Graph-Based Cross-Modal Learning for Drug–Target Affinity Prediction

Aakriti Attri, Jashn Khemani, Arush Gumber, Anubha GuptaAI4Science @ ICML 2026 · OpenReview

Predicting how strongly a drug molecule binds to a target protein—known as Drug-Target Affinity (DTA)—is the "holy grail" of early-stage drug discovery. While deep learning has made massive strides in this area, most models treat the drug and the protein like two people in separate rooms trying to describe each other through a wall.

A new paper, "Graph-Based Cross-Modal Learning for Drug–Target Affinity Prediction," introduces HeteroBindNet, a framework that finally lets these two modalities "talk" to each other at the atomic level.

TL;DR

HeteroBindNet is a heterogeneous graph neural network that predicts drug-target affinity by explicitly modeling the interaction between individual drug atoms and protein residues. By using a novel cross-modal attention mechanism, it outperforms existing baselines in "cold-start" scenarios (unseen drugs/targets) and identifies binding sites with high accuracy—even without being trained on 3D structural data.

What Problem Does This Paper Address?

In the world of drug discovery, we usually have two pieces of information: the chemical structure of the drug (often a 2D graph) and the amino acid sequence of the protein.

The standard deep learning approach is to encode the drug and the protein independently into two separate vectors, concatenate them at the very end, and run a regression to guess the affinity score. The authors argue this is a fundamental limitation. Real-world binding is a "handshake"—a local, physicochemical interaction where specific atoms meet specific residues. Existing models lack this explicit cross-modal modeling, making them less effective at generalizing to entirely new drugs or targets (the "cold-start" problem) and leaving them as "black boxes" that offer little insight into why a drug binds.

Methods

HeteroBindNet moves away from simple concatenation toward a structured communication approach. Its architecture is a multi-pronged ensemble:

  1. Drug Backbone: Uses a GINE (Graph Isomorphism Network with Edge features) to capture local molecular topology, supplemented by Morgan fingerprints for global chemical context.
  2. Protein Backbone: Combines ESM-2 (evolutionary scale language model) embeddings with a GCN (Graph Convolutional Network) operating over predicted residue contact maps. It also uses a multi-scale 1D CNN to capture sequential motifs.
  3. The HeteroBindNet Module (The Innovation): This is the "cross-modal" heart of the model. It projects drug atoms and protein residues into a shared 64-dimensional alignment space. It then computes a temperature-gated interaction matrix—essentially a heat map of which atoms are likely interacting with which residues.
  4. Bidirectional Message Passing: Using this matrix, the model performs "relational message passing," allowing the drug's representation to be updated by the protein's features and vice versa before the final prediction is made.

Datasets

The authors evaluated the model on two gold-standard benchmarks:

  • Davis: A smaller, sparser dataset (68 drugs, 442 proteins) focused on kinase inhibitors.
  • KIBA: A larger, denser dataset (2,111 drugs, 229 proteins) that integrates multiple bioactivity measurements into a single score.

Crucially, they tested these using "Cold Splits" (Cold-Drug, Cold-Target, and Cold-Both), where the test set contains molecules or proteins never seen during training.

Key Insights and Ideas

  • Emergent Interpretability: Perhaps the most striking finding is that the model learns to identify binding sites "for free." Even though it was only trained on affinity scores (not 3D coordinates), the top 5% of entries in its learned interaction matrix consistently align with known binding regions.
  • Division of Labor: Through ablation studies, the authors found that the cross-modal bonds are responsible for "ranking consistency" (getting the relative order of drugs right), while global auxiliary features (like fingerprints) are necessary for "point-wise accuracy" (getting the exact affinity number right).
  • Sequence-Driven, Structure-Aware: By using ESM-2 to predict contact maps, the model gains the benefits of structural awareness without requiring expensive, experimentally resolved 3D crystal structures, which are unavailable for many proteins.

Key Results

HeteroBindNet demonstrated superior generalization, particularly in the most difficult scenarios:

  • KIBA Cold-Drug Split: It achieved a 14.95% improvement in Mean Squared Error (MSE) and a 5.05% improvement in Concordance Index (CI) over the state-of-the-art DMFF-DTA baseline.
  • Davis Benchmark: It attained the lowest MSE (0.176) among all compared methods, including those that use specialized 3D structural data.
  • Structural Validation: In a test of five protein-ligand complexes, the model recovered ~92% of real atomic contacts at a permissive threshold, proving the interaction matrix isn't just noise—it’s capturing biology.

Limitations and Future Work

While HeteroBindNet excels on KIBA, the authors noted that the Davis Cold-Drug setting remains a challenge. Because Davis is small and has low affinity variance, the model’s MSE is sensitive to outliers.

Future Directions: * 3D Integration: Moving beyond 2D graphs to include explicit 3D structural graphs and binding pocket geometries. * Scaling Up: Utilizing even larger protein language models. * Broadening Scope: Extending the cross-modal architecture to other biological interactions, such as protein-protein or RNA-ligand binding, where the "handshake" between different modalities is equally critical.

Summary updated

Abstract

Drug-target binding affinity prediction is a fundamental task in structure-based drug discovery. But, in general, sequence- and structure-based deep learning methods lack explicit modeling of the physicochemical interactions between drug atoms and protein residues at the binding site. We present HeteroBindNet, a heterogeneous graph neural network that combines 1) GINE-based convolutions over RDKit-derived atomic graphs, 2) GCN-based protein contact graph learning over ESM-2 contact maps, 3) auxiliary global features enriched by Morgan fingerprints, 4) multi-scale 1-D CNN sequence encoders within a shared embedding space, and 5) a novel cross-modal HeteroBindNet which learns an atom-residue interaction matrix through scaled, temperature-gated attention and bidirectional gated message passing, capturing the local structural complementarity that governs binding. Even without any binding-site supervision, this interaction matrix is able to identify a subset of ligand atoms and protein residues (top 5%) that are enriched near known binding regions, consistent with established pharmacophoric contacts. Across multiple evaluation splits, HeteroBindNet achieves a CI of 0.791 and MSE of 0.347 on the KIBA benchmark cold-drug split, outperforming the baseline model, DMFF-DTA, by 14.95% on MSE and improving CI by 5.05%, with consistent performance on the Davis benchmark and strong generalization in cold-start scenarios, demonstrating its utility for real-world drug discovery applications.
BibTeX
@inproceedings{attri2026graph,
  author = {Aakriti Attri and Jashn Khemani and Arush Gumber and Anubha Gupta},
  title = {{Graph-Based Cross-Modal Learning for Drug–Target Affinity Prediction}},
  booktitle = {AI4Science @ ICML},
  year = {2026},
  url = {https://openreview.net/forum?id=aakl14l0sJ},
}
Figure