Prediction of Compound Bioactivities Using Heat-Diffusion Equation.

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning applications

Background:

• Cell-based phenotypic screening faces challenges like low throughput, high costs, imbalanced data, and unpredictable new chemical structures.
• Existing machine learning models struggle with the complexity and non-linear predictions inherent in phenotypic screening.

Purpose of the Study:

• To develop an advanced prediction model, the Prediction Model based on the Heat-Diffusion Equation (PM-HDE), to overcome limitations in machine learning for phenotypic screening.
• To validate the PM-HDE model's efficacy in virtual compound screening and its application in actual drug discovery pipelines.

Main Methods:

• Developed the Prediction Model based on the Heat-Diffusion Equation (PM-HDE) algorithm.
• Validated PM-HDE using biotest data from 946 assay systems in PubChem for virtual compound screening.
• Applied PM-HDE to supervised learning on ~50,000 compounds from motor neuron screening (ALS-patient-induced pluripotent stem cells) for large-scale virtual screening (>1.6 million compounds).

Main Results:

• The PM-HDE algorithm demonstrated feasibility for virtual compound screening.
• PM-HDE successfully enriched hit compounds in large-scale virtual screening efforts.
• The model identified novel chemical structures (chemotypes), indicating its potential for discovering new therapeutic agents.

Conclusions:

• The PM-HDE model offers a solution to the inflexibility and challenges associated with applying machine learning to complex phenotypic screening.
• This novel approach provides a powerful platform for accelerating drug discovery by improving the efficiency and effectiveness of screening processes.