STORM: A Lightweight Structure-Free Model for Binding Affinity Prediction
Organizations: Stanford University
Abstract
Accurately predicting a small molecule's binding affinity to a protein, a critical task in drug discovery, remains challenging. Physics-based approaches like free-energy perturbation (FEP) have very high computational costs and require expert setup. Recent machine learning--based cofolding models have reported accuracy approaching that of FEP---and at least one of them, Boltz-2, has already been widely adopted for affinity prediction---but they remain expensive for large computational screens and even more expensive for retraining on newly collected experimental data. To address these challenges, we introduce STORM: Similarity Transfer from Observed Reference Molecules. STORM is a fast, simple machine learning model that ensembles many small model fusions of diverse protein and molecule representations for binding affinity prediction. Trained on the same data, STORM matches or exceeds the accuracy of Boltz-2 on multiple evaluation datasets, while delivering a speed improvement of over five orders of magnitude for both inference and training. Our results also suggest opportunities to substantially improve affinity prediction accuracy by integrating historical experimental affinity data with structural and biophysical information.
Figures & tables
| Feature | Description | Input size |
|---|---|---|
| Target ID | Code derived from a hash of the target identifier. | 128 |
| GO | Gene Ontology annotations ( The Gene Ontology Consortium, 2026 ; Ashburner et al., 2000 ) . | 2,048 |
| InterPro | Protein-family and domain annotations ( Blum et al., 2025 ) . | 2,048 |
| Species | Human, nonhuman, or unknown indicator. | 3 |
| Representation | Description | Input size |
|---|---|---|
| Morgan count | Circular atom neighborhoods of radius 2 ( Rogers and Hahn, 2010 ) . | 2,048 |
| RDKit path count | Atom and bond paths of length 1–7 ( RDKit contributors, ) . | 2,048 |
| Atom pair count | Atom-type pairs and their graph distances ( Carhart et al., 1985 ) . | 2,048 |
| Topological torsion count | Sequences of four connected atoms ( Nilakantan et al., 1987 ) . | 2,048 |
| Avalon binary | Avalon molecular fingerprint ( RDKit contributors, ) . | 2,048 |
| 2D pharmacophore count | Pharmacophore feature pairs, triplets, and graph distances ( scikit-fingerprints contributors, ) . | 2,048 |
| Layer | Input size | Output size | Operation |
|---|---|---|---|
| Input fusion | 384 | Linear projection; add projected endpoint encoding | |
| Hidden layer 1 | 384 | 384 | Layer normalization, SiLU, dropout ( ) |
| Hidden layer 2 | 384 | 192 | Linear projection, layer normalization, SiLU, dropout ( ) |
| Prediction | 192 | 1 | Linear projection to pActivity |
| Model | ChEMBL35 | OpenFE | FEP+4 | CASP16 |
|---|---|---|---|---|
| STORM | ||||
| Boltz-2 (local, corrected) | ||||
| Structure-free DTA | ||||
| AttentionDTA | ||||
| PSICHIC | ||||
| BALM |
Appendix figures & tables15 assets
Supplementary material from the paper’s appendix.
Appendix
| Target representation | Ligand input | Votes |
|---|---|---|
| Canonical (ID128 + GO + InterPro + species) | Atom pair count | 1 |
| Canonical | Morgan count | 1 |
| Canonical | 2D pharmacophore count | 1 |
| Canonical | RDKit/Mordred graph-collapsed 1,212 descriptors | 2 |
| Canonical | Topological torsion count | 1 |
| Target blind | Atom pair count | 1 |
| Models left | Votes left | Component removed | OpenFE | ChEMBL35 |
|---|---|---|---|---|
| 15 | 18 | Full ensemble | ||
| 14 | 17 | Target-blind atom pair | ||
| 13 | 16 | Canonical 2D pharmacophore | ||
| 12 | 15 | ID-only atom pair | ||
| 11 | 14 | InterPro-only Klekota–Roth | ||
| 10 | 13 | InterPro-only merged descriptors |
| Evaluation set | 15 RDKit path fits | Diverse STORM |
|---|---|---|
| CASP16 | ||
| FEP+4 | ||
| OpenFE | ||
| ChEMBL35 Time Split |
| Condition | ChEMBL35 | OpenFE | FEP+4 | CASP16 |
|---|---|---|---|---|
| STORM | ||||
| Matched ligand-only STORM | ||||
| Repeated target-aware MLP | ||||
| Repeated ligand-only MLP |
| Benchmark | Unsmoothed | Smoothed | [95% CI] |
|---|---|---|---|
| ChEMBL35 | |||
| OpenFE | |||
| FEP+4 | |||
| CASP16 |
| Model | Overall ( ) | Enzyme ( ) | GPCR ( ) | Ion Channel ( ) | Kinase ( ) | Other ( ) | Transcription Factor ( ) |
|---|---|---|---|---|---|---|---|
| STORM | |||||||
| Boltz-2 | |||||||
| Structure-free DTA | |||||||
| AttentionDTA | |||||||
| PSICHIC | |||||||
| BALM |
| Benchmark | All rows | Matched rows | All assays | Matched assays |
|---|---|---|---|---|
| ChEMBL35 Time Split | 15,304 | 15,145 | 449 | 445 |
| OpenFE | 873 | 873 | 58 | 58 |
| FEP+4 | 87 | 87 | 4 | 4 |
| CASP16 | 140 | 140 | 2 | 2 |
| Model | ChEMBL35 | OpenFE | FEP+4 | CASP16 |
|---|---|---|---|---|
| STORM | ||||
| Boltz-2 (MW corrected) | ||||
| Structure-free DTA | ||||
| AttentionDTA | ||||
| PSICHIC | ||||
| BALM |
| Metric | Comparator | [95% CI] | Raw | Holm |
|---|---|---|---|---|
| Pearson | Boltz-2 (MW corrected) | |||
| Structure-free DTA | ||||
| AttentionDTA | ||||
| PSICHIC | ||||
| BALM | ||||
| Mattsson–Walters |
| Metric | Comparator | [95% CI] | Raw | Holm |
|---|---|---|---|---|
| Pearson | Boltz-2 (MW corrected) | |||
| Structure-free DTA | ||||
| AttentionDTA | ||||
| PSICHIC | ||||
| BALM | ||||
| Mattsson–Walters |
| Benchmark | Pearson | MAE | RMSE |
|---|---|---|---|
| ChEMBL35 | |||
| OpenFE | |||
| FEP+4 | |||
| CASP16 |
| Target | Compounds | Pearson |
|---|---|---|
| CDK2 | 16 | |
| JNK1 | 21 | |
| p38 | 34 | |
| TYK2 | 16 | |
| Overall | 87 |
| Benchmark | Model | No same-target ligands | ||||
| ChEMBL35 | STORM | |||||
| Boltz-2 | ||||||
| OpenFE | STORM | — | — | — | — | |
| Boltz-2 | — | — | — | — | ||
| FEP+4 | STORM | — | — | — | — | |
| Boltz-2 | — | — | — | — |
| Benchmark | Model | ||||
|---|---|---|---|---|---|
| ChEMBL35 | STORM | ||||
| Boltz-2 | |||||
| OpenFE | STORM | ||||
| Boltz-2 | |||||
| FEP+4 | STORM | — | — | ||
| Boltz-2 | — | — |
| Benchmark | Model | ||||||||
| ChEMBL35 | STORM | ||||||||
| Boltz-2 | |||||||||
| OpenFE | STORM | — | |||||||
| Boltz-2 | — | ||||||||
| FEP+4 | STORM | — | — | — | — | — | — | ||
| Boltz-2 | — | — | — | — | — | — |
Protein-Ligand Binding Affinity Prediction
Protein-ligand binding affinity prediction estimates how strongly a small molecule binds a protein to help prioritize compounds for drug discovery.
See all 30 papers