tomaarsen/mammal-entity-retrieval-eval

MAMMAL entity-retrieval evaluation

Evaluation of ibm-research/biomed.omics.bl.sm.ma-ted-458m (IBM MAMMAL, T5 encoder-decoder 458M) as an embedding/retrieval model over its native modalities — protein sequences and SMILES molecules. Embeddings are mean-pooled T5-encoder hidden states from the model's own modular-tokenizer entity prompts.

Two models are reported: the zero-shot base checkpoint, and a contrastively aligned finetune (tomaarsen/biomed-mammal-ted-458m-dti-contrastive) trained with symmetrized in-batch InfoNCE (temp 0.07, lr 1e-4, 3 epochs, 483 steps, effective batch 32) on 5,156 (target protein, ligand SMILES) pairs from amirhallaji/bindingdb_kd (pKd ≥ 7, capped 50/protein). The 200 DTI eval targets are held out from training entirely (the training script recovers the eval set by replicating the eval script's RNG sequence, verified identical).

Results

Benchmark	Metric	Zero-shot	DTI-contrastive
BindingDB per-target ranking (194 held-out targets)	mean AUROC	0.512 (= chance)	0.640
"	hit@10	0.302	0.458
SMILES↔SMILES Tanimoto retrieval (1,000 queries / 17,900 corpus)	Recall@10 (random = 0.0006)	0.354	0.408
ProteinGym DMS, 10 assays (~24k variants)	mean Spearman	0.356	0.354 (unchanged)

Key finding: contrastive alignment fixes cross-modal retrieval without degrading same-modality structure — molecule-space actually improved on Tanimoto recall, and the ProteinGym fitness proxy is unchanged within noise. Finetune loss fell 1.39 → 0.11.

Benchmarks

Benchmark	Protocol	Metric
SMILES↔SMILES	rank molecules by embedding cosine; ground truth = Tanimoto top-10 (Morgan FP r=2) over 20k ChEMBL-derived molecules (from amirhallaji/bindingdb_kd)	Recall@10 (random baseline reported)
BindingDB per-target ranking	given a target protein, rank ~18k candidate molecules by cosine(protein, mol); positives = measured pKd ≥ 7.0 (dataset: amirhallaji/bindingdb_kd)	mean AUROC, hit@10 over eligible targets
ProteinGym DMS subset	10 single-substitution assays from genbio-ai/ProteinGYM-DMS-zeroshot; variant score = cosine(mutant, WT)	per-assay Spearman, NDCG@10

Caveats: DTI positives are "measured pKd ≥ 7" against a candidate pool of all measured ligands (unmeasured pairs count as negatives); proteins truncated at ~986 residues. The ProteinGym score is a zero-shot embedding proxy, not a trained fitness predictor; published context is Protriever (arXiv 2506.08954), where raw ESM embeddings used as a retrieval score reach Spearman 0.432 — different protocol, not directly comparable.

Files

• eval.py — the evaluation script (--mode smoke|full, --model <repo-or-dir> to evaluate a finetuned checkpoint, --mock for plumbing tests)
• train.py — the contrastive alignment finetune (--mode smoke|full); full mode pushes the model and re-runs the full eval
• results/ — raw metric JSON per run (full_results.json, smoke_results.json, finetuned_results.json)

Reproduce

pip install biomed-multi-alignment==0.2.6 fuse-med-ml==0.4.1 datasets==5.0.1 rdkit==2026.3.6 scikit-learn scipy pandas
python eval.py --mode full --results results.json                                   # zero-shot
python train.py --mode full && python eval.py --mode full --model mammal_dti_contrastive \
    --results finetuned_results.json                                               # finetuned

Note: the tokenizer loads from ibm/biomed.omics.bl.sm.ma-ted-458m (no -research), the model from ibm-research/biomed.omics.bl.sm.ma-ted-458m. Training needs ~13 GB GPU at micro-batch 4 (effective batch 32 via accumulation; batch 32 direct OOMs 24 GB in T5 attention).