BALM

Picking hit compounds out of a billion-molecule library needs a scoring function that is both fast and able to rank. Docking is fast but ranks poorly; alchemical free energy methods rank well but cost far too much to run at library scale. Sequence-based deep learning sits between them, yet most such models are trained and evaluated on random data splits that leak near-identical compounds between train and test, so their reported accuracy overstates what happens when a genuinely new target arrives.

BALM, from the Mey group at the University of Edinburgh with collaborators at the University of Barcelona, attacks this at the model, data, and evaluation levels at once. Its mechanism is metric learning: a protein sequence is encoded by ESM-2, a ligand SMILES string by ChemBERTa-2, both are projected into one shared latent space, and the cosine similarity between the two vectors is the predicted binding affinity. There is no concatenation-and-regress head and no 3D complex; affinity is read off the geometry of the embedding space.

Alongside the model, the authors argue for a stricter way to report performance. Because a single global correlation over a pooled test set hides the fact that a model may be excellent on one target and useless on the next, they evaluate per target and aggregate with Fisher-transformed correlations, exposing variance that aggregate metrics conceal. The work appeared as a preprint in late 2024 and was subsequently published in the Journal of Chemical Information and Modeling.

#Key Features

• Affinity as embedding distance: Training maximizes cosine similarity for strong binders and minimizes it for weak ones, so the learned space directly encodes pKd rather than feeding a separate regression head.
• Parameter-efficient dual fine-tuning: LoKr adapters on the protein encoder and LoHa adapters on the ligand encoder adapt both language models while their base weights stay frozen, tuning only a small fraction of parameters.
• Stress-tested data splits: Evaluation spans random, cold-target, cold-drug, and Murcko-scaffold splits on a de-duplicated BindingDB subset, plus the leak-proof LP-PDBBind benchmark built to eliminate train-test similarity.
• Per-target reporting: Fisher-transformed per-target correlations replace pooled aggregates, surfacing the wide spread in zero-shot quality across individual proteins.
• Cheap enough for screening: Zero-shot scoring of 2,000 ligands against a target takes about 90 seconds on a single A100, and few-shot adaptation to a new target takes tens of minutes.

#Technical Details

The protein encoder is the 150M-parameter ESM-2 checkpoint (640-dimensional embeddings) and the ligand encoder is ChemBERTa-77M-MTR (384-dimensional); linear layers with ReLU project both into a shared 256-dimensional space. Training uses the Kd subset of BindingDB — about 48,000 interactions over 1,090 targets and 9,900 ligands, reduced to roughly 25,000 interactions after discarding measurements pinned at assay limits, a cleaning step that shrinks the data but removes a large artificial spike near pKd 5.

Adapter choice was tuned separately per encoder: LoHa at rank 16 improved ligand-side Pearson correlation by 9.4% and LoKr at rank 8 improved protein-side correlation by 18.2%, with the combination giving a 23.4% improvement over unadapted BALM on the random split, roughly 20% on cold target and 10% on cold drug. On the LP-PDBBind CL2 split (2,100 complexes across 12 protein families), zero-shot BALM beats rescoring with AutoDock Vina on Transferase targets by a wide margin (Pearson 0.65 versus 0.12, RMSE 1.82 versus 6.22 kcal/mol) and edges it on Chaperones (0.72 versus 0.65); on Oxidoreductases Vina ranks better (Spearman 0.66 versus 0.59) while BALM keeps the lower RMSE. Few-shot fine-tuning of just the projection layer lifts Pearson correlation on USP7 from 0.64 to 0.81 and on the SARS-CoV-2 main protease from 0.11 to 0.66 using 20% of the target's data.

#Applications

BALM is aimed at early-stage hit identification: triaging large on-demand libraries, re-ranking docked poses before committing compute to free energy calculations, and scoring targets for which no reliable three-dimensional structure exists, since it never sees a complex. Its few-shot mode fits campaigns where a handful of measurements against a new target already exist and a project team wants a target-specific scorer within the hour.

#Impact

The clearest contribution beyond the architecture is methodological: by showing how far per-target correlations scatter around a flattering pooled average, the paper makes a concrete case for how binding affinity models should be benchmarked. The training and evaluation code is released under the MIT license, and the fine-tuned checkpoints and curated benchmark splits are on Hugging Face, making the results directly reproducible. The authors are equally clear about where the model falls short: zero-shot quality varies sharply between individual targets, and on congeneric series from lead optimization benchmarks the predictions cluster in too narrow a range to separate closely related analogues, which keeps the model in the screening stage rather than in lead optimization.