Fig. 1: Protein foundation models can be fine-tuned to predict a protein’s homo-oligomer symmetry.
From: Rapid and accurate prediction of protein homo-oligomer symmetry using Seq2Symm
a Schematic showing our modeling setup for multi-label prediction of homo-oligomer symmetry, illustrated for the bovine seminal ribonuclease protein (PDB id: 11bg). The input can be either the protein amino-acid sequence and/or the multiple sequence alignment (MSA). The ‘protein foundation model’ (pFM) can be ESM-MSA, ESM2, or RoseTTAFold2 (RF2). We experiment with various architectures for the ‘classifier head’ (see Methods). We vary the number of layers we fine-tune in the pFM, from a fully frozen model with a single trainable prediction head (i.e., “pre-trained only”) to a model with all weights freely tunable (i.e., “fine-tuned”). b The homo-oligomer symmetry prediction can then be supplied to a structure prediction algorithm (e.g., AlphaFold, RoseTTAFold2, or ESMFold) to guide the generation of an atomic-resolution homo-oligomer structure.
