Computational chemogenomics: is it more than inductive transfer?

Area of Science:

• Computational chemistry
• Cheminformatics
• Pharmacology

Background:

• High-throughput screening generates vast multi-ligand, multi-target activity data.
• Quantitative Structure-Activity Relationship (QSAR) models traditionally focus on single targets.
• Computational chemogenomics (CG) offers a holistic approach to learning from complex biological data.

Purpose of the Study:

• To investigate the interplay between inductive transfer (IT) and explicit learning (EL) in CG models.
• To determine if including protein information (EL) enhances predictive capabilities beyond IT alone.
• To assess the effectiveness of CG models in deorphanization challenges compared to classical QSAR.

Main Methods:

• Support Vector Regression was employed using over 9,400 pKi values for 31 GPCRs.
• Compound-protein interactions were modeled using concatenated vectorial descriptions.
• Models were compared based on cross-validation and deorphanization performance, with and without explicit protein descriptors.

Main Results:

• EL-enabled CG models showed improved performance over classical QSAR but were not significantly better than IT-enhanced models.
• Explicit learning strategies did not demonstrate superiority in deorphanization tasks.
• The primary benefit of CG models appears to stem from IT rather than the explicit inclusion of protein information.

Conclusions:

• The benefits of CG models may be largely attributed to IT, facilitated by simultaneous learning across multiple targets.
• Current protein descriptors may not be sufficiently informative to unlock the full potential of EL in CG.
• Further advancements in protein descriptor development are crucial for realizing the anticipated advantages of explicit learning in chemogenomics.