heuristic over the candidate's proxy molecular weight, lipophilicity (logP) and hydrogen-bond donors flags it green (likely drug-like) or red (likely to fail absorption/toxicity screening).
- AlphaFold pipeline — the pocket geometry itself stands in for a predicted receptor structure: in a real pipeline this cavity comes from an AlphaFold (or crystallographic) protein model, and this scene is what docking the generated ligands into that predicted fold would look like.
Selection pressure (elite fraction) controls how quickly the population converges: low elite fractions with high mutation explore broadly but converge slowly; high elite fractions with low mutation converge fast but risk getting stuck on a mediocre scaffold — the same exploration/exploitation tension real generative drug-design pipelines have to tune.