Abstract
Motivation: Computational methods for compound–protein affinity and contact (CPAC) prediction aim at facilitating rational drug discovery by simultaneous prediction of the strength and the pattern of compound–protein interactions. Although the desired outputs are highly structure-dependent, the lack of protein structures often makes structure-free methods rely on protein sequence inputs alone. The scarcity of compound–protein pairs with affinity and contact labels further limits the accuracy and the generalizability of CPAC models.
Results: To overcome the aforementioned challenges of structure naivety and labelled-data scarcity, we introduce cross-modality and self-supervised learning, respectively, for structure-aware and task-relevant protein embedding. Specifically, protein data are available in both modalities of 1D amino-acid sequences and predicted 2D contact maps, that are separately embedded with recurrent and graph neural networks, respectively, as well as jointly embedded with two cross-modality schemes. Furthermore, both protein modalities are pretrained under various self-supervised learning strategies, by leveraging massive amount of unlabelled protein data. Our results indicate that individual protein modalities differ in their strengths of predicting affinities or contacts. Proper cross-modality protein embedding combined with self-supervised learning improves model generalizability when predicting both affinities and contacts for unseen proteins.
Introduction
Rational drug discovery relies on predicting how small molecules interact with target proteins. This involves understanding both the binding affinity (the strength of the interaction) and the specific intermolecular contacts (the physical points of touch between atoms and residues). This dual task is known as Compound–Protein Affinity and Contact (CPAC) prediction.
However, current computational methods face two significant hurdles:
- Structure Naivety: Most models rely solely on 1D protein sequences because experimental 3D structures are difficult to obtain, even though binding is a physical, 3D process.
- Supervision Starvation: While millions of protein sequences are available, there are very few "labeled" examples where both the exact binding affinity and contact points are known.
The CPAC Framework
The researchers developed CPAC, a model that addresses these challenges through two primary innovations: cross-modality protein embedding and self-supervised learning (SSL).
Model Architecture
The overall framework, illustrated in Figure 1, consists of separate encoders for the compound and the protein. The compound ($X_{\text{comp}}$) is processed into an embedding ($H_{\text{comp}}$), while the protein ($X_{\text{prot}}$) is represented through two modalities: 1D amino-acid sequences and predicted 2D contact maps (representing structural proximity).
Cross-Modality Protein Embedding
A core contribution of this work is how these protein modalities are combined. As shown in Figure 2, the authors tested two fusion schemes. The first is a simple Concatenation of sequence and graph features. The second is a more sophisticated Cross-Interaction module, which allows the 1D sequence representation (from an HRNN) and the 2D graph representation (from a GAT) to mutually inform each other during the encoding process.
Self-Supervised Pre-training
To solve the data scarcity problem, the encoders are pre-trained on massive unlabeled datasets. For 1D sequences, the model uses Masked Language Modeling (MLM). For 2D graphs, it uses Graph Completion, where the model learns to predict missing structural information. This allows the model to learn general protein features before ever seeing a labeled drug-binding example.
Experimental Setup and Datasets
The model was evaluated using a curated set of 4,446 compound-protein pairs from PDBbind and BindingDB. Crucially, the test sets were split into "Seen" and "Unseen" proteins and compounds to evaluate the model's ability to generalize to entirely new drugs or targets. For SSL pre-training, the authors leveraged a high-quality set of ~60,000 structures and a massive dataset (Pfam-A RP15) containing nearly 12.8 million protein sequences.
Key Results and Insights
- Modality Specialization: The study found that 1D sequences are surprisingly robust for predicting binding affinity, while 2D graphs are significantly better at predicting physical intermolecular contacts.
- Generalization Boost: The "Cross-Interaction" model with SSL pre-training outperformed existing benchmarks. For the hardest test cases—where both the protein and the compound were "unseen"—the contact prediction accuracy (AUPRC) jumped from roughly 4% in standard models to over 12%.
- Quality over Quantity: Interestingly, pre-training on the smaller dataset of high-quality structural information often yielded better results than pre-training on the massive 12.8M sequence dataset, suggesting that structural "awareness" is more valuable than raw sequence volume for CPAC tasks.
Conclusion
CPAC demonstrates that integrating structural awareness through 2D contact maps and leveraging unlabeled data via self-supervised learning significantly enhances the prediction of protein-ligand interactions. While there is a slight trade-off between optimizing for affinity versus contact, the joint approach provides a robust foundation for virtual drug screening. Future work will focus on better balancing these two objectives and incorporating increasingly accurate structural predictions from tools like AlphaFold2.
Summary updated