Identification of Active Molecules against Thrombocytopenia through Machine Learning.

Area of Science:

• Computational chemistry
• Pharmacology
• Bioinformatics

Background:

• Thrombocytopenia, often linked to thrombopoietin (TPO) deficiency, has limited treatment options and severe health risks.
• Traditional drug screening methods are inefficient for discovering novel thrombocytopenia therapeutics.
• Machine learning (ML) offers a promising approach to accelerate drug candidate discovery by exploring chemical structures.

Purpose of the Study:

• To computationally model drug-induced megakaryocyte differentiation and platelet production using ML.
• To identify key molecular features and structural characteristics associated with hematopoietic activity.
• To discover novel therapeutic agents for thrombocytopenia through in silico screening.

Main Methods:

• Developed 112 ML classifiers by integrating eight algorithms with 14 molecular features.
• Utilized 5-fold cross-validation and external validation for model performance assessment.
• Employed Shapley additive explanations (SHAP) for feature importance analysis and an ensemble strategy for integrated predictions.

Main Results:

• The top ML model achieved high accuracy (81.6% cross-validation, 83.1% external validation) and MCC (0.589, 0.642).
• SHAP analysis provided quantitative insights into molecular properties driving predictive accuracy.
• Ensemble models successfully predicted potential thrombocytopenia-active molecules from traditional Chinese medicine and drug repurposing databases.

Conclusions:

• ML models can effectively predict thrombopoiesis inducers, aiding in the discovery of thrombocytopenia treatments.
• Computational modeling provides valuable insights into structure-activity relationships for hematopoietic agents.
• This approach accelerates the identification of novel drug candidates for challenging conditions like thrombocytopenia.