bio.rodeo
ModelsOrganizationsProvidersLeaderboardAboutSign in
bio.rodeo

The authoritative source for evaluating biological foundation models. No hype, just honest analysis.

Categories
  • DNA & Gene
  • RNA
  • Protein
  • Small molecule
  • Single-cell
  • Spatial omics
  • Pathology
  • Imaging
  • Metabolomics
  • Biosignals
  • Language model
bio.rodeoModelsOrganizationsProvidersLeaderboardAboutFAQSubmit a modelContact
© 2026 Pulsatance. All rights reserved. ~
Built by Pulsatance
models / small-molecule / medsage
Small moleculeProtein
Stanford UniversityReleased May 2025

MedSAGE

Structure-based drug design by diffusing medicinal-chemistry fragments into a binding pocket, yielding synthesizable, selective, drug-like molecules.

5Openness

Where to run it

No providers recorded yet. Browse all providers

MedSAGESmall moleculeStanford University

Diffusion models have transformed de novo protein design, but their record on structure-based small-molecule design is weaker. Methods that diffuse individual atoms into a binding pocket routinely produce molecules a medicinal chemist would reject on sight: strained or bridged ring systems, reactive groups such as epoxides, excessive stereocenters, and scaffolds with no plausible synthetic route. They also tend to bind promiscuously, because optimizing a docking score alone rewards bulk and hydrophobicity.

MedSAGE, from the Dror lab at Stanford University, reframes the generative problem around a different unit. Rather than atoms or SMILES strings, it generates molecules as arrangements of chemical fragments — the functional groups and ring systems that constitute the working vocabulary of drug design. A library of several thousand fragments is curated from drug-like ligands, and each is assigned a low-dimensional vector encoding its Extended Three-Dimensional Fingerprint together with charge, polarity, hydrogen bond donors and acceptors, and atom count. Those embeddings are learned unsupervised with t-SNE, giving a smooth latent space in which nearby points are chemically similar fragments.

Generation proceeds in two phases. A denoising diffusion probabilistic model places fragment centroids and assigns their latent labels inside a fixed protein pocket, encoded in the same fragment representation. The second phase decodes those labels back to concrete fragments and searches for chemically valid bonds between neighbors, scoring candidate connectivities with a physics-based docking function.

#Key Features

  • Fragment-level diffusion: Representing a benzene ring as one labeled point rather than six positioned carbons cuts degrees of freedom more than tenfold, and structurally forbids the malformed rings that plague atom-level generators.
  • Synthesizability by construction: Generated molecules average a synthetic complexity score of 3.5 on a 1-10 scale, statistically indistinguishable from the 3.2 of reference drugs, while DiffSBDD, IPDiff, and PMDM score 4.9-5.3.
  • Selectivity, not just affinity: Scored as the docking gap between intended target and off-targets, MedSAGE molecules reach 3.6 against 4.2 for reference ligands, where all-atom baselines manage 1.3-1.5 — barely above random compounds.
  • Chemically valid connectivity search: A dedicated algorithm enumerates bonds between adjacent fragments under valence constraints and ranks candidates by docking score, plateauing after roughly 30-40 candidates.

#Technical Details

The denoiser is an equivariant graph neural network operating on the fragment point cloud, with context features distinguishing ligand points from pocket points; the pocket stays fixed throughout noising and denoising. Training data was built from every ligand-annotated Protein Data Bank structure, excluding lipids, peptides, carbohydrates, nucleotides, and ligands outside drug-like property ranges, then adding hydrogens, completing side chains, and assigning protonation states — roughly 35,000 protein-ligand complexes. The split is 85% training, 10% validation, and 5% test, with no protein sharing more than 30% sequence identity across splits.

The benchmark comprises 25 therapeutically relevant targets from the test set, each bound to an approved drug, clinical candidate, or optimized preclinical candidate with sub-micromolar affinity (mean 30 nM), with 400 molecules generated per target and evaluated on 31 physicochemical and pharmacological properties. Median Glide docking score is -9.2 against -9.5 for reference ligands (not significantly different), while DiffSBDD, IPDiff, and PMDM range from -5.5 to -6.3 and randomly sampled Enamine REALSpace compounds score -5.7. MedSAGE molecules reproduce 35% of the protein-ligand interactions made by reference ligands versus 12-23% for the baselines. Generated sets are diverse — average pairwise Tanimoto similarity below 0.1, with 197 unique scaffolds per 400 molecules — and drug-like, averaging 340 Da and logP 2.0. Against exhaustive virtual screening of 30-400 million compound libraries, the top 1% of MedSAGE molecules match the docking scores of the top 0.001% of screened hits.

#Applications

The intended workflow is hit generation for a target with a known holo pocket structure, the common starting point for early drug discovery. Because output molecules already satisfy medicinal chemistry constraints, they can feed directly into analog search: the authors match generated compounds to structurally similar, commercially available REALSpace compounds that retain comparable docking scores, turning a de novo design into an orderable one. Generating a few thousand molecules per target also substitutes for docking campaigns over hundreds of millions of compounds, making the approach attractive for teams without ultra-large-scale screening infrastructure.

#Impact

MedSAGE is a preprint awaiting peer review, and all reported validation is computational — affinity and selectivity are estimated by Glide docking scores rather than measured, so no synthesized compound has been assayed. Data and code are stated to be forthcoming after peer-reviewed publication, and no repository or weights are currently released; Stanford has filed a patent application on the work. The authors acknowledge two limitations: the model does not account for protein flexibility or induced fit, having been trained and benchmarked exclusively on holo structures solved with bound ligands, and the split controls sequence identity but not pocket similarity, so performance on genuinely novel pocket architectures is untested.

At a glance

Released
May 2025
Category
Small molecule
Organization
Stanford University

Links

Research Paper

Tags

de_novo_designdiffusiongenerativegraph_neural_networkmolecule_generation

Something wrong?

Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.