Unbiased descriptor and parameter selection confirms the potential of proteochemometric modelling

Area of Science:

• Computational chemistry
• Bioinformatics
• Cheminformatics

Background:

• Proteochemometrics predicts protein function using interaction data, bypassing 3D structures.
• Existing models offer insights into biomolecular interactions but lack rigorous statistical evaluation.
• Variable selection in current methods often uses all data, similar to gene expression analysis.

Purpose of the Study:

• To implement a methodology for unbiased evaluation of proteochemometric model predictive power.
• To assess the performance of proteochemometric models on large datasets.
• To address limitations in current statistical evaluation and interpretation of these models.

Main Methods:

• Developed and applied a double cross-validation (CV) loop procedure.
• Estimated expected performance of proteochemometric design methods.
• Utilized two large-scale proteochemometric datasets for evaluation.

Main Results:

• Unbiased performance estimates (P2) confirm useful predictive power of well-designed single proteochemometric models.
• Standard cross-validation may yield models with limited performance.
• Commercial software can produce misleading performance estimates; chemical interpretation of single models is uncertain with small datasets.

Conclusions:

• The double CV loop provides unbiased estimates, identifying non-predictive proteochemometric models.
• Chemical interpretations should be based on multiple models from the double CV loop, not single models.
• This approach enhances the reliability and interpretability of proteochemometric modeling.