A semi-supervised learning framework for quantitative structure-activity regression modelling.

Area of Science:

• Computational chemistry
• cheminformatics
• drug discovery

Background:

• Quantitative Structure-Activity Relationship (QSAR) methods are vital for preclinical small molecule drug discovery.
• Supervised learning models in QSAR are trained on molecular structure representations and activity data to predict novel compound activities.

Purpose of the Study:

• To develop methods addressing key challenges in QSAR modeling.
• To compare information content of molecular fingerprints relative to a specific target.
• To quantify prediction accuracy degradation based on data distance and adjust for selection bias.

Main Methods:

• A novel method for comparing information content between molecular fingerprints.
• A method to quantify prediction accuracy degradation as a function of data distance.
• A method to adjust for screening-dependent selection bias in training datasets.
• A semi-supervised learning framework integrating data distance and bias adjustment.

Main Results:

• The developed methods were illustrated using the Tres Cantos AntiMalarial Set (TCAMS) data.
• The study demonstrates improved QSAR modeling by addressing fingerprint information content and data biases.
• The semi-supervised framework provides predictions considering compound similarity and reporting bias.

Conclusions:

• The presented methods offer significant advancements for QSAR modeling in drug discovery.
• These techniques can lead to more accurate predictions of novel compound activities.
• The study provides tools to overcome common limitations in QSAR training datasets.