Abstract
De novo molecular design has facilitated the exploration of large chemical space to accelerate drug discovery. Structure-based de novo method can overcome the data scarcity of active ligands by incorporating drug-target interaction into deep generative architectures. However, these strategies are bottlenecked by the small fraction of experimentally determined protein or complex structures. In addition, the cost of molecular generation is computationally expensive due to 3D representations of both molecule and protein. Here, we demonstrate a widely used and fast protein sequence-based reinforcement learning (RL) model for drug discovery. In the generative model, one of the reward components, a binding affinity predictor, is based on 1D protein sequence and molecular SMILES. As a proof of concept, the RL model was utilized to design molecules for four targets. The generated compounds showed bioactivities by the validation of both QSAR and molecular docking with experimental 3D binding pockets. We also found that the performance of generated molecules depends on the selection of data source training for the binding predictor. Furthermore, drug design for a kinase without any experimental structure, CDK20, was studied by our model. With only 1D protein sequence as input, the generated novel compounds showed favorable binding affinity based on the AlphaFold predicted structure.
Introduction
The chemical universe of drug-like molecules is estimated to contain between $10^{23}$ and $10^{60}$ possibilities. Traditional virtual screening is limited to the tiny fraction of these molecules that already exist in databases. While deep learning-based de novo design allows us to explore this "dark" chemical space, current structure-based methods face two major hurdles: the lack of high-resolution 3D structures for approximately 82% of human protein residues, and the high computational cost of 3D docking simulations.
To address these challenges, this research presents a reinforcement learning (RL) framework that is both fast and structure-agnostic, relying solely on the 1D amino acid sequence of the target protein.
Methodology
The model utilizes a generative Recurrent Neural Network (RNN) agent that learns to construct molecular SMILES strings. The agent is guided by a reward function centered on a Drug-Target Affinity (DTA) predictor.
The DTA Architecture
Instead of complex 3D voxels, the model employs two Convolutional Neural Network (CNN) blocks to process the 1D protein sequence and the 1D molecular SMILES. These representations are concatenated and fed into a fully connected network to predict binding affinity. This approach allows the model to bypass the need for 3D coordinates while still capturing essential evolutionary and chemical information.
Training and Validation
The DTA predictors were trained on major bioactivity databases including Davis, Kiba, and BindingDB. To validate the 1D-generated molecules, the researchers used traditional QSAR models and high-fidelity 3D molecular docking (Glide) as independent benchmarks.
Key Results
Speed and Efficiency
By transitioning from 3D docking to 1D sequence prediction, the inference time for 5,000 molecules dropped from 45 minutes to less than 20 milliseconds. This speedup enables the RL agent to explore significantly larger swaths of chemical space.
Target Success and Chemical Space
The model successfully designed potential inhibitors for four well-known targets: EGFR, GSK3β, 5HT1A, and D3R. As shown in the t-SNE analysis, the generated molecules (green) effectively cover and expand upon the chemical space occupied by known inhibitors (blue).
The CDK20 Case Study
To test the model's performance on targets without experimental structures, the researchers targeted CDK20. The model generated novel compounds that showed high docking scores when tested against an AlphaFold-predicted structure, suggesting the model can effectively "pre-design" drugs as soon as a gene sequence is known.
Conclusion
This study demonstrates that 1D protein sequences contain sufficient information to guide generative models in finding relevant chemical neighborhoods. While the model's accuracy is linked to the diversity of its training data, its speed and structure-agnostic nature make it a powerful tool for rapid drug discovery. Future work may involve integrating this fast 1D approach with AlphaFold’s structural predictions for a multi-stage discovery pipeline.
Summary updated