MetaRT: Retention time (RT) prediction based on 3D structure of peptides from limited training dataset

MetaRT is a prediction tool that can predict the retention times of hydrophobic peptides by exploiting limited training dataset. MetaRT employs stacked ensemble machine learning approach that combines multiple base algorithms through meta-learner. The role of meta-learner is to optimally combine the predictions of base learners to produce the final predictions.

Overview

• Prediction of peptide retention time by stacked ensemble learning
• Accurate prediction with limited training data samples of peptides
• Feature extraction from 3D structures of peptide using graph neural network (GNN)
• Prediction precision (R²): 0.991
• Discriminate isomeric peptides using RT

Library

DeepChem V2.8.0, Scikit-learn V1.5.2, RDKit 2024.3.6, Torch-geometric V2.6.1

Datasets

Training and validation datasets are kept same while performing the entire predictions. The testing datasets are different for hydrophobic peptides, isomeric sequences and less hydrophobic peptides.

• Structure-based dataset contains seq, smiles, inchi, M+H, rt.
• The descriptor dataset contains hydrophobicity, binding_energy, solvation_energy_wat, solvation_energy_acn, total_sasa, side_chain_total_sasa, nAtom, nHeavyAtom, nC, nBonds, nBondsKS, nBondsKD, GeomDiameter, GeomRadius, FilterItLogS, apol, SLogP

Structure-based RT Prediction

● ΔRT prediction error range of hydrophobic peptides: 0.07 - 1.4 min

● ΔRT prediction error range of sequence isomers: 0.2 - 1.3 min

Questions

If you have any question, please contact at metart.pred@gmail.com