Bayesian Optimization in the Latent Space of a Variational Autoencoder for the Generation of Selective FLT3 Inhibitors

Raghav Chandra, Robert I. Horne, Michele VendruscoloJournal of Chemical Theory and Computation 2023 · DOI

TL;DR

This paper introduces a novel computational pipeline that combines deep learning generative models with Bayesian optimization to design highly selective drug candidates. By exploring the latent space of a variational autoencoder, the method efficiently identifies molecules with strong binding to a target while minimizing binding to related off-targets, demonstrating promising results for FLT3 inhibitors.

What Problem Does This Paper Address?

Drug discovery is a notoriously slow, expensive, and inefficient process. A major bottleneck is identifying molecules that not only bind strongly to their intended biological target (high affinity) but also avoid binding to other, similar targets (high selectivity). This selectivity is crucial to minimize side effects and improve drug efficacy. Traditional methods often rely on large datasets and extensive screening, which can be time-consuming and costly. While deep learning has shown promise in generating novel molecules, achieving high selectivity has remained a significant challenge, often due to the vastness of chemical space and the difficulty in representing molecular properties continuously. This paper aims to address this by developing a more efficient and targeted approach to generate selective drug candidates with reduced data requirements.

Methods

The core of this work is a five-component pipeline designed for generating selective drug candidates:

  1. Binding Affinity Prediction: The researchers employed a variant of AutoDock Vina called "Vinardo" to predict the binding energies of small molecules to protein targets. This is a standard docking method, but the use of Vinardo and its specific parameter tuning is highlighted.
  2. Off-Target Identification: To ensure selectivity, the pipeline first identifies structurally similar protein pockets to the intended target. This was achieved using the PoSSuM database, which uses a fast 3D pocket alignment method to find homologous targets based on geometric and physicochemical properties.
  3. Molecular Representation: A Junction Tree Variational Autoencoder (JTVAE) was used to encode small molecules into a 56-dimensional latent space. This latent space is continuous and allows for smooth exploration, overcoming the discontinuity issues often found in molecular representations. The JTVAE was pretrained on a large dataset of drug-like molecules.
  4. Objective Function: A crucial component is the objective function, which quantifies the desirability of a generated molecule. It's designed to maximize binding affinity to the target while simultaneously penalizing binding to identified off-targets. An "anybind" term is also included to discourage weak binding to the target.
  5. Bayesian Optimization: Given the computational cost of predicting binding affinities for many molecules, Bayesian optimization was employed to efficiently search the JTVAE's latent space. This method intelligently selects the next points to evaluate based on an acquisition function, aiming to maximize the objective score with fewer evaluations than random sampling.

Datasets

  • ZINC Database: Used for pretraining the Junction Tree Variational Autoencoder (JTVAE) on approximately 250,000 drug-like molecules.
  • Protein Data Bank (PDB): Used by the PoSSuM database to identify protein binding sites.
  • FLT3 Target Data: The pipeline was demonstrated using FMS-like tyrosine kinase 3 (FLT3) as the target. Data on known FLT3 inhibitors and their off-target bindings were used for validation and comparison.

Key Insights and Ideas

  • Latent Space Exploration for Selectivity: The paper demonstrates that exploring the continuous latent space of a VAE, rather than the discrete chemical space directly, can be a more efficient strategy for optimizing multiple objectives like affinity and selectivity.
  • Bayesian Optimization as an Efficiency Driver: The high computational cost of binding affinity predictions makes exhaustive searches infeasible. Bayesian optimization proves to be a powerful tool for navigating this complex, high-dimensional space and finding optimal solutions with significantly fewer evaluations.
  • Integrated Selectivity Mechanism: By explicitly incorporating off-target binding penalties into the objective function and using a systematic method for identifying relevant off-targets, the pipeline actively guides the generation process towards selective molecules.
  • Bridging Generative Models and Predictive Tools: The seamless integration of a generative model (JTVAE) with predictive tools (Vinardo, PoSSuM) creates a robust end-to-end drug design pipeline.

Key Results

  • The pipeline successfully generated small molecules with predicted binding affinities and selectivities comparable to, and in some cases superior to, clinically approved drugs for FLT3.
  • Bayesian optimization significantly increased the hit rate (molecules with objective scores > 0.5) from approximately 4.5% with random sampling to around 10%.
  • The generated molecules exhibited high Quantitative Estimate of Drug-likeness (QED) scores, suggesting good drug-like properties.
  • The study identified several promising FLT3 inhibitor candidates with high predicted affinity and selectivity, some of which had fewer than 5 rotatable bonds, a desirable trait for drug development.

Limitations and Future Work

  • Experimental Validation: The authors acknowledge that the generated molecules are predictions and require experimental validation to confirm their actual binding affinities and selectivities. This is stated as the next crucial step.
  • Synthesizability: While the JTVAE generates chemically valid molecules, it doesn't guarantee synthesizability or stability. The paper notes that some generated molecules might be difficult to synthesize.
  • Off-Target Identification Complexity: The paper mentions that for targets with highly similar off-targets (like FLT3 and its close relatives), distinguishing binding affinities can be particularly challenging. Future work could explore more sophisticated methods for off-target identification or weighting.
  • Generalizability: While demonstrated for FLT3, the generalizability of this pipeline to other targets and protein families would be an important area for future research. The authors suggest that the approach could be adapted to other targets where selectivity is a concern.

Summary updated

Abstract

The process of drug design requires the initial identification of compounds that bind their targets with high affinity and selectivity. Advances in generative modeling of small molecules based on deep learning are offering novel opportunities for making this process faster and cheaper. Here, we propose an approach to achieve this goal, where predictions of binding affinity are used in conjunction with the Junction Tree Variational Autoencoder (JTVAE) whose latent space is used to facilitate the efficient exploration of the chemical space using a Bayesian optimization strategy. The exploration identifies small molecules predicted to have both high affinity and high selectivity by using an objective function that optimizes the binding to the target while penalizing the binding to off-targets. The framework is demonstrated for FMS-like tyrosine kinase 3 (FLT3) and shown to predict small molecules with predicted affinity and selectivity comparable to those of clinically approved drugs for this target.
BibTeX
@article{chandra2023bayesian,
  author = {Raghav Chandra and Robert I. Horne and Michele Vendruscolo},
  title = {{Bayesian Optimization in the Latent Space of a Variational Autoencoder for the Generation of Selective FLT3 Inhibitors}},
  journal = {Journal of Chemical Theory and Computation},
  year = {2023},
  doi = {10.1021/acs.jctc.3c01224},
}
Figure