BindPred: a framework for predicting protein–protein binding affinity from language model embeddings

Motivation: Reliable predictions of protein–protein binding affinities are essential for molecular biology and therapeutic discovery. However, most computational methods rely on three-dimensional structural models, which are often unavailable for many complexes. Results: We introduce BindPred, a structure-agnostic input framework that predicts affinities directly from amino acid sequences by combining embeddings from large protein language models with gradient boosting trees. On the protein–protein binding (PPB)-Affinity benchmark, which comprises 11 919 diverse complexes, BindPred achieves a Pearson correlation coefficient of 0.86 in random split five-fold cross-validation. Ablation analysis indicates that evolutionary embeddings alone capture most of the predictive signals, while augmenting with physics-based energy terms from PyRosetta and BindCraft increases the correlation only by 0.01. A more stringent protein-level split that places entire protein families (wild-type and all mutants) exclusively in either training or testing sets, resulting in only a modest decline in performance, demonstrating robust generalization to novel interaction pairs. Because BindPred operates exclusively on sequence input, it enables rapid inference [approximately 3 million complexes per GPU (T4) hour], making proteome-scale screening computationally feasible. Availability: The pretrained model and inference pipeline are available in a Google Colab notebook: BindPred Colab notebook. The training dataset, code, and model weights are available on the hugging face: https://huggingface.co/hbp5181/BindPred.