Learning to Predict Drug Target Interaction From Missing Not at Random Labels.

Area of Science:

• Bioinformatics
• Computational Biology
• Drug Discovery

Background:

• Drug-Target Interaction (DTI) prediction is vital for accelerating new drug development.
• Current DTI prediction methods use binary classification but suffer from incomplete supervision, leading to issues with overwhelming negative samples and ambiguous labels.
• The assumption of random label distribution is flawed, as negative DTI labels are often missing due to research prioritization.

Purpose of the Study:

• To address the challenge of learning from incomplete labels in DTI prediction.
• To develop a model that accounts for the non-random nature of missing DTI labels.
• To improve the accuracy and robustness of DTI prediction, especially for imbalanced datasets.

Main Methods:

• Introduced a novel probabilistic model, Factorization with Non-Random Missing Labels (FNML).
• Modeled the generative process of DTI labels and their observation status, linking label missingness to the label's sign.
• Developed an ensemble scheme, FNML-EN, to enhance prediction accuracy and reduce variance on imbalanced DTI data.

Main Results:

• Demonstrated superior and robust performance of the proposed FNML and FNML-EN models.
• Showcased the effectiveness of accounting for non-random missing labels in DTI prediction.
• Validated the models on a comprehensive, up-to-date DTI database.

Conclusions:

• The proposed FNML model effectively handles incomplete labels in DTI prediction by considering non-random missingness.
• The FNML-EN ensemble further improves prediction accuracy and robustness, particularly for imbalanced datasets.
• This approach offers a significant advancement in computational drug discovery by providing more reliable DTI predictions.