BindPred

How tightly two proteins bind is a central quantity in antibody engineering, enzyme design, and interaction-network mapping, yet nearly every computational estimator of binding affinity demands a three-dimensional model of the complex. Free-energy perturbation protocols take hours per complex, and structure-based machine learning inherits both the cost and the errors of whatever produced the input coordinates — co-folding systems such as AlphaFold-Multimer and AlphaFold 3 still misplace partners or distort interfaces for assemblies with weak coevolutionary signal. BindPred, from Costas Maranas's group at Pennsylvania State University, removes the structural dependency outright: it predicts absolute log10 Kd for a protein-protein complex directly from the two amino acid sequences.

The architecture is deliberately spare. Each chain is encoded by a frozen protein language model, the pooled receptor and ligand vectors are concatenated, and a gradient boosting regressor maps that fixed-length feature vector onto affinity. Nothing in the language model is fine-tuned, so the learned component amounts to roughly 6,000 trainable parameters operating on precomputed embeddings. Unlike ΔΔG predictors, which score a mutation against a known wild-type complex, BindPred estimates affinity for complexes it has never seen, which is what makes it usable for ranking entirely novel pairs.

The paper's most consequential result is an ablation. Appending physics-based energy terms from PyRosetta and BindCraft to the embeddings improves the Pearson correlation by only about 0.01, and the advantage does not survive a stricter protein-level split. Sequence embeddings, in other words, already encode most of the interfacial energetic information that explicit force-field descriptors supply.

#Key Features

• Structure-agnostic input: Affinity is predicted from sequence only, so complexes with no crystal structure, no reliable homology model, and no confident co-folded prediction remain in scope.
• Proteome-scale throughput: Because inference is a single language model forward pass plus a tree ensemble evaluation, the pipeline scores roughly 3 million complexes per T4 GPU-hour.
• Absolute affinity rather than ΔΔG: The model outputs log10 Kd values on an absolute scale, enabling cross-complex ranking rather than only mutation-relative comparisons within a fixed scaffold.
• Physics terms shown to be redundant: Nineteen PyRosetta ref2015 energy terms and fourteen BindCraft interface-energy features were each tested as augmentations and contributed marginal gains over embeddings alone.
• Released checkpoints and inference notebook: A general-purpose model and an ACE2-RBD-specific model are distributed under an MIT license on Hugging Face, with a Colab notebook covering the full inference path.

#Technical Details

Sequence features are mean-pooled residue-level hidden states from the final transformer layer of either ESM-2 650M (esm2_t33_650M_UR50D, pretrained on UniRef50) or MINT, a cross-chain-attention variant of the same backbone trained on STRING interaction pairs. Receptor and ligand each yield a 1,280-dimensional vector, concatenated into a 2,560-dimensional input to a CatBoost regressor trained with an RMSE objective. Training uses "Corpus PPB-Affinity", 11,919 valid complexes assembled from SKEMPI v2.0, SAbDab, PDBbind 2020, Affinity Benchmark v5.5, and ATLAS.

Under random-split five-fold cross-validation the model reaches a Pearson correlation of 0.86, about 0.16 above the structure-dependent PPB-Affinity baseline; the ESM-2 and MINT variants differ by 0.003. Per-subset performance spans 0.898 on SKEMPI and 0.895 on Affinity Benchmark v5.5 down to 0.632 on the heterogeneous ATLAS set. Cross-subset "stranger" validation, where four subsets train and the fifth is held out, is substantially harder, with ESM-2 correlations from 0.194 on ATLAS to 0.667 on Affinity Benchmark v5.5. The authors document three failure modes: extreme-affinity outliers are compressed toward the mean, mutations with weak evolutionary signal are predicted to behave like wild type, and underrepresented complex families degrade.

#Applications

The combination of sequence-only input and high throughput suits screens that structure-based pipelines cannot reach: sweeping candidate interaction partners across a proteome, triaging antibody affinity maturation libraries, and prioritizing de novo binder designs before committing to expensive structural scoring. The ACE2-RBD checkpoint illustrates the specialization pattern, targeting receptor-binding-domain variants across host species. The authors note that BindPred ranks designed sequences by predicted affinity but does not reliably classify binders from non-binders, so it functions as a prioritization filter ahead of experimental or physics-based validation.

#Impact

BindPred contributes a clean empirical argument to an active debate about how much structural information sequence embeddings actually carry: a tree ensemble with ~6,000 parameters on frozen embeddings matches structure-augmented variants and beats a structure-dependent baseline, suggesting feature quality matters more than architectural sophistication for this task. The work was published in Bioinformatics after peer review, with weights, training code, and an inference notebook openly released. Its evaluation remains confined to splits of a single curated corpus, and the sharp drop under cross-subset validation indicates that accuracy is sensitive to assay heterogeneity and to how far a target sits from the training distribution.