Fig. 3: Annotating and aligning proteins in the Malidup benchmark.

a, Comparison of different sequence and structural alignment methods with DeepBLAST and TM-Vec. DeepBLAST, Needleman–Wunsch and Smith–Waterman are sequence alignment methods, whereas Fast, Dali, Mammoth and TM-align are structural alignment methods. The y axis represents the predicted TM-score (for the alignment methods, this is given by a predicted alignment), and the x axis represents the TM-score from a manually curated alignment. The performance of TM-Vec was comparable with that of structural alignment methods, and its trend line overlapped with that of TM-align. The performance of DeepBLAST was similar to that of Mammoth, a structure alignment method, and it outperformed the other sequence alignment method, Needleman–Wunsch. Data are presented as mean values estimated with a locally estimated scatterplot smoothing fit with 95% confidence intervals. b, A predicted alignment of two duplicated Annexin domains from Malidup, where DeepBLAST could accurately align (TM-score = 0.81) and Needleman–Wunsch struggled to align (TM-score = 0.33). c, Manual alignment of the two duplicated Annexin domains; the agreement with DeepBLAST is highlighted. d, Visualization of the manual structural alignment of the Malidup; the chains that DeepBLAST aligned correctly are highlighted in yellow.