Binding peptide generation for MHC Class I proteins with deep reinforcement learning.

Area of Science:

• Immunology
• Computational Biology
• Bioinformatics

Background:

• Major Histocompatibility Complex (MHC) Class I proteins are crucial in immunotherapy for presenting peptides to anti-tumor immune cells.
• Distinct peptide repertoires and binding motifs characterize different MHC Class I proteins.
• In vitro methods for characterizing MHC Class I binding motifs are infeasible for the vast number of known MHC Class I proteins.

Purpose of the Study:

• To develop a de novo generation framework for characterizing MHC Class I binding motifs.
• To generate peptide repertoires for any given MHC Class I protein.
• To provide a more efficient and scalable alternative to in vitro experiments for motif characterization.

Main Methods:

• Introduced PepPPO, a de novo generation framework utilizing a reinforcement learning agent.
• Employed a mutation policy within the reinforcement learning agent to transform random peptides into positively presented ones.
• Applied PepPPO to characterize binding motifs for approximately 10,000 human MHC Class I proteins.

Main Results:

• Computed MHC Class I binding motifs showed high similarity to experimentally derived motifs.
• Successfully characterized binding motifs for a large cohort of human MHC Class I proteins.
• Demonstrated the utility of computed motifs for rapid neoantigen screening at reduced computational cost compared to existing deep learning methods.

Conclusions:

• PepPPO offers an efficient and scalable computational approach for MHC Class I binding motif characterization.
• The framework accurately predicts peptide repertoires and binding motifs, complementing experimental data.
• This method accelerates neoantigen discovery, significantly impacting immunotherapy development.