🤖 AI

Structure-Aware Compound-Protein Affinity Prediction via Graph Neural Network with Group Lasso Regularization

This paper proposes an explainable graph neural network framework enhanced with group lasso and sparse group lasso regularizations to predict compound-protein affinity by leveraging activity cliff pairs, thereby improving prediction accuracy and identifying critical molecular substructures for drug discovery.

Original authors: Zanyu Shi, Yang Wang, Pathum Weerawarna, Jie Zhang, Timothy Richardson, Yijie Wang, Kun Huang

Published 2026-05-29

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Zanyu Shi, Yang Wang, Pathum Weerawarna, Jie Zhang, Timothy Richardson, Yijie Wang, Kun Huang

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

The Big Picture: Finding the "Magic Ingredient" in Medicine

Imagine you are a chef trying to figure out why two very similar soups taste completely different. One is delicious and cures a cold, while the other is bland and does nothing. Both soups share the same base broth and most of the same vegetables (the scaffold), but one has a pinch of a special spice (the substituent) that the other lacks.

In the world of drug discovery, scientists face this exact problem. They have pairs of molecules that look almost identical but have vastly different powers to fight a specific disease target (like a virus or a cancer protein). These pairs are called "Activity Cliffs."

The problem is that traditional AI models often get confused by the massive amount of "soup base" (common parts) and miss the tiny "spice" (the specific change) that actually makes the drug work. This paper proposes a new way to train AI to spot that magic spice.

The Problem: Too Much Noise, Not Enough Signal

Current AI models for drug discovery are like students trying to study for a test by reading an entire encyclopedia. They see millions of inactive compounds (the "noise") and only a few active ones. Furthermore, tiny changes in a molecule's shape can completely change how it works, making it hard for the AI to learn the rules.

When scientists try to explain why an AI made a decision (a field called Explainable AI), the AI often points to the wrong parts of the molecule. It might highlight the common broth instead of the special spice, leading to unstable and confusing explanations.

The Solution: A Specialized Training Camp

The authors built a new framework using Graph Neural Networks (GNNs). Think of a molecule as a map where atoms are cities and chemical bonds are roads. The AI travels this map to learn the molecule's properties.

Here is how their new method works:

The "Cliff" Training: Instead of feeding the AI random molecules, they feed it pairs of "Activity Cliffs"—molecules that are 90% identical but have a huge difference in power. This forces the AI to focus on the 10% that is different.
Separating the Wheat from the Chaff: The AI is trained to look at two things separately:
- Common Nodes: The shared "scaffold" (the soup base).
- Uncommon Nodes: The unique "substituents" (the special spice).
The "Pruning Shears" (Regularization): This is the paper's secret sauce. They added a mathematical tool called Group Lasso and Sparse Group Lasso.
- Analogy: Imagine the AI is a gardener. Without these tools, the gardener might water every plant equally. With Group Lasso, the gardener is forced to either water an entire group of plants or cut them all out. With Sparse Group Lasso, the gardener can be even more precise, cutting out specific weeds within a group while keeping the good flowers.
- Result: This "pruning" forces the AI to ignore the unimportant parts of the molecule and highlight only the specific atoms that cause the drug to work (or fail).

The Results: Sharper Focus, Better Predictions

The team tested this on three specific proteins (Src kinases) involved in diseases like Alzheimer's and cancer.

Better Accuracy: By using their new "pruning" rules and looking at both the common and uncommon parts of the molecules, the AI made fewer mistakes. The error rate dropped significantly (by about 15-17%), and the predictions matched real-world data much more closely.
Better Explanations: When the AI was asked to point out which part of the molecule was responsible for the drug's power, it did a much better job.
- Without the special tools, the AI was like a blurry camera, highlighting large, vague areas.
- With the Sparse Group Lasso, the AI acted like a high-definition camera, accurately coloring the specific atoms that mattered. This "atom coloring" matched the ground truth (what scientists knew was true) much better.

The Bottom Line

This paper doesn't claim to have cured a disease yet. Instead, it offers a better microscope for drug discovery.

By teaching AI to ignore the "noise" of common molecular structures and focus intensely on the tiny, critical differences that create "Activity Cliffs," the researchers have created a model that is not only more accurate at predicting drug strength but also more honest about why it made that prediction. This helps scientists understand which specific parts of a molecule to tweak during the "lead optimization" phase (fine-tuning a drug candidate) to make it more effective.

In short: They taught the AI to stop looking at the whole soup and start tasting the specific spice that makes the difference.

Technical Summary: Structure-Aware Compound-Protein Affinity Prediction via Graph Neural Network with Group Lasso Regularization

Problem Statement
The application of deep learning to Structure-Activity Relationship (SAR) modeling for compound property prediction faces significant challenges. These include data imbalance, where inactive compounds vastly outnumber active ones for specific protein targets, and the complexity introduced by subtle molecular configuration changes that drastically alter properties. While Explainable Artificial Intelligence (XAI) is increasingly used to rationalize predictions, existing end-to-end models often highlight inconsistent molecular substructures, leading to unstable explainability. Furthermore, many models fail to optimally leverage structural information to identify key moieties responsible for compound-protein affinity, particularly when dealing with "activity cliffs"—pairs of structurally similar compounds with large potency differences.

Methodology
The authors propose a novel framework utilizing Graph Neural Networks (GNNs) and Message Passing Neural Networks (MPNNs) trained on matched molecular pair (MMP) activity cliffs. The core methodology involves:

Data Construction: The study utilizes activity-cliff pairs derived from the BindingDB validation datasets, specifically targeting three Src family tyrosine-protein kinases (PDB IDs: 1O42, 2H8H, 4MXO). Pairs are defined as compounds sharing a molecular scaffold (via Maximum Common Substructure, MCS) but differing at substituent sites with at least a 1 log unit difference in activity (e.g., IC50).
Model Architecture: The framework processes molecular graphs where nodes and edges represent atoms and bonds. The model learns latent representations for both "common" nodes (shared scaffolds) and "uncommon" nodes (substituents).
Loss Function Design:
- Base Loss: The primary objective minimizes the Mean Squared Error (MSE) between predicted and experimental binding affinities.
- Node-Specific Loss: A specific node loss ( $L_N$ ) is designed to capture features from both common and uncommon substructures, moving beyond previous approaches that focused solely on uncommon motifs.
- Regularization: To enhance explainability and feature selection, the authors integrate Group Lasso and Sparse Group Lasso penalties. These regularizations prune and highlight molecular subgraphs by treating common and uncommon structures as distinct groups. Group Lasso enforces group-level sparsity (selecting or discarding entire groups), while Sparse Group Lasso adds within-group sparsity, allowing for finer selection of individual features within a group.
Feature Attribution: The study evaluates model interpretability using multiple attribution methods (CAM, DeepLIFT, Grad-CAM, Integrated Gradients). Ground-truth labels for atom-level coloring are derived from the activity difference signs between paired compounds.

Key Contributions

Integrated Structural Learning: The framework successfully integrates both common (scaffold) and uncommon (decoration) node information, demonstrating that considering structural integrity alongside motif differences improves prediction accuracy.
Regularization for Explainability: The application of Group Lasso and Sparse Group Lasso is shown to not only improve predictive performance but also significantly enhance feature attribution. These methods prune irrelevant subgraphs and highlight vital molecular substructures, providing stable subgraph-level explainability.
Activity Cliff Utilization: By focusing training on MMP-cliff pairs, the model specifically targets the structural determinants of large potency differences, addressing the limitations of training on imbalanced, broad datasets.

Results
The proposed framework was evaluated using 5-fold cross-validation on the three Src kinase targets:

Predictive Performance: The model utilizing the loss function with both common and uncommon substructures, combined with Sparse Group Lasso, achieved the best performance. It recorded the lowest average Root Mean Squared Error (RMSE = 0.2143) and the highest Pearson Correlation Coefficient (PCC = 0.9615). Compared to a baseline loss function considering only uncommon nodes without penalties, this approach reduced the average RMSE by 17.48%.
Explainability Metrics: The inclusion of regularization significantly improved feature attribution.
- Global Direction Scores: Models with Group Lasso showed substantial increases in global direction scores (e.g., a 47.13% increase for CAM, 14.83% for Grad-CAM), indicating better preservation of the activity difference direction across compound pairs.
- Atom-Level Accuracy: Visual comparisons of atom coloring (Grad-CAM) revealed that models trained with Sparse Group Lasso produced predictions more consistent with ground-truth feature attribution labels compared to non-regularized models.

Significance and Claims
The paper claims that this approach strengthens model interpretability within drug discovery pipelines, particularly for lead optimization where identifying critical molecular substructures is essential. By leveraging activity cliffs and regularization, the framework offers a more stable and accurate method for predicting compound-protein affinity and rationalizing drug property differences. The authors note that while the current study focuses on Src kinases due to data availability and relevance to Alzheimer's disease and cancer, the framework is designed to be applicable to other protein targets. They acknowledge that the use of average pooling may slightly weaken signals from vital nodes and that broader validation across more protein families would require further computational resources.