Abstract
Proteins are dynamic entities whose functional behavior is dictated by their molecular context. Traditional machine learning models often represent proteins as static vectors, failing to capture this inherent adaptability. We present Docking-Aware Attention (DAA), a novel architecture that integrates 3D molecular docking predictions directly into a transformer's attention mechanism. By leveraging physics-based scoring from sampled binding poses, DAA generates dynamic, context-dependent protein embeddings. Evaluated on the ECREACT dataset, DAA achieves state-of-the-art performance in enzymatic reaction prediction, particularly for complex and novel substrates, demonstrating the power of combining deep learning with physical constraints.
Introduction
The prediction of enzyme-catalyzed reactions, or biocatalysis prediction, is a cornerstone of modern drug discovery and green chemistry. Enzymes are highly efficient but versatile; a single enzyme can catalyze distinct chemical transformations depending on the specific substrate it interacts with.
Current computational approaches struggle with this adaptability. Most models rely on static protein representations—either discrete Enzyme Commission (EC) labels or fixed embeddings from protein language models (e.g., ESM) or structural models (e.g., GearNet). These static vectors do not change regardless of the molecular context, ignoring the physical reality of enzyme-substrate interactions. To accurately predict chemical products, models require dynamic representations that physically adapt to the input molecule.
Methodology: Docking-Aware Attention (DAA)
The DAA framework bridges physics-based molecular docking with sequence-to-sequence language models to create context-aware embeddings. The architecture follows a multi-stage process:
- Protein Encoding: Standard per-amino-acid embeddings are generated using ESM3, a state-of-the-art protein language model.
- Interaction Simulation: Multiple 3D binding poses between the protein and the substrate are sampled using DiffDock, a generative diffusion model.
- Physics-Based Scoring: For each pose, interaction strengths are calculated using the Lennard-Jones potential, modeling atomic attraction and repulsion.
- Attention Integration: These physical scores are injected into the transformer's attention mechanism. This biases the model to "pay attention" to amino acids that physically interact with the specific substrate.
- Sequence-to-Sequence Prediction: The resulting dynamic protein representation is used by a T5 model to translate the substrate's SMILES string into the product's SMILES string.
Experimental Results
The model was evaluated on the ECREACT dataset (60,000 enzymatic reactions) and pre-trained on the USPTO dataset (1 million non-enzymatic reactions) to learn general chemical transformation rules.
DAA consistently outperformed all baseline methods, including static sequence-based (ESM3, ProtBERT) and structure-based (GearNet) models.
Key Performance Metrics
- Overall Accuracy: DAA achieved a Top-1 accuracy of 49.96%, surpassing the previous best (ReactEmbed) at 47.41%.
- Complexity Handling: On complex molecules (Bertz Complexity > 1500), DAA's Top-1 accuracy reached 41.62%, a 13% relative improvement over baselines.
- Generalization to Novelty: For reactions involving molecules not seen during training, DAA achieved 35.24% Top-1 accuracy, significantly outperforming static methods like ESM3 (27.39%).
Discussion and Conclusion
The success of DAA highlights that physics and deep learning are complementary. By forcing the protein representation to change based on the substrate, the model avoids simple memorization and instead learns the underlying principles of enzyme-substrate interactions. Visualizations confirm that the attention mechanism dynamically shifts focus to relevant binding pockets for different substrates.
While the use of diffusion-based docking models like DiffDock increases computational overhead, the gains in accuracy for complex and novel reactions are substantial. Future work will explore integrating DAA into multi-step retrosynthesis planning and modeling the impact of environmental factors like pH and temperature on enzymatic activity.
Summary updated