Can Boltz‑2 Predict Protein‑Protein Affinity? Let's See What Science Says

Predicting how tightly two proteins bind is key to understanding cellular processes and designing new drugs. A research team tried to fine‑tune Boltz‑2, an AI model already famous for estimating protein‑ligand affinity, to work on protein‑protein interactions. They tested the new Boltz‑2‑PPI on two datasets: TCR3d, which contains T‑cell receptors, and PPB‑affinity, with experimental binding measurements. Even though the model can reconstruct structures very accurately, its affinity predictions are worse than those from sequence‑only methods. When they combined Boltz‑2‑PPI embeddings with those from sequence models, predictions improved, especially for weaker sequence models. This shows that the two model types capture different information: structure provides geometric detail, while sequence captures evolutionary patterns. The authors conclude that current structural representations are not yet optimized for affinity prediction, and biases from structural data limit performance. The work suggests continuing to integrate multiple information sources for more reliable predictions.

Boltz‑2 is built on a transformer‑type neural network that processes atomic coordinates of a protein‑ligand complex to generate a three‑dimensional embedding. In adapting it to Boltz‑2‑PPI, the authors kept the same architecture but altered the regression head to predict continuous ΔG binding values between two protein chains. The TCR3d and PPB‑affinity datasets contain respectively 1,200 and 5,000 protein‑protein pairs, with affinities measured by techniques such as SPR and ITC. A major challenge is the variability of interfaces: unlike ligands, proteins have larger and more flexible contact surfaces, making it harder to capture interaction forces with a single embedding. Moreover, available structural data are heavily biased toward high‑affinity complexes, introducing a learning bias. Sequence‑based models, like those using AlphaFold or ESM‑2 embeddings, exploit evolutionary information and residue co‑variation, which are especially useful for recognizing conserved interface patterns. Integration of structural and sequence embeddings was performed by concatenating the vectors and re‑training a small fully‑connected network. Results show an average increase of 0.12 kcal/mol in root‑mean‑square error (RMSE) compared to pure sequence models, highlighting a non‑negligible complementarity. However, the improvement remains modest, suggesting current geometric representations fail to capture all energetic factors (such as side‑chain entropy or solvation effects) influencing affinity. Future work could involve pre‑training on molecular dynamics simulations or employing deeper graph neural networks to model atom‑level interactions, aiming to close the gap between structure‑ and sequence‑based predictions.