PANDORA v2.0: Benchmarking peptide-MHC II models and software improvements

Abstract

T-cell specificity to differentiate between self and non-self relies on T-cell receptor (TCR) recognition of peptides presented by the Major Histocompatibility Complex (MHC). Investigations into the three-dimensional (3D) structures of peptide:MHC (pMHC) complexes have provided valuable insights of MHC functions. Given the limited availability of experimental pMHC structures and considerable diversity of peptides and MHC alleles, it calls for the development of efficient and reliable computational approaches for modeling pMHC structures. Here we present an update of PANDORA and the systematic evaluation of its performance in modelling 3D structures of pMHC class II complexes (pMHC-II), which play a key role in the cancer immune response. PANDORA is a modelling software that can build low-energy models in a few minutes by restraining peptide residues inside the MHC-II binding groove. We benchmarked PANDORA on 136 experimentally determined pMHC-II structures covering 44 unique alpha;beta chain pairs. Our pipeline achieves a median backbone Ligand-Root Mean Squared Deviation (L-RMSD) of 0.42 Angstrom on the binding core and 0.88 Angstrom on the whole peptide for the benchmark dataset. We incorporated software improvements to make PANDORA a pan-allele framework and improved the user interface and software quality. Its computational efficiency allows enriching the wealth of pMHC binding affinity and mass spectrometry data with 3D models. These models can be used as a starting point for molecular dynamics simulations or structure-boosted deep learning algorithms to identify MHC-binding peptides. PANDORA is available as a Python package through Conda or as a source installation at https://github.com/X-lab-3D/PANDORA .

Improving generalizability for MHC-I binding peptide predictions through geometric deep learning

“…This configuration aims at simulating real-case scenarios in which the network has to provide a prediction on alleles it has never seen or seen infrequently. We build 3D models for all these binding affinity data using PANDORA 33 , which served as input of our GDL approaches ( Fig. 1E ).…”

“…E) Data enrichment through 3D modeling. We used PANDORA 33 to generate 20 3D models of each of the 100 178 pMHC data points, thus enriching the sequence information with physics-derived information, such as geometry and physico-chemical features.…”

“…Expected polar interactions and hydrogen bonds are shown as yellow dashed lines. Structural models are generated by PANDORA 33 . Visualizations are done in PYMOL 46 .…”

“…The BA dataset was processed with PANDORA (v2.0beta2) software 33 to generate 3D models of the pMHC complexes for these binding affinity data. For every case, PANDORA uses NetMHCpan4.1 10 to predict the peptide’s binding core, using it to select the anchor residues to restraint.…”