Pair Potentials as Machine Learning Features.

Area of Science:

• Computational biology and bioinformatics
• Structural bioinformatics
• Machine learning in chemistry

Background:

• Atom pairwise potential functions are crucial for protein decoy detection in scoring functions.
• Extracting potential function parameters is feasible, but obtaining calculated atom pairwise energies is challenging.
• Machine learning (ML) offers new avenues for combining potential functions into novel models.

Purpose of the Study:

• To directly encode Amber force field parameters (ff94 and ff14SB) to calculate atom pairwise energies.
• To develop and evaluate machine learning models, specifically Random Forest (RF), combined with these encoded Amber potentials for protein decoy detection.
• To compare the performance of the developed RF models against other established scoring functions.

Main Methods:

• Directly utilized ff94 and ff14SB force field parameters from Amber to encode atom pairwise energies.
• Validated the encoded Amber potentials using single amino acid and dipeptide sets, achieving energy differences within ±0.06 kcal/mol.
• Integrated Random Forest (RF) models with the encoded Amber force field potentials and trained/tested on 224 protein native-decoy systems.

Main Results:

• The RF models incorporating ff94 and ff14SB force field parameters demonstrated superior performance in native structure detection compared to KECSA2, RWplus, DFIRE, dDFIRE, and GOAP.
• These novel RF models showed comparable performance to other scoring functions in identifying the best decoy.
• Including best decoy comparisons in RF model building enhanced both accuracy and best decoy detection capabilities.

Conclusions:

• The developed method effectively encodes Amber force field parameters to calculate atom pairwise energies, enabling the creation of high-performing ML scoring functions.
• Combining the Random Forest algorithm with Amber force field potentials yields a powerful and flexible approach for protein decoy detection.
• Both the Random Forest algorithm and the specific force field potentials are essential components for achieving optimal performance in ML scoring functions.