Impact of Chemist-In-The-Loop Molecular Representations on Machine Learning Outcomes

Area of Science:

• Cheminformatics
• Computational Chemistry
• Bioactivity Prediction

Background:

• Molecular descriptors are crucial for cheminformatics, with current methods relying on algorithms or machine learning.
• A key question is how these computational approaches compare to the expertise of trained chemists.
• The CAS fingerprint uniquely incorporates curated chemical motifs identified by expert chemists for potential bioactivity influence.

Purpose of the Study:

• To benchmark the performance of the CAS fingerprint against commonly used molecular fingerprints.
• To evaluate the predictive power of expert-curated features in cheminformatics.
• To explore the novelty and information content of the CAS fingerprint.

Main Methods:

• Benchmarking the CAS fingerprint against established molecular fingerprints.
• Utilizing a well-validated benchmark dataset comprising 88 biological targets.
• Analyzing the features selected by expert chemists for the CAS fingerprint.

Main Results:

• The CAS fingerprint demonstrated superior performance compared to most commonly used molecular fingerprints.
• Expert-selected features for the CAS fingerprint included novel motifs not frequently reported in scientific literature.
• The CAS fingerprint provides distinct and complementary information compared to other fingerprints.

Conclusions:

• Expert human insight, or anthropomorphic insights, significantly contributes to predictive power in molecular descriptor development.
• A chemist-in-the-loop approach is valuable, even in the age of advanced machine learning.
• The CAS fingerprint represents a promising cheminformatics tool leveraging expert knowledge for enhanced bioactivity prediction.