TL;DR
This paper introduces SPE-MONN, a deep learning model that significantly improves predictions of where compounds interact with proteins and how strongly they bind. By leveraging self-supervised protein embeddings, SPE-MONN captures richer protein information than previous methods, leading to more accurate drug discovery and design.
What Problem Does This Paper Address?
Identifying how and how strongly small molecules (compounds) interact with proteins is a cornerstone of drug discovery. This knowledge helps researchers design new drugs that can effectively target disease-causing proteins or modulate their function. Traditional methods often struggle with the complexity of protein structures and the vast chemical space of potential compounds. While deep learning has shown promise, effectively representing proteins, especially their intricate 3D structures and evolutionary information, remains a challenge. Existing models often rely solely on protein sequences, missing crucial structural and functional nuances. This paper aims to bridge this gap by developing a more sophisticated protein representation that enhances the accuracy of predicting both interaction sites and binding affinities.
Methods
The core of this work is the SPE-MONN model, which builds upon the existing MONN architecture. The key innovation lies in how it represents proteins. Instead of solely relying on primary sequences, SPE-MONN incorporates self-supervised protein embeddings. These embeddings are generated by models trained on massive protein datasets without explicit labels, allowing them to learn intrinsic patterns and features. The paper explores three such embedding types: PtsRep (trained on protein tertiary structures), UniRep (trained on protein sequences using mLSTM), and TAPE-BERT (trained on protein sequences using a BERT-like format).
SPE-MONN uses a multi-objective neural network to simultaneously predict two tasks:
- Interaction Site Prediction: Identifying specific residues on a protein that are likely to interact with a compound.
- Binding Affinity Prediction: Quantifying the strength of the interaction between a compound and a protein.
The model architecture consists of:
- A protein convolution module that processes the self-supervised protein embeddings and evolutionary information (derived from BLOSUM62 matrices) to create a combined protein representation.
- A compound graph convolution module that represents compounds as graphs and extracts features using graph neural networks.
- A prediction module that takes the combined protein and compound representations to output predictions for both interaction sites and binding affinity.
The paper also investigates different ways to combine protein embedding and evolutionary information, finding that a specific combination yields optimal results for each task.
Datasets
The primary dataset used for training and evaluation is PDBbind v2018. This dataset contains binding affinity data for biomolecular complexes with known 3D structures. Interaction sites were extracted using the PLIP tool. For comparison, the authors also used data from PDB-2016, PDB-2013, and PDB-2007 datasets.
Key Insights and Ideas
- Self-supervised protein embeddings are powerful: The study strongly suggests that embeddings derived from protein structures (PtsRep) and sequences (UniRep, TAPE) capture richer information than sequence alone, leading to significant performance improvements.
- Structure matters for interaction site prediction: PtsRep, which leverages protein structure, appears particularly beneficial for accurately predicting interaction sites.
- Evolutionary information is crucial for binding affinity: While embeddings are good, combining them with evolutionary information (Pevo) significantly boosts binding affinity prediction accuracy, suggesting that historical protein adaptations play a role in interaction strength.
- A hybrid approach is best: The optimal strategy involves using a combined protein representation (Pcombine) for interaction site prediction and evolutionary information (Pevo) for binding affinity prediction, demonstrating task-specific optimization.
Key Results
- Interaction Site Prediction: SPE-MONN-PtsRep achieved an AUC of 0.811 and 0.849, outperforming the baseline MONN (0.729, 0.811) by 11.2% in the new-compound and new-protein settings, respectively.
- Binding Affinity Prediction: In the new-compound setting, SPE-MONN-PtsRep showed a Pearson correlation coefficient of 0.687 and an RMSE of 1.491, outperforming MONN, DeepDTA, and GraphDTA. Across various settings, SPE-MONN consistently showed improvements in Pearson correlation and reductions in RMSE compared to established methods.
- Case Study: In a case study predicting interaction sites for SARS-CoV-2 protease inhibitors, SPE-MONN-PtsRep identified 19 of the top 10 interaction sites, significantly more than MONN (13 sites).
Limitations and Future Work
- Hyperparameter tuning: The paper notes that the optimal combination ratio for protein representation (alpha) was found to be 0.5, and deviations from this value led to performance degradation. This suggests that careful tuning of such parameters is important.
- Graph Convolution Module: While the current graph convolution module performed best, the authors explored other graph neural network architectures (GAT, GIN, GCN-GAT) and found the existing one to be superior for their tasks. Future work could involve further exploration of advanced graph neural networks.
- Emergence of AlphaFold2: The authors acknowledge that the advent of AlphaFold2 will likely lead to more advanced protein embedding methods based on 3D structures, which could further enhance predictive accuracy and accelerate drug discovery. They anticipate that future research will leverage these advancements.
Summary updated