Two protein design models redesigned T cell receptor contact points on solved structures

What this research was about

T cells patrol the body looking for signs of trouble. Each one carries a receptor on its surface that reads short fragments of protein held up by a molecule called the MHC, a kind of display case that shows the cell's contents to the immune system. The receptor grips this fragment-and-case pairing through a handful of flexible loops, known as CDRs. Designing a receptor that grips a chosen target is hard because the loops are small, the contacts are subtle, and the same shape can be made from many different amino acid sequences. Immune receptors also carry a further constraint: real ones are assembled by a genetic shuffling process, so not every sequence is one the body could make.

Some deep learning tools now work backwards from shape to sequence, a task called inverse folding: given the three-dimensional skeleton of a protein, they suggest amino acids that would fit it. The researchers set out to test how well two such tools behave when pointed at T cell receptors. They collected solved receptor-MHC structures from a database, keeping only ones dated after the models' training data ended, so the models could not simply be recalling them.

Where AI came in

ProteinMPNN and ESM-IF1 were run with their released settings, nothing retrained, to generate new receptor sequences on the fixed native skeleton. Design scopes ranged from the CDR3 residues closest to the fragment or the MHC up to the whole variable domain; the rest of the sequence was held fixed, and the bound target was supplied as context except in one control where it was removed. In place of these models, the usual approach is a physics-based design program, and Rosetta was run over the same positions for comparison. Sequence recovery at the designed CDR3 contact positions averaged 43.9 per cent for ProteinMPNN and 50.1 per cent for ESM-IF1 on MHC-I cases, against 31.7 per cent for Rosetta.

A third model, TCRModel2, a receptor-focused version of AlphaFold2, was then used to predict structures for the designed sequences so their backbones and interfaces could be compared with the original crystal structures. Everything else in the assessment was non-learned: counts of germline and natural CDR3 variability, Rosetta interface scoring, and molecular dynamics simulations with free energy calculations. In those simulations, covering seven test cases, most designs showed no significant difference in binding free energy from the native receptor, with lower values than native in one case for ProteinMPNN and three for ESM-IF1. Removing the target from the design input cut maximum sequence recovery by roughly 20 per cent for both models.

Written by AIxSci from the checked record below, to give context for readers outside the field. It is not part of the record.