Fig. 2: TM-Vec structural similarity prediction.

a–d, Two TM-Vec models were built and benchmarked against protein pairs from SWISS-MODEL and CATHS40. a, SWISS-MODEL TM-score prediction errors (absolute value of difference between the known TM-score from running TM-align on structures and the TM-Vec-predicted TM-score) for 1.01 million pairs with different sequence identities. Sequence similarity as measured by sequence identity ranges from [0, 0.1) (least similar) to (0.9, 1.0] (most similar). b, TM-Vec absolute value of prediction error obtained from protein sequences compared with TM-scores from TM-align obtained from protein structures. Prediction errors were stratified across 681,000 proteins from three test benchmarking datasets: pairs, domains and folds. The pairs test dataset included protein sequence pairs that were left out of model training and/or validation. Similarly, the domains and folds test dataset included protein pairs derived from domains and folds that were never seen in model training and/or validation. Bounds of the boxplots denote 25% and 75% percentiles, the center is the 50% percentile and the whiskers denote the 1.5× interquartile range. c, t-SNE (t-distributed stochastic neighbor embedding) visualization of protein embeddings from the top five most represented categories from each CATH classification tier (class, topology, architecture, homology) within the test dataset. For each CATH classification tier, TM-Vec embeddings were observed to separate structural categories better than the default protein sequence embeddings generated by ProtTrans. d, Quantitative benchmarks of the ability of TM-Vec to predict CATH labels. We compared with ProtTrans and five structure-based methods: cliques, GRAFENE, ORCA, CNN (influenced by DeepFRI) and GCN (influenced by the Kipf and Welling GAE). Adjusted mutual information was computed by comparing spectral clustering assignments with structural label assignments for each CATH classification tier. Triplet-scoring AUPR is a metric that determines how often cosine embedding distances from within structural categories are smaller than cosine embedding distances across structural categories.