Generative Molecular Design and Experimental Validation of Selective Histamine H1 Inhibitors

Kevin McLoughlin, Da Shi, Jeffrey E. Mast, John Bucci, John P. Williams, William D Jones, Derrick Miyao, Luke Nam, Heather L. Osswald, L. A. Zegel'man, Jonathan Allen, Brian J. Bennion, Amanda K. Paulson, Ruben Abagyan, Martha S. Head, James M. BrasebioRxiv 2023 · DOI

TL;DR

This paper introduces ATOM-GMD, a scalable platform for generative molecular design that uses a junction tree variational autoencoder and a genetic algorithm to explore vast chemical spaces. They successfully applied it to discover potent and selective histamine H1 receptor antagonists, synthesizing and testing 103 compounds, with six showing promising activity and selectivity.

What Problem Does This Paper Address?

Drug discovery often involves optimizing molecules to possess multiple desirable properties simultaneously, such as high efficacy against a target, low off-target effects, and good pharmacokinetic profiles. This multi-objective optimization is challenging because the chemical space of possible molecules is enormous. Traditional methods can be slow and inefficient. Generative molecular design (GMD) aims to address this by using machine learning to propose novel molecules that are predicted to meet these criteria. However, many existing GMD approaches focus on optimizing only a few simple properties or lack robust experimental validation. This paper tackles the challenge of developing a GMD platform that can handle multiple complex properties, scale to large populations of molecules, and demonstrate its effectiveness through experimental validation in a real-world drug discovery scenario.

Methods

The core of the ATOM-GMD platform is a combination of a junction tree variational autoencoder (JT-VAE) and a genetic algorithm.

  • JT-VAE: This model learns a continuous "latent space" representation of molecules. It can encode complex molecular structures into numerical vectors and decode these vectors back into molecules. The JT-VAE used here is designed to handle the tree-like structure of molecules, capturing both substructures and their connections.
  • Genetic Algorithm: This algorithm operates on the latent space representations. It maintains a population of latent vectors (analogous to chromosomes), applies "crossover" (combining parts of vectors) and "mutation" (random changes) to generate new vectors, and then uses a "cost function" to score how well the molecules decoded from these vectors meet the desired design criteria. The best-scoring molecules are selected as parents for the next generation, iteratively driving the population towards optimal solutions.
  • Multi-Parameter Optimization: The platform is designed to optimize many properties simultaneously. This is achieved through a flexible cost function that assigns penalties based on predicted molecular properties (e.g., H1 binding affinity, M2 binding affinity, hERG inhibition, molecular weight, etc.).
  • Scalable Framework: The entire process is implemented on a high-performance computing cluster using a message broker (RabbitMQ) to coordinate distributed workers, enabling parallel processing of large molecular populations.

Datasets

  • Training Data:
    • H1 and M2 Binding: Data was compiled from Neurocrine Biosciences' single-concentration binding measurements, the ChEMBL database, and the proprietary GoStar database.
  • hERG Inhibition: Data was sourced from ChEMBL, Drug Target Commons, and ExCAPE-DB.
  • Generative Model Training: A set of 24,741 compounds was used to train the JT-VAE, derived from known H1 inhibitors and expanded using the Rapid Isostere Discovery Engine (RIDE) to search the Enamine REAL virtual library for similar molecules.

Key Insights and Ideas

  • Latent Space Genetic Algorithm: Operating a genetic algorithm directly on the latent space of a JT-VAE is an effective way to explore diverse chemical structures while simultaneously optimizing multiple properties. This approach avoids direct manipulation of complex molecular representations.
  • Hybrid Training for Predictive Models: Incorporating both precise pKi values and less precise single-concentration binding data into model training (especially for hERG) was crucial. This allowed the models to cover a broader chemical space and improve predictive accuracy, even if it introduced some noise.
  • Applicability Domain (AD) Constraints: Explicitly including AD constraints in the cost function is vital. This prevents the GMD from generating molecules that are too dissimilar to the training data, ensuring that the predictions from the ML models remain reliable.
  • Scalability is Key: The ATOM-GMD platform's ability to handle millions of compounds and optimize numerous parameters simultaneously is a significant advantage for tackling complex drug discovery problems.

Key Results

  • Successful Lead Optimization: The ATOM-GMD platform was successfully applied to design potent and selective histamine H1 receptor antagonists.
  • Experimental Validation: 103 compounds were synthesized and tested.
  • Potent H1 Binders: Six of the tested compounds bound to H1 receptors with Ki values between 10 and 100 nM.
  • High Selectivity: These six compounds were also at least 100-fold selective against muscarinic M2 receptors, a key off-target to avoid.
  • Diversity Maintained: Analysis of compound similarity and clustering showed that the GMD process maintained a diverse population of molecules throughout the optimization, avoiding convergence to a small set of chemotypes.

Limitations and Future Work

  • hERG Model Limitations: The authors noted that the hERG model might have been operating outside its domain of applicability for some generated compounds, potentially leading to underprediction of binding. They suggest that using a model trained purely on binding data or incorporating more diverse training data could improve predictions.
  • BBB Penetration Not Addressed: For this specific study, blood-brain barrier penetration was not a design criterion, although it is an important factor for many H1 antagonists.
  • Future Work: The authors suggest that incorporating BBB penetration as a design criterion would be a logical next step. They also highlight the platform's flexibility for applying it to other drug discovery targets and optimizing different sets of properties. The ability to combine GMD with virtual screening of large databases and targeted analog design is presented as an attractive alternative to brute-force methods.

Summary updated

Abstract

ABSTRACT Generative molecular design (GMD) is an increasingly popular strategy for drug discovery, using machine learning models to propose, evaluate and optimize chemical structures against a set of target design criteria. We present the ATOM-GMD platform, a scalable multiprocessing framework to optimize many parameters simultaneously over large populations of proposed molecules. ATOM-GMD uses a junction tree variational autoencoder mapping structures to latent vectors, along with a genetic algorithm operating on latent vector elements, to search a diverse molecular space for compounds that meet the design criteria. We used the ATOM-GMD framework in a lead optimization case study to develop potent and selective histamine H1 receptor antagonists. We synthesized 103 of the top scoring compounds and measured their properties experimentally. Six of the tested compounds bind H1 with K i ’s between 10 and 100 nM and are at least 100-fold selective relative to muscarinic M2 receptors, validating the effectiveness of our GMD approach.
BibTeX
@misc{mcloughlin2023generative,
  author = {Kevin McLoughlin and Da Shi and Jeffrey E. Mast and John Bucci and John P. Williams and WILLIAM D Jones and Derrick Miyao and Luke Nam and Heather L. Osswald and L. A. Zegel'man and Jonathan Allen and Brian J. Bennion and Amanda K. Paulson and Ruben Abagyan and Martha S. Head and James M. Brase},
  title = {{Generative Molecular Design and Experimental Validation of Selective Histamine H1 Inhibitors}},
  year = {2023},
  doi = {10.1101/2023.02.14.528391},
  howpublished = {bioRxiv},
}
Figure