Predicting clinically promising therapeutic hypotheses using tensor factorization.

Area of Science:

• Computational biology
• Machine learning in drug discovery

Background:

• Drug target identification is a costly and error-prone initial step in therapeutic development.
• Machine learning and informatics can computationally predict promising drug targets based on target-indication evidence.
• Data from 17 sources on target-indication associations were compiled into a tensor.

Purpose of the Study:

• To benchmark machine learning models for predicting clinical outcomes of therapeutic hypotheses.
• To evaluate tensor factorization's performance against other machine learning methods.

Main Methods:

• Compiled a gold-standard dataset of 6140 clinical outcomes from 875 targets and 574 indications.
• Benchmarked Logistic Regression, LASSO, Random Forest, Tensor Factorization, and Gradient Boosting Machine models.
• Utilized 10-fold cross-validation and biologically-motivated cross-validation schemes.

Main Results:

• Tensor factorization achieved an Area Under the Receiver Operating Characteristic Curve (AUROC) of 0.82 ± 0.02 and an Area Under the Precision-Recall Curve (AUPRC) of 0.71 ± 0.03.
• Tensor factorization's performance was comparable or superior to other evaluated machine learning models.
• Demonstrated application in predicting outcomes for novel indications of approved drug targets.

Conclusions:

• Tensor factorization is an effective machine learning technique for predicting clinical outcomes of therapeutic hypotheses.
• The method shows significant potential for identifying novel drug targets and indications, advancing drug discovery.
• Biologically-motivated cross-validation provides robust insights into prediction performance for future methods.