Rprot-Vec

Structural similarity tells you more about protein function than sequence identity does — two proteins with divergent sequences and the same fold usually share a job and an evolutionary origin. The problem is that measuring it has traditionally meant having structures. Fewer than 0.1% of known protein sequences have an experimentally solved one, so any method requiring structure leaves most of the sequence universe unreachable.

Rprot-Vec — Rapid Protein Vector — from Yichuan Zhang at the University of Tokyo and Wen Zhang at Florida Atlantic University predicts the TM-score between two proteins directly from their amino acid sequences. Each sequence is embedded into a fixed-length vector; the cosine similarity between two such vectors is trained to reproduce the TM-score that structural alignment would have returned. Homology search then reduces to encoding a query once and scanning cosine similarities against a precomputed vector database.

The design goal is deliberately narrow: a small, fast model that trains well on modest data. Rather than stacking transformer blocks, the network puts a bidirectional GRU and multi-scale convolutions on top of a frozen protein language model, giving a trainable head of 7.2 million parameters. It occupies the sequence-only structural search niche also served by TM-Vec 2, from an independent author group; Rprot-Vec's own comparisons are against the original TM-Vec. Posted to bioRxiv in January 2025, it was published in BMC Bioinformatics in July 2025.

#Key Features

• Frozen language model front end: ProtT5-XL-UniRef50 encodes each residue into a 1,024-dimensional context-aware vector and is not updated during training, so the trainable network stays small.
• Recurrent plus convolutional head: A bidirectional GRU with an attention layer captures global sequence dependencies, while convolution kernels of size 3 and 7 pick up local features at two scales.
• Cosine similarity as the score: Adaptive average pooling and a fully connected layer produce a fixed-length embedding whose positive-half cosine similarity is the predicted TM-score, so identical folds overlap in vector space and unrelated ones fall orthogonal.
• Released training corpora: Three CATH-derived, TM-score-balanced datasets ship alongside the model, filling a gap left by earlier work whose training data was never released.

#Technical Details

Training data was generated by downloading roughly 120,000 CATH sequences with corresponding PDB structures and running US-align over pairs for ground-truth TM-scores — about 1.08 billion raw pairs totaling 160 GB, produced over three months. Because most random protein pairs are structurally unrelated, the raw set was filtered to spread evenly across the 0–1 TM-score range, yielding CATH_TM_score_S, M, and L at 3.0M, 8.5M, and 23.3M entries (492 MB, 1.3 GB, 3.6 GB). All reported experiments use the M dataset, split 95:5. Optimization used Adam at a learning rate of 1e-4, batch size 64, and mean-squared-error loss, on a single RTX 4070 Ti.

Across the full TM-score range, Rprot-Vec's average prediction error is 0.0561, against 0.0981 for the published TM-Vec model and 0.0717 for a TM-Vec retrained locally on the same data for a fair comparison. In the homologous region (TM-score above 0.8, where fold-family membership is near certain) the error is 0.0438, against 0.0738 and 0.0462, with 65.3% of pairs predicted accurately and the published TM-Vec showing a 5.7% failure rate at error above 0.2. The overall improvements are 21.8% across the full interval and 5.19% in the homologous region, achieved with 41% of TM-Vec's parameter count. Predicted TM-scores correlate with true values at Pearson 0.965, while absolute error correlates with sequence identity at only -0.096 — evidence the model captures structural signal that sequence similarity alone does not carry, particularly in the low-to-middle band where remote homologs live.

#Applications

The intended use is fast structural screening where computing or aligning structures is impractical: annotating uncharacterized proteins by structural neighborhood, building an encoded database that turns homology search into a cosine scan, and comparing protein therapeutics to one another. The authors demonstrate the last of these on DrugBank protein drugs, where pairwise embedding similarity flagged ecallantide and aprotinin as structurally close — both serine protease inhibitors, one narrowly selective and one broad-spectrum — illustrating how the method could nominate repositioning candidates for experimental follow-up.

#Impact

Rprot-Vec is a compact contribution to a useful niche, and its most durable piece may be the datasets: three balanced CATH-derived TM-score corpora that later models can train and benchmark on, released precisely because their absence made earlier comparisons difficult. The accuracy figures are modest in absolute terms — 65.3% within the homologous region is a starting point, not a solved problem — and all comparisons are against TM-Vec on the authors' own data, without an independent benchmark or a structure-alignment baseline such as Foldseek. The journal article is open access under CC BY-NC-ND; the GitHub repository provides training and testing notebooks with the CATH download and dataset-generation scripts, and carries no license file.