Ligand Fingerprints, Not Expert Routing or Pocket Conditioning, Drive Cross-Family Affinity Prediction in Protein Foundation Models: A Controlled Study on 56,637 Measured ChEMBL Records

Protein foundation models such as ESM-2 encode transferable structural and evolutionary features, and sparse expert-routing overlays with router re-anchoring have been proposed as a route to modular drug–receptor specialization—accompanied by simulated reports of large accuracy margins. In this study we test what survives measurement. We curate 56,637 measured Ki/IC50/Kd records for eleven human receptors (dopamine D2/D3, serotonin 5-HT1A/5-HT2A, muscarinic M1, GABA-A α1, adrenergic α2A/β2, histamine H1, and the SERT/DAT transporters) from ChEMBL, reduce them to 41,271 unique receptor–ligand pairs (23,872 records → 17,177 pairs on the four-receptor core), and combine 2048-bit Morgan fingerprints with embeddings from the genuine frozen ESM-2 650M checkpoint. Under Bemis–Murcko scaffold-disjoint splits, five-seed Pearson correlation reaches r = 0.50–0.80 for the ten data-rich receptors (not the simulated 0.834); random splits overstate performance by 0.10–0.21 in r on the core, confirming that naive splits largely measure analogue recall; and a temporal split (train on pre-2015 records, test on 2015+) lowers performance further (shared-model r = 0.11–0.72). Three controls are decisive, and all three persist across the expanded set. First, the trained router remains near-uniform (gate entropy ≥ 93% of maximum) and never outperforms a frozen-uniform router—measuring router blindness rather than curing it. Second, cross-attention receptor conditioning improves the shared model only marginally and not significantly (paired Wilcoxon across seeds, p ≥ 0.0625), and a wrong-pocket control that deliberately supplies the wrong pocket residues matches it: the small gains do not come from pocket structure. Third, a Ridge regression on the fingerprints alone is competitive throughout (r = 0.39–0.83) and is the only model with signal on the low-resource receptor (r = 0.53 where the deep models collapse to r ≈ 0.01). We conclude that, on measured data, expert routing and contrastive router re-anchoring are not the operative mechanism of cross-family affinity prediction, and that fingerprint baselines remain the bar such architectures must clear. All curated data artefacts, SHA-256 checksums, and evaluation scripts are publicly available at: github.com

Authors

Publication Details

Journal
Figshare
Published
2026-11-13
DOI
https://doi.org/10.6084/m9.figshare.35002837
Primary Topic
Computational Drug Discovery Methods
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Ligand Fingerprints, Not Expert Routing or Pocket Conditioning, Drive Cross-Family Affinity Prediction in Protein Foundation Models: A Controlled Study on 56,637 Measured ChEMBL Records

Md Mushfiqur Rahim
2 links
Figshare
Computational Drug Discovery Methods
preprint

Ligand Fingerprints, Not Expert Routing or Pocket Conditioning, Drive Cross-Family Affinity Prediction in Protein Foundation Models: A Controlled Study on 56,637 Measured ChEMBL Records

Md Mushfiqur Rahim
preprint en

Abstract

Protein foundation models such as ESM-2 encode transferable structural and evolutionary features, and sparse expert-routing overlays with router re-anchoring have been proposed as a route to modular drug–receptor specialization—accompanied by simulated reports of large accuracy margins. In this study we test what survives measurement. We curate 56,637 measured Ki/IC50/Kd records for eleven human receptors (dopamine D2/D3, serotonin 5-HT1A/5-HT2A, muscarinic M1, GABA-A α1, adrenergic α2A/β2, histamine H1, and the SERT/DAT transporters) from ChEMBL, reduce them to 41,271 unique receptor–ligand pairs (23,872 records → 17,177 pairs on the four-receptor core), and combine 2048-bit Morgan fingerprints with embeddings from the genuine frozen ESM-2 650M checkpoint. Under Bemis–Murcko scaffold-disjoint splits, five-seed Pearson correlation reaches r = 0.50–0.80 for the ten data-rich receptors (not the simulated 0.834); random splits overstate performance by 0.10–0.21 in r on the core, confirming that naive splits largely measure analogue recall; and a temporal split (train on pre-2015 records, test on 2015+) lowers performance further (shared-model r = 0.11–0.72). Three controls are decisive, and all three persist across the expanded set. First, the trained router remains near-uniform (gate entropy ≥ 93% of maximum) and never outperforms a frozen-uniform router—measuring router blindness rather than curing it. Second, cross-attention receptor conditioning improves the shared model only marginally and not significantly (paired Wilcoxon across seeds, p ≥ 0.0625), and a wrong-pocket control that deliberately supplies the wrong pocket residues matches it: the small gains do not come from pocket structure. Third, a Ridge regression on the fingerprints alone is competitive throughout (r = 0.39–0.83) and is the only model with signal on the low-resource receptor (r = 0.53 where the deep models collapse to r ≈ 0.01). We conclude that, on measured data, expert routing and contrastive router re-anchoring are not the operative mechanism of cross-family affinity prediction, and that fingerprint baselines remain the bar such architectures must clear. All curated data artefacts, SHA-256 checksums, and evaluation scripts are publicly available at: github.com

Figshare
Computational Drug Discovery Methods
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon