TL;DR
This paper introduces OptiMol, a novel framework that combines ligand-based generative models with structure-based molecular docking to efficiently discover drug candidates with high binding affinities. By using docking as a costly oracle to guide a generative model, OptiMol significantly accelerates the identification of promising molecules, achieving a tenfold increase in high-scoring compounds within a fixed computational budget.
What Problem Does This Paper Address?
The search for new drugs involves navigating an astronomically vast chemical space (estimated at 10^60 compounds) to find molecules that bind effectively to specific biological targets. Traditional methods often rely on either ligand-based approaches (predicting activity based on known active molecules) or structure-based approaches (simulating how molecules interact with a target's 3D structure). While deep generative models have revolutionized ligand-based drug design by enabling exploration of this chemical space, they often fail to incorporate crucial target structure information, limiting their ability to discover novel binding modes. Conversely, structure-based methods like molecular docking are computationally expensive and can only realistically screen a fraction of the available chemical space. This paper addresses the challenge of efficiently and effectively exploring the chemical space to find molecules with high binding affinity to a target, bridging the gap between ligand-based and structure-based approaches.
Methods
The core of OptiMol is a closed-loop optimization pipeline that integrates ligand-based generative modeling with structure-based docking.
- Graph-to-Selfies VAE: The authors introduce a new Variational Autoencoder (VAE) that uses molecular graphs as input and generates molecules represented by the Selfies (Self-referencing Embedded Strings) format. This approach offers several advantages:
- Graph Input: Using molecular graphs as input helps learn a more chemically organized latent space and is more robust to minor structural changes compared to string-based representations like SMILES.
- Selfies Output: Selfies ensure that all generated molecules are chemically valid, overcoming a common issue with SMILES generation.
- Efficiency: This graph-to-Selfies VAE boasts a significantly faster decoding time (17x faster) compared to graph-to-graph methods while maintaining competitive performance.
- Structure-Based Docking as an Oracle: Molecular docking (specifically using Autodock Vina) is employed as a "black-box" oracle to estimate binding affinity. This is a computationally intensive step, making it impractical to dock every generated molecule.
- Conditioning by Adaptive Sampling (CbAS): To handle the costly and non-differentiable nature of the docking oracle, the authors leverage a recently proposed method called Conditioning by Adaptive Sampling (CbAS). CbAS iteratively updates a generative model to produce samples that maximize an objective function (in this case, docking score). It uses an importance sampling scheme and an alternating phase of sampling and model tuning to efficiently explore the chemical space and adapt the generative model towards higher-scoring compounds.
Datasets
- Moses Benchmark Set: Used for training and evaluating the graph-to-Selfies VAE. This dataset contains approximately 1.5 million molecules from the ZINC database.
- ExCAPE Database: Used for evaluating the docking performance. This database provided 380 clustered active compounds and 20,000 decoys for the Dopamine Receptor D3 (DRD3) target.
- ZINC Database: Used as a source for generating random samples for comparison.
Key Insights and Ideas
- Bridging Ligand and Structure-Based Approaches: The paper demonstrates that effectively combining generative models with docking can overcome the limitations of each individual approach, leading to more efficient and targeted drug discovery.
- Docking as a Guiding Force: Using docking not just for final evaluation but as an iterative oracle to guide the generative process is key to discovering molecules with improved binding affinities.
- CbAS for Costly Oracles: The successful application of CbAS to a costly, non-differentiable oracle like molecular docking is a significant contribution, opening doors for optimizing other complex objectives in drug discovery.
- Efficiency Gains: The new graph-to-Selfies VAE offers a substantial speedup in molecule generation, making large-scale exploration more feasible.
- Target-Specific Optimization: The framework allows for optimizing molecules for specific targets without prior knowledge of existing active compounds, which is crucial for novel targets.
Key Results
- Improved Docking Scores: The OptiMol pipeline, using CbAS, successfully optimized molecule generation to achieve significantly enhanced docking scores, shifting the average score from -7.5 to -8.5 kcal/mol.
- Ten-Fold Increase in Hits: With a fixed computational budget, the method achieved a tenfold increase in the number of high-scoring compounds (docking score better than -10 kcal/mol) compared to sampling from the original chemical space. Specifically, they were able to sample a distribution where 7.4% of compounds scored better than -10, compared to 0.62% from the prior distribution.
- Competitive VAE Performance: The graph-to-Selfies VAE achieved state-of-the-art performance on the Moses benchmark metrics while being 17 times faster for sampling and training.
- Diversity Preservation: The optimization process did not lead to a collapse in chemical diversity, as indicated by similar mean pairwise Tanimoto distances between generated samples and baseline datasets.
- Efficient Implementation: The CbAS implementation was parallelized, allowing training of the model in approximately 15 hours using hundreds of CPU cores.
Limitations and Future Work
- Computational Cost: While significantly more efficient than traditional methods, docking remains a bottleneck. The authors note that further optimization of the docking phase could be beneficial.
- Objective Function Design: The paper suggests that the smoothness of the objective function might impact the success of the method. Future work could explore composite objective functions that penalize docking score by drug-likeliness or synthetic accessibility.
- Latent Space Reshaping: The authors hypothesize that reshaping the latent space and incorporating 3D molecular properties could further improve performance.
- Retraining on Diverse Datasets: Extending the validation by retraining the prior generative model on different datasets is suggested.
- Exploration of Other Optimization Methods: While CbAS proved effective, exploring other optimization techniques for non-differentiable and costly oracles could be valuable.
Summary updated