Discrimination vs. Generation: The Machine Learning Dichotomy for Dopaminergic Hit Discovery

Temitope Sobodu,Adeshina Yusuf,Dan Kiel,Dong Kong

2024-09-19

Abstract:Virtual screening plays a pivotal role in early drug discovery, traditionally dominated by physics-based methods. While these approaches offer detailed insights, they are often hindered by high computational costs, limited sampling, and forcefield inaccuracies. Advances in Machine Learning (Ml)and Deep Learning (DL) present resource-efficient alternatives, with approaches like predictive geometric ML (EQUIBIND) and generative geometric ML (DIFFDOCK)showing promise in enhancing both efficiency and predictive capability. Here, we compare these two strategies, retrospectively and prospectively, for identifying novel agonists targeting the dopamine D2 receptor. To complement DIFFDOCK's dual functionality in protein-ligand conformer generation and confidence estimation, we adopted a complementary atom-type-based confidence model for EQUIBIND. This pipeline, termed the discriminative model, integrates a featurization step and an XGBoost classifier to differentiate between active and inactive ligands. The top-ranked compounds from both models were evaluated using an ultrafast dopaminergic biosensor assay, dLight. Our results demonstrate that the generative model achieved a higher hit rate, notably leading to the discovery of Compound 1, a nanomolar dopamine D2 receptor agonist with a novel scaffold.

Biomolecules

What problem does this paper attempt to address?

The aim of this paper is to address the problem of comparing and evaluating the effectiveness of two machine learning (ML) strategies in the discovery process of dopamine D2 receptor agonists: discriminative models and generative models. Specifically, the objectives of the study include:

1. **Enhancing the efficiency of virtual screening**: By utilizing machine learning methods (such as predictive geometric deep learning and generative geometric deep learning), improve the efficiency and predictive capability of virtual screening, reducing the high computational cost, limited sampling, and force field inaccuracies associated with traditional physics-based methods.

2. **Discovering novel agonists**: By comparing the performance of the two models (discriminative models and generative models) in identifying novel dopamine D2 receptor agonists, explore their potential applications in real-world drug discovery.

3. **Model validation**: Validate the effectiveness of these machine learning methods in experimental settings, particularly in real-world drug discovery tasks, by evaluating the compounds predicted by the models through biosensor experiments (dLight).

Through this research, the authors hope to provide more efficient and accurate methods for drug discovery and to identify new chemical types, thereby offering potential new therapies for diseases related to dopamine dysregulation (such as attention deficit hyperactivity disorder, addiction, depression, schizophrenia, Parkinson's disease, etc.).