Molecular Assays Simulator to Unravel Predictors Hacking in Goal-Directed Molecular Generations

Area of Science:

• Computational chemistry and cheminformatics
• Machine learning in drug discovery
• Bioinformatics and computational biology

Background:

• Generative models are widely used in drug discovery, often integrated with ADME/QSAR models for property optimization.
• A significant challenge is the generation of false positives, where predicted active molecules are inactive upon testing, due to predictive model 'hacking'.
• Existing in silico oracles for evaluating generated molecules are often too simplistic and limited to single objectives, failing to capture complex biological realities.

Purpose of the Study:

• To introduce a novel simulator for multitarget assays using a neural network (NN) as a more realistic in silico oracle.
• To replicate a lead optimization (LO) scenario using generative models and the developed oracle.
• To investigate and propose methods for mitigating the 'hacking' issue in generative model-based drug discovery pipelines.

Main Methods:

• Developed a neural network-based simulator for multitarget assays, providing continuous real-valued outputs for molecular properties.
• Trained predictive models on an initial dataset to predict oracle values for molecules.
• Utilized the open-source GuacaMol package to generate optimized molecules, coupled with predictive models, and evaluated them using the NN-based oracle.

Main Results:

• Even with high predictive model performance metrics, the final selected compounds contained numerous false positives when evaluated by the NN-based oracle.
• The 'hacking' phenomenon was observed in both mono- and bi-objective optimization scenarios.
• Several strategies were proposed and evaluated to address and mitigate the 'hacking' issue in generative model optimization.

Conclusions:

• The proposed NN-based multitarget assay simulator provides a more robust evaluation of generated molecules compared to simpler oracles.
• The 'hacking' of predictive models remains a critical challenge in generative drug discovery, even with advanced simulation tools.
• Further research into mitigation strategies is crucial for the reliable application of generative models in drug discovery.