InteractBind

A physically grounded, large-scale protein–ligand interaction dataset
for interpretable and interaction-aware binding prediction

---

Links

• Paper: A Large-Scale Dataset and Benchmark: Do Protein-Ligand Models Learn Binding Sites or Just Binding Likelihood?
• Github: ZhaohanM/InteractBind

---

Motivation

Most existing protein–ligand binding datasets provide only coarse-grained supervision, such as binary labels or scalar affinity values. While effective for prediction, these signals compress complex molecular interaction processes into a single outcome, limiting interpretability and mechanistic understanding.

InteractBind addresses this limitation by explicitly modelling non-covalent interaction patterns derived from experimentally resolved protein–ligand complexes.

It enables token-level supervision, bridging sequence-based representations with physically meaningful interaction structures.

---

Dataset Overview

InteractBind is constructed from high-quality experimentally resolved complexes and includes:

• Protein sequences (FASTA and structure-aware sequence)
• Ligand molecular representations (SMILES and SELFIES)
• Binding labels and affinity annotations
• Token-level non-covalent interaction maps

The dataset is designed to support both prediction accuracy and mechanistic interpretability.

---

Dataset

This repository provides benchmark CSVs with true residue-level interaction maps for PLI prediction evaluation.

Files

The Hugging Face Dataset Viewer is configured to read the CSV subsets under data/:

• affinity: the full InteractBind affinity table.
• p_ood_25, p_ood_28, p_ood_31, p_ood_33: protein OOD benchmark subsets with train, validation, and test splits.

Each CSV includes seven residue-level binding-site fingerprint columns derived from the interaction maps:

• Hydrogen bonding_binding_site
• Salt Bridges_binding_site
• π–π Stacking_binding_site
• Cation–π_binding_site
• Hydrophobic_binding_site
• Van der Waals_binding_site
• Overall_binding_site

Each value is a binary list aligned to the protein FASTA sequence. For example, [0,0,1,0] marks the third residue as a binding-site residue. Negative protein-ligand pairs without contact-map entries are encoded as all-zero fingerprints.

Supported Interaction Types

Structured annotations are provided for major non-covalent interaction categories:

• Hydrogen bonds
• Hydrophobic interactions
• Salt bridges
• π–π stacking
• π–cation interactions
• Van der Waals contacts

Each interaction channel can be used independently or combined for multi-channel supervision.

---

Key Features

• Physically grounded supervision
  Derived from experimentally resolved complexes rather than heuristic attention signals.

• Token-level interaction maps
  Enables fine-grained modelling of residue–atom interactions.

• Model-agnostic integration
  Compatible with sequence-based encoders (e.g., ESM, SELFormer, and other protein–ligand models).

• Interpretability support
  Facilitates binding residue identification and interaction pattern analysis.

• Scalable design
  Allows large-scale training without requiring full structural modelling during inference.

---

Research Applications

InteractBind supports a broad range of research directions:

• Protein–ligand binding prediction
• Binding site/pocket localisation
• Interaction-aware representation learning
• Mechanistic hypothesis generation
• Drug discovery and virtual screening
• Explainable AI for molecular modelling

---

Citation

If you use InteractBind in your research, please cite:

@misc{meng2026largescaleinteractbind,
  title = {A Large-Scale Dataset and Benchmark: Do Protein-Ligand Models Learn Binding Sites or Just Binding Likelihood?},
  author = {Meng, Zhaohan and Bai, Zhen and Yuan, Ke and Ounis, Iadh and Meng, Zaiqiao and Xu, Hao and Loscalzo, Joseph},
  year = {2026},
  eprint = {2605.24045},
  archivePrefix = {arXiv},
  primaryClass = {cs.LG},
  url = {https://arxiv.org/abs/2605.24045}
}