Exploration and Exploitation Approaches Based on Generative Machine Learning to Identify Potent Small Molecule Inhibitors of α-Synuclein Secondary Nucleation

Robert I. Horne, Mhd Hussein Murtada, Donghui Huo, Z. Faidon Brotzakis, Rebecca C. Gregory, Andrea Possenti, Sean Chia, Michele VendruscoloJournal of Chemical Theory and Computation 2023 · DOI

TL;DR

This paper introduces a novel two-pronged machine learning strategy, combining generative modeling and reinforcement learning, to discover potent small molecule inhibitors of alpha-synuclein aggregation, a key factor in Parkinson's disease. By first exploring novel chemical spaces for blood-brain barrier penetration and then exploiting known chemical spaces for enhanced potency, they successfully identified promising drug candidates that could significantly impact Parkinson's disease therapeutics.

What Problem Does This Paper Address?

Parkinson's disease (PD) is a devastating neurodegenerative disorder for which no disease-modifying drugs currently exist. A major culprit is the aggregation of the alpha-synuclein (aS) protein into toxic oligomeric species. While many clinical trials have aimed to inhibit this aggregation, they have largely failed, partly due to the difficulty in identifying potent and effective small molecule inhibitors in preclinical stages. Specifically, targeting the secondary nucleation step of aS aggregation, which is a primary driver of oligomer production, is crucial. This paper addresses the challenge of efficiently discovering such inhibitors by leveraging advanced machine learning techniques to navigate complex chemical spaces and predict molecular efficacy.

Methods

The authors developed a two-pipeline approach:

  • Exploration Pipeline: This pipeline uses a generative model (GraphINVENT) to create novel, drug-like molecules predicted to cross the blood-brain barrier (BBB). This generative model is then fine-tuned using reinforcement learning, guided by two complementary QSAR (Quantitative Structure-Activity Relationship) models. These QSAR models predict a molecule's ability to inhibit aS secondary nucleation and its BBB penetrance. The reinforcement learning agent learns to generate molecules that maximize the scores from these QSAR models.
  • Exploitation Pipeline: This pipeline focuses on improving the potency of existing chemical structures. It employs a chemical language model (CLM) designed for low-data regimes. This model is trained using transfer learning (pre-trained on a large bioactive molecule dataset) and data augmentation techniques to generate molecules with enhanced anti-aggregation properties. The goal here is to refine known scaffolds into more potent inhibitors.

Both pipelines integrate QSAR models as filters, allowing the generative models to learn the structure-activity relationship in silico, thereby reducing the need for extensive experimental testing.

Datasets

  • CNS Drug Libraries (ChemDiv): Used as a source for molecules with predicted blood-brain barrier (BBB) penetrance.
  • B3DB dataset: A benchmark dataset of BBB molecules compiled from 50 published resources.
  • ChEMBL24: A large dataset of bioactive molecules used for pre-training the CLM in the exploitation pipeline.
  • MEGx natural product library: Used as a target space in the exploitation pipeline to indirectly optimize pharmacokinetics.
  • Experimental aggregation inhibition data: A dataset of 453 molecules previously tested for their ability to inhibit aS secondary nucleation. This dataset was crucial for training the QSAR models.

Key Insights and Ideas

  • Synergistic Exploration and Exploitation: The paper demonstrates the power of combining generative exploration (finding novel structures) with targeted exploitation (optimizing known structures) in a sequential manner. This hybrid approach can overcome limitations of each method alone.
  • Targeting Secondary Nucleation is Key: The research emphasizes the critical role of inhibiting secondary nucleation as the most effective strategy for reducing toxic aS oligomer production.
  • QSAR as Reward Functions: Integrating QSAR models directly into the generative process as reward functions significantly enhances the efficiency of drug discovery by guiding the generation of molecules with desired properties.
  • Importance of BBB Penetrance: For CNS disorders like PD, ensuring that potential drug candidates can cross the blood-brain barrier is a fundamental requirement, and the exploration pipeline effectively addresses this.
  • Descriptor-Based QSAR Outperforms SMILES Embeddings: For the anti-aggregation QSAR model, using chemical descriptors proved more robust and yielded better generalization and fewer false negatives compared to models based on SMILES embeddings.

Key Results

  • The exploration pipeline generated small molecules predicted to be BBB-penetrant and possess anti-aggregation activity. One such molecule showed mild inhibition, comparable to existing clinical inhibitors.
  • The exploitation pipeline identified a novel compound, CLM.1, with a KIC50 of 0.42 µM, which represents a significant improvement over previously identified hits (KIC50 of 0.52 µM) and existing compounds like Anle-138b (extrapolated KIC50 of 18.2 µM).
  • The descriptor-based QSAR model achieved a high accuracy of 0.98 and an average AUC score of 0.85, with no false positives.
  • The study successfully demonstrated that machine learning can predict small molecules with a balance of drug-likeness, BBB penetrance, and aggregation inhibition.

Limitations and Future Work

  • Limited Experimental Data: The authors acknowledge that the size and imbalance of the experimental aggregation inhibition dataset presented challenges. While data augmentation and transfer learning were employed, a larger, more diverse dataset would further improve model performance.
  • Availability of Generated Molecules: Many molecules generated by the exploration pipeline were not readily available and would require custom synthesis, limiting immediate experimental validation.
  • In Vitro vs. In Vivo Efficacy: The aggregation assay used is an in vitro proxy for disease mechanisms. While it showed promise, further validation in in vivo models is necessary.
  • Future Work: The authors suggest that the performance of the models could improve with more varied training data. They also highlight the potential for this combined exploration-exploitation strategy to be applied to other protein misfolding diseases and drug discovery challenges. The ability to optimize for multiple parameters simultaneously in the exploration pipeline is also a promising avenue for future development.

Summary updated

Abstract

The high attrition rate in drug discovery pipelines is an especially pressing issue for Parkinson's disease, for which no disease-modifying drugs have yet been approved. Numerous clinical trials targeting α-synuclein aggregation have failed, at least in part due to the challenges in identifying potent compounds in preclinical investigations. To address this problem, we present a machine learning approach that combines generative modeling and reinforcement learning to identify small molecules that perturb the kinetics of aggregation in a manner that reduces the production of oligomeric species. Training data were obtained by an assay reporting on the degree of inhibition of secondary nucleation, which is the most important mechanism of α-synuclein oligomer production. This approach resulted in the identification of small molecules with high potency against secondary nucleation.
BibTeX
@article{horne2023exploration,
  author = {Robert I. Horne and Mhd Hussein Murtada and Donghui Huo and Z. Faidon Brotzakis and Rebecca C. Gregory and Andrea Possenti and Sean Chia and Michele Vendruscolo},
  title = {{Exploration and Exploitation Approaches Based on Generative Machine Learning to Identify Potent Small Molecule Inhibitors of α-Synuclein Secondary Nucleation}},
  journal = {Journal of Chemical Theory and Computation},
  year = {2023},
  doi = {10.1021/acs.jctc.2c01303},
}
Figure