Pre-training on noncovalent interactions in synthetic protein-ligand structures to better predict binding affinity

Arun Raja, Garrett A. Stevenson, Garrett M Morris, Jonathan E Allen, Hyojin Kim

TL;DR

The PLIP (Protein-Ligand Interaction Pre-training) framework significantly improves binding affinity predictions by pre-training a heterogeneous graph transformer on 5.1 million synthetic protein-ligand structures. By explicitly modeling aromatic rings and specific noncovalent interactions (like $\pi$-$\pi$ stacking), PLIP captures the complex chemical hierarchy that standard atom-level models often ignore.

What Problem Does This Paper Address?

In computational drug discovery, predicting how tightly a small molecule (ligand) binds to a target protein—its binding affinity—is the "holy grail." While Graph Neural Networks (GNNs) have become the standard tool for this, they typically treat protein-ligand complexes as "flat" collections of atoms.

This approach has two major flaws:

  1. Ignoring Hierarchy: It neglects the crucial role of aromatic rings and the specific types of noncovalent interactions (e.g., hydrogen bonds, cation-$\pi$ contacts) that actually drive binding.
  2. Data Scarcity: High-quality experimental 3D structures are rare. Most existing models pre-train on 2D molecular graphs or SMILES strings, which lack the spatial context necessary to understand 3D binding physics.

Methods

The authors introduce PLIP, a heterogeneous equivariant graph transformer designed to "see" the 3D world more like a medicinal chemist does.

  • Heterogeneous Graph Representation: Instead of just atoms, PLIP uses four node types: ligand atoms, protein atoms, ligand rings, and protein rings.
  • Interaction-Specific Edges: Edges aren't just distances; they are labeled with specific interaction types (hydrophobic, salt bridges, $\pi$-stacking, etc.) identified via the Open Drug Discovery Toolkit (ODDT).
  • Atomic Environment Vectors (AEV): To capture the local 3D neighborhood, the model uses a 352-dimensional descriptor based on radial symmetry functions, moving beyond simple one-hot encodings of element types.
  • Multi-Task Pre-training: The model is pre-trained on three simultaneous objectives:
  • Classification: Identifying the type of noncovalent interaction.
  • Regression: Predicting interatomic distances.
  • Affinity: Predicting the binding strength (pIC50).
  • PLEC-Butina Splitting: To ensure rigorous evaluation, the authors used a novel splitting strategy based on interaction fingerprints. This prevents "data leakage" where the model might simply memorize similar binding motifs seen during training.

Datasets

  • Pre-training: The authors utilized the Structural and Interaction Repository (SAIR), containing 5.1 million synthetic structures generated by the Boltz-1x AI model. They tested three subsets: "All," "Best" (highest confidence), and "Best Positive" (moderate-to-high affinity).
  • Evaluation: The model was benchmarked on four diverse drug targets:
  • AChE (Alzheimer’s target)
    • SARS-Mpro (COVID-19 protease)
    • Zika protease
    • $\mu$-Opioid receptor (Pain management)

Key Insights and Ideas

  • Rings are First-Class Citizens: By treating aromatic rings as explicit nodes, the model gains a biologically motivated inductive bias. This is particularly effective for capturing $\pi$-$\pi$ stacking, which is thermodynamically vital but geometrically complex to model with atom-to-atom messages alone.
  • Synthetic Data is a Valid Teacher: One of the most significant findings is that pre-training on AI-generated structures (from Boltz-1x) translates effectively to experimental downstream tasks. This suggests we can overcome the experimental data bottleneck using high-quality structural simulations.
  • The "Why" Helps the "How Much": Learning to classify interaction types (the "why" of binding) significantly boosts the model's ability to predict the affinity (the "how much").
  • Robustness to Small Data: Pre-training reduced the model's sensitivity to the size of the downstream dataset, showing particularly strong gains on the smallest dataset ($\mu$-opioid receptor).

Key Results

PLIP consistently outperformed training from scratch and surpassed established baselines like EIGN and PLAIG on three out of four targets.

  • AChE: Achieved a Pearson correlation of 0.758 (+26% improvement over training from scratch).
  • SARS-Mpro: Achieved 0.255 (+39% improvement).
  • Zika: Achieved 0.490 (+6% improvement).
  • $\mu$-Opioid: Achieved 0.374 (+32% improvement).

The authors noted that while PLIP was slightly behind EIGN on the SARS-Mpro target, it exhibited much lower variance across different data folds, indicating a more stable and reliable model.

Limitations and Future Work

  • Narrow Dynamic Range: The model struggled slightly with SARS-Mpro, which has a very narrow range of affinity values in the dataset, making correlation-based metrics difficult.
  • Scaling: While 5.1 million structures is large, the authors suggest scaling even further while focusing more on the quality and diversity of the synthetic structures rather than just the raw count.
  • Generalization: Future work will focus on "out-of-distribution" generalization—testing how well the model performs on entirely new protein families it hasn't encountered during pre-training.

Summary updated

Abstract

Accurate prediction of protein-ligand binding affinity is a central challenge in computational drug discovery, yet existing graph neural network approaches typically represent protein-ligand complexes as homogeneous atom-level graphs, neglecting the role of aromatic ring systems and the rich hierarchy of noncovalent interactions that govern binding. In this work, we introduce the Protein-Ligand Interaction Pre-training or PLIP approach, a heterogeneous equivariant graph transformer that explicitly encodes four node types - ligand atoms, protein atoms, ligand rings, and protein rings connected by interaction- specific edge relations such as hydrogen bonds, hydrophobic contacts, π-π stacking, and cation-π interactions. We pretrain on interactions found in 5.1 million synthetic protein-ligand structures from the Structural and Interaction Repository (SAIR) using a multi-task self-supervised objective comprising interaction-type classification, interatomic distance regression, and binding affinity prediction. Systematic evaluation on four drug targets - acetylcholinesterase (AChE), SARS- CoV-2 main protease (SARS-Mpro), Zika protease, and μ-opioid receptor - demonstrates that multi-task pretraining on all three pretraining objectives (interaction type, distance and affinity) achieves Pearson correlations of 0.667 on AChE (+26% over training from scratch), 0.490 on Zika (+6%), 0.374 on μ-opioid receptor (+32%), and 0.255 on SARS-Mpro (+39%). Comparisons against competing baselines demonstrate the benefits of pre-training on protein-ligand interactions for structure-based binding affinity prediction.
BibTeX
@inproceedings{raja2026pre,
  author = {Arun Raja and Garrett A. Stevenson and Garrett M Morris and Jonathan E Allen and Hyojin Kim},
  title = {{Pre-training on noncovalent interactions in synthetic protein-ligand structures to better predict binding affinity}},
  booktitle = {AI4Science @ ICML},
  year = {2026},
  url = {https://openreview.net/forum?id=h8yARiHZeW},
}
Figure