🤖 machine learning

h-MINT: Modeling Pocket-Ligand Binding with Hierarchical Molecular Interaction Network

The paper introduces h-MINT, a hierarchical molecular interaction network that utilizes a novel "OverlapBPE" tokenization method to capture complex, overlapping chemical contexts at both the atom and fragment levels, significantly improving the accuracy of pocket-ligand binding affinity predictions and virtual screening.

Original authors: Yanru Qu, Yijie Zhang, Wenjuan Tan, Xiangzhe Kong, Xiangxin Zhou, Chaoran Cheng, Mathieu Blanchette, Jiaxuan You, Ge Liu

Published 2026-04-28

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Yanru Qu, Yijie Zhang, Wenjuan Tan, Xiangzhe Kong, Xiangxin Zhou, Chaoran Cheng, Mathieu Blanchette, Jiaxuan You, Ge Liu

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to teach a computer how to play a high-stakes game of "Molecular Tetris."

In this game, the computer’s goal is to predict how perfectly a tiny "key" (a drug molecule) will fit into a specific "lock" (a protein in your body). If the fit is perfect, the drug works; if it’s a bad fit, the drug fails or causes side effects.

For years, scientists have been trying to teach computers how to "see" these molecules. This paper, H-MINT, introduces a much smarter way for the computer to look at the world.

Here is the breakdown of how they did it, using three simple analogies.

1. The Problem: The "Lego" vs. "Dust" Dilemma

Until now, computers usually looked at molecules in one of two ways:

The "Dust" Approach (Atom-level): The computer sees the molecule as a giant pile of individual atoms. It’s like trying to understand a beautiful painting by looking at it through a microscope that only shows individual grains of pigment. You see the colors, but you miss the "face" in the portrait.
The "Lego" Approach (Fragment-level): The computer sees the molecule as pre-made Lego chunks. This is faster, but it’s rigid. If you have a special piece that is half-red and half-blue, the old way forces you to pick one or the other, breaking the "magic" of the piece.

The result? The computer misses the subtle "vibes" of the molecule—like its electrical charge or its specific 3D shape (chirality)—which are the very things that make a drug work.

2. The Solution Part 1: OverlapBPE (The "Smart Puzzle" Method)

The researchers created a new way to break molecules down called OverlapBPE.

Imagine you have a photograph of a person. Instead of cutting it into rigid squares (where a person's eye might be split awkwardly between two squares), you use overlapping translucent slides. One slide might focus on the eyes, another on the nose, and another on the mouth. Because the slides overlap, the "eye" is captured perfectly in its context, and the "nose" is captured in its context.

By allowing these "fragments" to overlap, the computer can see the "Aromatic Rings" (the sturdy, circular parts of a molecule) and the "Ionic Charges" (the electrical magnetism) without breaking them apart. It sees the "soul" of the chemical structure, not just a pile of parts.

3. The Solution Part 2: h-MINT (The "Dual-Lens" Microscope)

Once they had these smart, overlapping pieces, they needed a brain to process them. They built h-MINT.

Think of h-MINT as a scientist looking through a dual-lens microscope:

Lens 1 (The Atom Lens): Sees the tiny, granular details (the individual atoms).
Lens 2 (The Fragment Lens): Sees the big picture (the functional groups and shapes).

Most old models could only use one lens at a time. h-MINT uses both simultaneously. It allows information to flow back and forth between the tiny atoms and the big fragments. It’s like a conductor in an orchestra who can hear both the individual violinists and the entire melody of the symphony at the same time. This "bidirectional" hearing allows the model to understand that a tiny change in an atom can change the entire "song" of the molecule.

Why does this matter? (The "So What?")

In the real world, discovering a new drug is like searching for a needle in a haystack the size of the solar system. It takes years and billions of dollars.

The researchers tested H-MINT against the current "world champions" of AI, and H-MINT won. It was significantly better at:

Predicting Strength: How tightly the "key" turns in the "lock."
Virtual Screening: Scanning through millions of potential drugs to find the winners much faster and more accurately.

In short: By teaching computers to see molecules as overlapping, meaningful patterns rather than just piles of atoms or rigid blocks, they’ve given us a much faster, more accurate "map" to navigate the world of medicine.

Technical Summary: H-MINT: Modeling Pocket-Ligand Binding with Hierarchical Molecular Interaction Network

Published at: ICLR 2026

1. Problem Statement

Accurate modeling of protein-ligand interactions is essential for drug discovery (e.g., binding affinity prediction and virtual screening). Existing deep learning approaches face two primary limitations:

Atom-level Graph Representations: While common, atom-level graphs struggle to capture higher-order chemical contexts such as stereochemistry (chirality), lone pairs, and conjugation (aromaticity) solely from atomic tokens.
Fragment-based Methods: Existing fragmentation techniques (e.g., Principal Subgraph) typically partition molecules into disjoint sets. This "naive partitioning" breaks essential chemical structures, such as aromatic rings or ionic states, failing to preserve the "fuzzy" boundaries of molecular substructures.

2. Methodology

The authors propose a two-pronged solution involving a novel tokenization scheme and a specialized hierarchical architecture.

A. OverlapBPE Tokenization
To move beyond disjoint partitioning, the authors introduce OverlapBPE, a data-driven, frequency-based tokenization algorithm:

Overlapping Fragments: Unlike standard methods, OverlapBPE allows fragments to share atoms. This preserves the integrity of aromatic systems (e.g., a naphthalene molecule can be represented by two overlapping benzene rings).
Chemical Enrichment: The algorithm incorporates 3D conformational information and enriched atomic properties (e.g., formal charges, aromaticity) into the tokenization process. This ensures that chirality and ionic states (like $[N^+]$ ) are preserved as unique identifiers.
Bottom-Up Merging: Inspired by Byte Pair Encoding (BPE), it iteratively merges the most frequent adjacent token pairs to build a vocabulary of composite tokens, ranging from single atoms to large functional motifs.

B. h-MINT (Hierarchical Molecular Interaction NeTwork)
Because OverlapBPE creates a many-to-many mapping between atoms and fragments, standard hierarchical models (which assume 1-to-1 mapping) are insufficient. The authors developed h-MINT:

Bilevel Attention Mechanism: An SE(3)-equivariant graph Transformer that enables bidirectional information flow between atoms and their overlapping fragments.
Hierarchical Graph Construction: It builds a "Token-expanded Atom Graph" where token-level edges (based on KNN of tokens) are expanded into atom-level geometric edges. This allows the model to capture both fine-grained local physics and coarse-grained long-range interactions (like $\pi$ -stacking).
Equivariance: The model maintains SE(3)-equivariance through specialized equivariant feed-forward networks and layer normalization, ensuring consistent geometric reasoning.

3. Key Contributions

OverlapBPE: A novel tokenization method that preserves chemical context (chirality, aromaticity, charges) by allowing overlapping substructures.
h-MINT Architecture: A new hierarchical, SE(3)-equivariant framework specifically designed to handle the many-to-many atom-fragment mappings induced by overlapping tokens.
Automated Motif Discovery: The ability to automatically mine complex, biologically relevant motifs (e.g., adenosine monophosphate or sulfonamide protease inhibitor motifs) directly from data without manual rules.

4. Results

The model was evaluated across several critical benchmarks:

Binding Affinity Prediction (PDBBind & LBA): h-MINT achieved state-of-the-art performance, improving Pearson/Spearman correlation by 2–4% over existing models like GET and GET-PS.
Virtual Screening (DUD-E & LIT-PCBA): In zero-shot settings, the model outperformed leading baselines (e.g., DrugCLIP, LigUnity) by 1–3% in key metrics like AUC and BEDROC, demonstrating strong generalization.
High-Throughput Screening (PubChem): Using only tokenized bag-of-words with XGBoost, the OverlapBPE method outperformed complex GNN baselines (like MolKGNN) in early-enrichment performance (logAUC).
Robustness: The model showed high resilience to structural noise (up to 5 Å) and fragment masking.

5. Significance

This work shifts the paradigm of molecular representation from "disjoint partitioning" to "overlapping coverage." By successfully bridging the gap between fine-grained atomic physics and coarse-grained chemical functionality, H-MINT provides a more expressive and chemically accurate framework for drug discovery. Its ability to capture complex motifs and maintain chirality makes it a highly potent tool for real-world applications in enzyme engineering and pharmaceutical lead optimization.