Peking University Team Introduces MTPNet: The First Generalized Prediction Framework for Target-Perceived 'Activity Cliffs'}
Peking University unveils MTPNet, a novel target-aware model for predicting activity cliffs by integrating macro and micro target features, demonstrating strong performance and interpretability.
Editor | ScienceAI
"Two nearly identical molecules, one highly active, the other inactive"—this is one of the most perplexing puzzles in drug discovery for AI—Activity Cliffs, phenomena where tiny structural changes cause huge, unpredictable activity differences.
To address this, Peking University’s team proposed MTPNet (Multi-Grained Target Perception Network), the first target protein-aware universal prediction framework. It introduces a multi-granularity target semantic perception mechanism that combines macro-level receptor features with micro-level binding site details, modeling receptor-ligand interactions comprehensively. The model achieved significant performance improvements on 30 activity cliff datasets, demonstrating excellent generalization and interpretability.
Related paper: https://arxiv.org/pdf/2506.05427
Open-source code: https://github.com/ZishanShu/MTPNet
1. Why incorporate target proteins as condition information?
Previous models mainly focused on molecular structure or chemical features, often ignoring the deep root cause—namely, the key role of receptor-ligand interactions during binding. These interactions are fundamental drivers of activity cliffs.
Figure 1: Motivation and methodology of MTPNet. It posits that only when structural differences concentrate on receptor-sensitive regions, disrupting key binding modes or inducing conformational changes, will activity cliffs occur. MTPNet models molecular and receptor features at multiple granularities to better identify these critical changes.
2. The architecture of MTPNet: A multi-granularity target perception framework
To model receptor-ligand interactions at both receptor-level and binding site-level, MTPNet employs the MTP module, composed of the Macro-level Target Semantic (MTS) and Micro-level Pocket Semantic (MPS) networks. MTS uses pretrained protein models to extract receptor embeddings, guiding molecular features via conditional layer normalization and self-attention. MPS captures local interactions within binding pockets through cross-attention. The iterative fusion of these features enables precise activity cliff prediction.
Figure 2: MTPNet architecture design.
3. Comprehensive performance evaluation
The team evaluated MTPNet on the MoleculeACE benchmark, designed to assess activity cliffs caused by subtle molecular variations across 30 high-quality subsets with over 35,000 samples. Results showed an average RMSE reduction of 18.95%, PCC increase of 11.6%, and R2 improvement of 17.8%.
Figure 3: Prediction performance of MTPNet. It also excels in activity cliff classification on the CYP3A4 dataset, achieving an AUC of 0.924, surpassing state-of-the-art models like Mole-BERT (0.902) and MolCLR (0.896).
4. Plug-and-play capability of MTP module
The team tested the MTP module as a plug-in across baseline models such as GCN, GAT, GIN, MolCLR, and Mole-BERT. Results showed over 15% RMSE improvement and performance surpassing scaled-up models, demonstrating broad compatibility and robustness.
Figure 4: Plug-and-play effects of the MTP module.
5. Interpretability of MTPNet
The model pays significant attention to key functional groups such as amino (-NH₂), carbonyl (C=O), sulfonyl (O=S=O), carboxyl (-COOH), and halogens, as well as specific bonds like double and triple bonds. This focus aligns with chemical principles that functional groups dominate molecular properties and reactivity, reflecting strong interpretability. MTPNet not only accurately identifies influential regions but also reveals how functional groups and bonds affect solubility, hydrophobicity, and affinity.
Figure 5: Interpretability of MTPNet.
6. Summary and outlook
MTPNet explicitly incorporates receptor proteins as condition information, enabling universal activity cliff prediction. It has broad applications in drug discovery, molecular optimization, and chemical mechanism research, helping identify key functional groups and bonds that cause activity changes, thus improving screening efficiency and reducing trial-and-error. Its high interpretability also aids in understanding molecular mechanisms of protein-ligand binding, guiding rational drug design, and exploring activity cliff origins. Future extensions could include toxicity prediction, structure-activity relationship modeling, and in-depth studies of complex receptor-ligand interactions.