Extending in Silico Protein Target Prediction Models to Include Functional Effects.

Area of Science:

• Computational chemistry and cheminformatics
• Drug discovery and mechanism-of-action studies
• Machine learning in pharmacology

Background:

• In silico protein target deconvolution is vital for understanding drug mechanisms but often lacks functional effect prediction (activation/inhibition).
• Existing methods primarily focus on compound-target binding, neglecting the downstream functional consequences of this interaction.
• Integrating functional effects into target prediction models is essential for a comprehensive understanding of small molecule activity.

Purpose of the Study:

• To develop and evaluate computational models that predict the functional effects (activation or inhibition) of small molecules on protein targets.
• To compare the performance of different model architectures, including simple random forest and cascaded approaches, for functional effect prediction.
• To identify the most suitable architecture for predicting functional responses and assess its applicability domain for novel chemical spaces.

Main Methods:

• Assimilation of a large bioactivity dataset including binding, activating, and inhibiting data points across 332 targets.
• Development of three distinct model architectures: a baseline random forest (Arch1) and two cascaded models (Arch2, Arch3) with varying training set strategies.
• Rigorous evaluation using fivefold stratified cross-validation and prospective validation on a temporal dataset, including applicability domain analysis.

Main Results:

• Chemical space analysis revealed distinct clustering for binding, activating, and inhibiting compounds, supporting the need for functional prediction.
• Cascaded models (Arch2 and Arch3) demonstrated superior precision and recall compared to the baseline model in cross-validation.
• Arch3, utilizing inactive background sets, achieved the highest average precision (71%) and recall (53%) in prospective validation and showed better extrapolation capabilities.
• A case study on CHRM1 highlighted the model's ability to annotate unanticipated functional changes.

Conclusions:

• The developed cascaded model (Arch3) is proposed as the most suitable architecture for predicting small molecule functional effects due to its high performance and broad applicability.
• Incorporating functional effect prediction significantly enhances the utility of in silico target deconvolution for mechanism-of-action studies.
• This approach provides vital insights into potential off-target effects and functional changes, crucial for drug development and safety assessment.