ANTIPASTI: interpretable prediction of antibody binding affinity exploiting Normal Modes and Deep Learning

Kevin Michalewicz, Mauricio Barahona, Barbara Bravi

TL;DR

ANTIPASTI is a new deep learning framework that predicts antibody-antigen binding affinity by treating protein structures as networks of vibrating "balls and springs." By combining Normal Mode Analysis (NMA) with Convolutional Neural Networks (CNNs), the model achieves state-of-the-art accuracy while remaining highly interpretable, revealing that long-range structural synergies are the secret sauce behind high-affinity binding.

What Problem Does This Paper Address?

Antibodies are the "guided missiles" of the immune system, but designing them to bind tightly to a specific target (antigen) is a massive computational challenge. The strength of this bond is measured by binding affinity ($K_D$).

Current prediction methods usually fall into two camps: sequence-based methods are fast but miss the spatial "synergy" of a 3D structure, while structure-based methods are often computationally heavy or rely on "handcrafted" features that ignore the protein's dynamic flexibility. There is a critical need for a model that can look at the whole 3D structure, account for how it moves and vibrates, and explain why it predicts a certain affinity.

Methods

The authors introduce ANTIPASTI (ANTIbody Predictor of Affinity from STructural Information). The workflow is a clever blend of physics and AI:

  1. Elastic Network Models (ENM): They represent the antibody-antigen complex as a network of residues connected by harmonic springs.
  2. Normal Mode Analysis (NMA): Instead of running expensive molecular dynamics simulations, they use NMA to calculate how these residues vibrate and fluctuate in relation to one another.
  3. The "Antigen Imprint": They calculate a correlation map for the whole complex but then remove the antigen residues. This leaves an "imprint" on the antibody's correlation map—essentially capturing how the presence of the antigen changes the antibody's internal dynamics.
  4. CNN Architecture: This 2D correlation map is fed into a Convolutional Neural Network. Because the map is "image-like," the CNN can efficiently learn structural motifs and energetic patterns associated with high affinity.

Datasets

The model was trained and validated using:

  • SAbDab (Structural Antibody Database): 634 antibody-antigen structures with experimentally determined $K_D$ values.
  • AlphaFold Predictions: A separate test set of 21 recently published structures was used to see if the model could work with AI-predicted structures instead of just experimental "gold-standard" PDB files.

Key Insights and Ideas

  • Dynamics > Static Geometry: Normal Mode correlation maps proved much more informative than simple contact maps. It’s not just about which residues touch, but how they "wiggle" together.
  • The Power of Low Frequencies: The model performs best when using low-frequency modes (global fluctuations). This suggests that binding affinity is driven by large-scale, long-range structural movements rather than just local "jitter."
  • Latent Clustering: Even though the model only "sees" the antibody residues (with the antigen imprint), it naturally clusters antibodies by their target type (e.g., proteins vs. haptens) in its internal representations.
  • Long-Range Synergy: A staggering 80% of the most important correlations identified by the model were "long-range" (residues more than 10Å apart). This highlights that high affinity is a team effort across the entire antibody variable region, not just the tips of the CDR loops.

Key Results

  • State-of-the-Art Performance: ANTIPASTI achieved a Pearson correlation of $R = 0.86$ ($0.85$ mean across splits) on experimental data, outperforming existing sequence-based ($R \approx 0.45-0.66$) and structure-based ($R \approx 0.69-0.79$) predictors.
  • Efficiency: Despite using complex structural data, the model is incredibly lightweight, training in under 10 minutes on a standard CPU.
  • AlphaFold Compatibility: The model maintained a high correlation ($R = 0.9$) when using AlphaFold-predicted structures, though performance dropped significantly if AlphaFold was uncertain about the binding interface (high PAE/low pLDDT).

Limitations and Future Work

The authors identify a few hurdles for the future:

  • Data Scarcity: 634 structures is a small dataset for deep learning. As more structural data becomes available, the model's performance is expected to improve.
  • The Constant Region: Most PDB files only include the "variable" part of the antibody. Including the "constant" region could reveal even more long-range allosteric effects, but data is currently lacking.
  • Interface Accuracy: The model is sensitive to the quality of the 3D structure. If AlphaFold fails to predict the exact "docking" interface correctly, ANTIPASTI’s affinity prediction suffers.

Future work may involve integrating "docking" simulations to better handle cases where the exact bound structure isn't known beforehand.

Summary updated

Abstract

The high binding affinity of antibodies toward their cognate targets is key to eliciting effective immune responses, as well as to the use of antibodies as research and therapeutic tools. Here, we propose ANTIPASTI, a convolutional neural network model that achieves state-of-the-art performance in the prediction of antibody binding affinity using as input a representation of antibody-antigen structures in terms of normal mode correlation maps derived from elastic network models. This representation captures not only structural features but energetic patterns of local and global residue fluctuations. The learnt representations are interpretable: they reveal similarities of binding patterns among antibodies targeting the same antigen type, and can be used to quantify the importance of antibody regions contributing to binding affinity. Our results show the importance of the antigen imprint in the normal mode landscape, and the dominance of cooperative effects and long-range correlations between antibody regions to determine binding affinity.
BibTeX
@inproceedings{michalewicz2025antipasti,
  author = {Kevin Michalewicz and Mauricio Barahona and Barbara Bravi},
  title = {{ANTIPASTI: interpretable prediction of antibody binding affinity exploiting Normal Modes and Deep Learning}},
  booktitle = {AI4X},
  year = {2025},
  url = {https://openreview.net/forum?id=jpreMDvNOa},
}
Figure