On the Unreported-Profile-is-Negative Assumption for Predictive Cheminformatics.

Area of Science:

• Cheminformatics
• Machine Learning
• Bioinformatics

Background:

• Compound-target binding profiles are crucial data in cheminformatics research.
• Many repositories only provide positive binding profiles, leading to the assumption that unreported profiles are negative.

Purpose of the Study:

• To evaluate the effectiveness of assuming unreported binding profiles are negative.
• To demonstrate the impact of this assumption on predictive model performance.
• To introduce a novel framework for improving predictive models by recovering missing binding profile data.

Main Methods:

• Utilizing compound-target binding profiles as features for predictive model training.
• Empirically assessing prediction performance degradation when the negative profile assumption fails.
• Implementing a framework for joint profile recovery and predictive model learning.
• Applying matrix recovery techniques to address the missing feature problem.

Main Results:

• Prediction performance significantly degrades when the assumption of unreported profiles being negative is incorrect.
• Explicit recovery of unreported binding profiles enhances predictive model performance.
• The proposed joint recovery and learning framework yields further performance improvements.

Conclusions:

• The assumption that unreported compound-target binding profiles are negative is often ineffective and can hinder predictive accuracy.
• Recovering these unknown profiles is essential for building robust predictive models in cheminformatics.
• This study introduces 'Learning with Positive and Unknown Features,' a new challenge in machine learning for handling incomplete datasets.