ANTIPASTI: Interpretable Prediction of Antibody Binding Affinity Exploiting Normal Modes and Deep Learning

Kevin Michalewicz, Mauricio Barahona, Barbara Bravi

Introduction

Antibodies are critical components of the immune system, functioning as precise "guided missiles" that target specific antigens. Designing antibodies with high binding affinity ($K_D$) is a cornerstone of therapeutic development, yet it remains a significant computational challenge. Existing prediction methods typically fall into two categories: sequence-based approaches, which are computationally efficient but often overlook the 3D spatial synergies of protein structures, and structure-based methods, which can be computationally intensive or rely on static, "handcrafted" features that ignore the inherent flexibility of proteins. There is a pressing need for models that integrate 3D structural dynamics with high-performance predictive capabilities while maintaining interpretability.

Methodology

The authors present ANTIPASTI (ANTIbody Predictor of Affinity from STructural Information), a framework that conceptualizes protein complexes as networks of vibrating "balls and springs." The workflow integrates physics-based modeling with deep learning:

  1. Elastic Network Models (ENM): The antibody-antigen complex is represented as a network of residues connected by harmonic springs.
  2. Normal Mode Analysis (NMA): Instead of costly molecular dynamics simulations, NMA is employed to calculate the collective vibrations and fluctuations of the residues.
  3. The "Antigen Imprint": A correlation map is generated for the entire complex. The antigen residues are then removed, leaving an "imprint" on the antibody's correlation map that captures how the antigen's presence alters the antibody's internal dynamics.
  4. CNN Architecture: This 2D "image-like" correlation map is processed by a Convolutional Neural Network (CNN) to learn structural motifs and energetic patterns associated with binding affinity.
Overview of the ANTIPASTI framework, illustrating the transition from 3D structure to Elastic Network, Normal Mode correlation maps, and finally the CNN-based affinity prediction.

Results

ANTIPASTI achieves state-of-the-art performance, reaching a Pearson correlation of $R = 0.86$ on experimental data from the SAbDab database. This significantly outperforms standard sequence-based ($R \approx 0.45-0.66$) and structure-based ($R \approx 0.69-0.79$) predictors. Notably, the model is highly efficient, training in under 10 minutes on a standard CPU.

The framework also demonstrates robustness when using AI-predicted structures. On a test set of 21 recently published structures generated by AlphaFold, the model maintained a high correlation ($R = 0.9$), provided the predicted binding interface was accurate (high pLDDT/low PAE).

Performance evaluation of ANTIPASTI. (a) Correlation between true and predicted binding affinity ($K_D$). (c) Comparison of R scores across different models, highlighting the superior performance of Normal Mode-based features over simple contact maps.

Discussion and Key Insights

A central finding of this work is that dynamics are more informative than static geometry. Normal Mode correlation maps provide a richer representation of binding than simple contact maps, as they capture how residues "wiggle" together. The model's performance peaks when using low-frequency modes, suggesting that binding affinity is primarily driven by large-scale, global structural fluctuations rather than local atomic "jitter."

Furthermore, the model reveals the importance of long-range structural synergy. Approximately 80% of the most critical correlations identified by the model involved residues separated by more than 10Å. This indicates that high-affinity binding is a collective property of the entire antibody variable region, rather than being localized solely at the CDR loop tips. Interestingly, the model's internal representations naturally cluster antibodies by target type (e.g., proteins vs. haptens), even though it only "sees" the antibody residues via the antigen imprint.

Limitations and Future Directions

Despite its success, ANTIPASTI faces challenges related to data scarcity, as the current training set is limited to 634 structures. Additionally, the model's performance is sensitive to the quality of the input 3D structure; inaccuracies in the predicted docking interface can lead to poor affinity predictions. Future work will explore the inclusion of antibody constant regions to capture broader allosteric effects and the integration of docking simulations to improve predictions when bound structures are unavailable.

Summary updated

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