PURE

PURE (Policy-guided Unbiased REpresentations) generates molecules that stay structurally close to a given seed compound while improving a target property — the task known as structure-constrained molecular generation, and the computational analogue of medicinal-chemistry lead optimisation. It was developed by Abhor Gupta, Barathi Lenin, Sean Current, Rohit Batra, Balaraman Ravindran, Karthik Raman, and Srinivasan Parthasarathy at the Robert Bosch Centre for Data Science and AI, Indian Institute of Technology Madras, with The Ohio State University. The work was posted in May 2025 and published in the Journal of Cheminformatics in October 2025.

The paper's central critique is methodological. Most deep generative approaches embed molecules in a continuous latent space and optimise within it, creating an impedance mismatch: molecules are discrete, and interpolating between them yields points that decode to nothing chemically meaningful. These models are also typically trained against the very property metric they are later evaluated on — QED, penalised logP, a DRD2 activity classifier — which the authors call metric leakage.

PURE avoids both problems by never leaving discrete chemical space and never using an external property metric during training. Generation proceeds as a sequence of reaction-template applications, and similarity to the seed molecule emerges from the learned representations rather than being enforced by a hand-coded Tanimoto penalty. A single trained policy is then applied to new seeds and new property objectives at inference, without retraining per target.

#Key Features

• Discrete, reaction-grounded generation: Every step is the application of a reaction template mined from experimental chemistry, so intermediate and final molecules are chemically valid and synthetically reachable by construction rather than by post-hoc filtering.
• No metric supervision during training: The policy is trained under a goal-conditioned reinforcement-learning objective over simulated molecular transformations, so no property scorer leaks into the learned representation.
• Graph-native representations: Graph Isomorphism Networks encode both the molecular state and the reaction signature that defines each action, keeping the state-action space in the same structural language as the chemistry.
• One policy across objectives: The same trained model produces optimised molecules for each property benchmark, with objective-specific selection applied at generation time instead of per-target policy retraining.
• Emergent similarity constraint: Because the action space consists of single-step transformations from the seed, structural proximity is a property of the search rather than a penalty term balanced against the property reward.

#Technical Details

The action space is built from 84,968 reaction rules mined from the USPTO-MIT reaction dataset, with starting materials drawn from the Enamine Building Block Global Stock catalogue. The offline reinforcement-learning dataset comprises 100,000 samples derived from 20,000 source-target pairs with trajectories of length five, split 80,000 for training and 20,000 for validation. Actor and critic networks each use hidden layers of 256 units with batch normalisation and ReLU, trained with a policy-gradient loss for the actor and mean squared error for the critic. Molecule generation at inference uses beam search over the template action space.

Evaluation covers four standard structure-constrained benchmarks — QED (800 molecules), DRD2 (1,000), and penalised logP at similarity thresholds of 0.4 and 0.6 (800 each) — scored on validity, property improvement, Tanimoto similarity to the seed, novelty, and diversity, against baselines including JT-VAE, VJTNN, VJTNN+GAN, CORE, COMA, HierG2G, and UGMMT. PURE achieves competitive or superior property scores while producing notably higher diversity and novelty: 1.0 validity with 0.866 diversity on QED, a total score of 0.772 with 0.846 diversity on DRD2, and a property score of 6.373 with 0.849 diversity on penalised logP at threshold 0.4.

#Applications

The intended workflow is lead optimisation: start from a compound with known activity and generate close analogues with better drug-likeness, potency, or physicochemical behaviour. The authors demonstrate this on drug resistance, seeking sorafenib analogues that evade ABCG2 efflux-mediated resistance while retaining BRAF engagement. Sequential filtering narrowed 1,744 docking candidates to 1,134 with improved ABCG2 binding affinity, then 214 that reproduced the critical BRAF hydrogen bond, then 81 passing drug-likeness filters; 30 had one or fewer lead-likeness violations against sorafenib's three, and nine had none. Because generation follows reaction templates, the resulting analogues come with a plausible synthetic route rather than an unmade structure.

#Impact

PURE's most durable contribution may be its framing of metric leakage as a systematic problem in generative chemistry benchmarking — a model trained on the metric it is scored by inflates results in a way that leaderboards do not surface. Demonstrating competitive performance without any metric supervision is direct evidence for that argument. The caveats are the usual ones for this class of model: the sorafenib case study is entirely computational, resting on docking and rule-based filters rather than synthesis and assay, and reaction-template generation is bounded by the transformations present in USPTO-MIT. Code, datasets, and pretrained embedders are available on GitHub, though the repository carries no license file.