Explainable protein-protein binding affinity prediction via fine-tuning protein language models

This paper introduces a scalable, explainable, and data-efficient framework that reframes protein-protein binding affinity prediction as a metric learning task using fine-tuned protein language models, achieving state-of-the-art performance from sequence alone without requiring 3D structural inputs.

The Gist

The Big Problem: The "Lock and Key" Mystery

Imagine your body is a giant city where proteins are the workers. To get things done, these workers need to grab onto each other. Sometimes, they hold hands tightly (strong binding); other times, they just give a quick high-five (weak binding).

Scientists want to predict how tightly two proteins will hold hands just by looking at their "ID cards" (their amino acid sequences). This is crucial for designing new medicines, like antibodies that can grab onto a virus and stop it.

The Old Way:
Previously, scientists tried to solve this by building a 3D model of the proteins, like sculpting them out of clay. They would measure the distance between every atom to see how well they fit.

• The Problem: This is slow, expensive, and requires you to already have the 3D blueprint. But often, we don't have the blueprint yet! We only have the ID card (the sequence).

The New Solution: BALM-PPI

The authors of this paper created a new tool called BALM-PPI. Think of it as a "Compatibility Dating App" for proteins that works without needing 3D blueprints.

Here is how it works, broken down into three simple concepts:

1. The "Shared Language" (Metric Learning)

Imagine you have two people who speak different languages. Usually, to see if they get along, you need a translator.

• The Old Way: You would translate both of them into English, then compare their sentences.
• The BALM-PPI Way: Instead of translating them, you teach both of them a secret, shared language (a "latent space").
  • In this secret language, if two proteins "speak" in a way that is very similar (high cosine similarity), it means they will hold hands tightly in real life.
  • If their secret languages are very different, they won't stick together.
  • The Magic: The model learns that "similarity in this secret language = strong physical bond."

2. The "Fine-Tuning" (PEFT & LoRA)

The model uses a giant, pre-trained brain called ESM-2. This brain has read millions of protein books and knows the general rules of biology.

• The Problem: The brain is too big to retrain from scratch for every new job. It's like trying to re-teach a master chef how to cook every time you want a new recipe.
• The Solution (LoRA): Instead of retraining the whole chef, the authors attach a tiny, lightweight "training headset" (Low-Rank Adaptation) to the chef's ear.
  • This headset only changes a tiny fraction of the chef's brain (less than 1%).
  • It teaches the chef just enough to specialize in "protein dating" without forgetting everything else it knows.
  • Result: It's fast, cheap, and incredibly efficient. You can teach it a new specific job with very little data (like showing it just 30% of the usual examples).

3. The "Why" (Explainability)

Most AI models are "black boxes." You give them input, and they give an answer, but you don't know why.

• The Innovation: BALM-PPI comes with a "Highlighter".
• When it predicts that Protein A and Protein B will stick together, it can point to the exact amino acids (the letters in the sequence) that are responsible.
• The Metaphor: It's like a detective looking at a crime scene. Instead of just saying "The suspect did it," it points to the specific fingerprints on the door handle.
• Why this matters: Scientists can look at these highlighted spots and say, "Ah, the model thinks these two specific parts are the 'glue.' That makes sense biologically!" This builds trust so they can use the AI to design real drugs.

What Did They Prove?

The team tested this tool on some very tough challenges:

1. The "Stranger" Test: They tested it on proteins that are evolutionarily very different (like a human protein and a bacteria protein). Even though they look nothing alike, the model still guessed the bond strength correctly.
2. The "Data-Starved" Test: They gave the model very little data to learn from (just 30% of the usual amount). Even with this little info, it outperformed other models that had seen 90% of the data.
3. The "No-Blueprint" Test: It worked perfectly without ever seeing a 3D structure, proving you don't need the clay sculpture to know if the lock and key fit.

The Real-World Impact

Imagine you are a drug designer trying to stop a new virus.

• Before: You had to wait months to get 3D structures, then run slow simulations.
• With BALM-PPI: You type in the virus's sequence and your antibody's sequence. In seconds, the AI tells you: "These two will stick together very well." It also highlights exactly which parts of the antibody you should tweak to make it even stronger.

Summary

BALM-PPI is a smart, efficient, and transparent tool that predicts how well proteins stick together using only their text sequences. It learns a "secret language" of binding, uses a tiny "headset" to specialize quickly, and acts like a detective to show you exactly why it made its prediction. It turns a slow, complex scientific puzzle into a fast, accessible workflow for saving lives.

Technical Summary

1. Problem Statement

Predicting protein-protein binding affinity (typically expressed as pKd) is critical for antibody optimization, biologics design, and understanding biological pathways.

• Limitations of Structure-Based Methods: While physics-based methods (e.g., Rosetta, FoldX) and deep learning models using 3D structures (e.g., AlphaFold-Multimer) achieve high accuracy, they are not scalable. They require atomic-level 3D structural inputs, which are often unavailable or unreliable for de novo design and large-scale screening.
• Limitations of Existing Sequence-Based Methods: Current sequence-only approaches often rely on concatenating embeddings from Protein Language Models (PLMs) and passing them through a regressor. These methods struggle with generalization to evolutionarily distant proteins, require massive amounts of labeled data, and lack interpretability (residue-level explainability).
• The Gap: There is a need for a scalable, data-efficient, and explainable framework that predicts binding affinity using sequence alone, generalizes across distribution shifts (e.g., new assays or antigens), and provides residue-level rationales for predictions.

2. Methodology: The BALM-PPI Framework

The authors propose BALM-PPI, a framework that reframes affinity prediction as a metric learning problem rather than a standard regression task.

Core Architecture

• Backbone: The model utilizes ESM-2 (a large-scale Protein Language Model) as the encoder to generate sequence embeddings.
• Metric Learning Objective: Instead of concatenating embeddings, BALM-PPI projects the two interacting proteins (Target and Binder) into a shared latent space.
  • Two independent projection layers transform the ESM-2 embeddings into a lower-dimensional space (256 dimensions).
  • The embeddings are L2-normalized.
  • Cosine Similarity between the two projected vectors is calculated. This similarity score is directly mapped to the experimental binding affinity (pKd).
  • Hypothesis: In this shared space, higher cosine similarity correlates directly with stronger binding affinity.

Parameter-Efficient Fine-Tuning (PEFT)

To adapt the pre-trained ESM-2 model without catastrophic forgetting or high computational cost:

• LoRA (Low-Rank Adaptation): The authors inject trainable low-rank matrices into the Query (Q), Key (K), and Value (V) projection layers of the frozen ESM-2 attention blocks.
• Efficiency: Only <1% of the total parameters (approx. 0.31% or ~2 million parameters) are trained. The vast majority of the pre-trained weights remain frozen.
• Benefit: This allows the model to specialize for the affinity prediction task while retaining general protein knowledge, enabling rapid "warm-up" on new datasets with minimal data.

Explainability Mechanism

• Integrated Gradients (IG): The framework employs IG to compute residue-level attributions.
• Process: By integrating gradients from a baseline (zero embedding) to the actual input, the model identifies which specific amino acid residues contribute most to the predicted affinity.
• Visualization: Results are visualized as heatmaps on 3D structures and residue strips, highlighting "interaction hotspots."

3. Key Contributions

1. Metric Learning for Affinity: Successfully reframed binding affinity prediction as a cosine similarity task in a shared latent space, outperforming standard concatenation-based regression baselines.
2. Data Efficiency via Few-Shot Learning: Demonstrated that PEFT (LoRA) allows the model to adapt to new, strictly de-overlapped datasets (e.g., AB-Bind) using only 10–30% of the labeled data, outperforming models trained on 90% of the data.
3. Robust Generalization: Achieved strong performance on evolutionarily distant proteins (<30% sequence identity) and across diverse biological subgroups (Antibody-Antigen, TCR-pMHC) without 3D structural inputs.
4. Residue-Level Explainability: Validated that the model's attention aligns with experimentally known interaction hotspots (e.g., electrostatic anchors, hydrophobic patches) across enzyme-inhibitor and antibody-antigen systems.
5. Open-Source Tool: Released an interactive web server (BALM-PPI·predict) and open-source code, allowing users to input sequences and receive pKd predictions with 3D visualizations of attribution scores.

4. Key Results

Performance on PPB-Affinity Benchmark

• Random Split: Achieved Pearson $r = 0.89$ and RMSE = 0.994.
• Cold Split (Unseen PDBs): Maintained Pearson $r = 0.73$, outperforming regression baselines.
• Sequence-Similarity Split (<30% Identity): Achieved Pearson $r = 0.61$, demonstrating the ability to generalize to evolutionarily distant proteins where structure-based methods often fail or are inapplicable.
• Comparison: Consistently outperformed structure-based deep learning baselines across biological subgroups (TCR-pMHC, Antibody-Antigen) and data sources (SKEMPI, SAbDab, PDBbind).

Few-Shot Adaptation (AB-Bind & AbBiBench)

• AB-Bind Dataset: In a strictly de-overlapped setting (zero-shot transfer initially failed with negative correlation), few-shot adaptation with only 30% of the data achieved Pearson $r = 0.756$ and RMSE = 0.688. This surpassed the MVSF-AB model (trained on 90% of non-de-overlapped data) which achieved $r = 0.739$ and RMSE = 1.905.
• AbBiBench Assays: Across nine deep-mutational scanning assays, 10–30% labeled variants were sufficient to achieve strong positive correlations, with significant gains in influenza hemagglutinin assays (Pearson rising from ~0.3 to >0.95).

Explainability Validation

• Hotspot Identification: The model correctly identified known energetic hotspots in zero-shot settings for diverse complexes:
  • Barnase-Barstar: Identified electrostatic residues (Asp, Arg, Trp).
  • MDM2-p53: Highlighted hydrophobic anchors (Phe, Trp).
  • Antibody-Antigen (CR6261, CR9114): Focused on CDR loops and conserved stem epitopes.
• Refinement: Few-shot adaptation sharpened the attribution scores, increasing contrast at specific interface residues.

5. Significance and Impact

• Scalability: By removing the dependency on 3D structures, BALM-PPI enables high-throughput screening of protein interactions and antibody variants that would be impossible with structure-based methods.
• Therapeutic Design: The framework offers a practical workflow for therapeutic antibody optimization: calibrate with a small set of project-specific data, predict variants, and use residue-level attributions to verify biophysical plausibility before costly experimental validation.
• Interpretability: Unlike "black box" deep learning models, BALM-PPI provides residue-level rationales that align with structural biology, building trust for lead optimization campaigns.
• Data Efficiency: The success of few-shot learning suggests that PLMs encode reusable interaction signals, requiring only minor calibration for new assay landscapes, significantly reducing the experimental burden in drug discovery.

In conclusion, BALM-PPI establishes a new standard for sequence-only binding affinity prediction, combining the accuracy of metric learning, the efficiency of PEFT, and the interpretability required for real-world biological applications.