This Week In Cheminformatics: Issue #013

Saturn for sample-efficient molecular generation, minimum message length based fingerprint, AF3 for covalent virtual screening and a long list of papers.

Mar 22, 2026

Highlights

Compressing Chemistry Reveals Functional Groups

Sharma and King’s paper applies the Minimum Message Length principle to validate conventional functional groups. They developed an unsupervised algorithm, FGCOMPRESS, to identify valid SMILES substrings that maximally compress nearly three million ChEMBL molecules. What makes this compelling is that their results aligns almost perfectly with what most chemists would think about when we say functional groups. Beyond just validation, when these discovered substructures are used to generate unhashed, count-based vectors, they significantly outperformed MACCS, Morgan fingerprints, and even MolFormer-XL continuous embeddings in Ridge Regression across 24 bioactivity prediction data sets !

Discovery of Covalent Ligands with AlphaFold3

Shamir et al. evaluate AlphaFold3 for covalent virtual screening, demonstrating that it significantly outperforms classical docking algorithms like DOCKovalent and DOCK6. To quantify enrichment, the authors introduce COValid, a benchmark dataset comprising 874 active acrylamide protomers and 37,919 topologically dissimilar, property-matched decoys across ten target cysteine sites. Ranking AF3 cofolding predictions strictly by the minimum predicted aligned error (mPAE) yielded near-optimal classification of true binders, surpassing both traditional docking outputs and Rosetta-rescored AF3 models. This work is highly relevant for virtual screening workflows as it establishes mPAE as an effective empirical ranking function. It is worth noting the authors found mPAE to be unsuitable for precise cross-target selectivity prediction.

Sample-efficient generative molecular design using memory manipulation

Saturn by Guo and Schwaller is a Mamba state space model for goal-directed molecular generation. This paper focuses on sample efficiency. The authors demonstrate that combining Mamba with Augmented Memory’s experience replay and SMILES augmentation induces a “hop-and-locally-explore” sampling behavior. Under fixed multi-parameter optimization budgets, Saturn consistently outperforms strong baselines like GEAM, successfully generating molecules with strict QED and synthetic accessibility profiles in significantly fewer oracle calls. Good read !