Rprot-Vec: a deep learning approach for fast protein structure similarity calculation.

backgroundPredicting protein structural similarity and detecting homologous sequences remain fundamental and challenging tasks in computational biology. Accurate identification of structural homologs enables function inference for newly discovered or unannotated proteins. Traditional approaches often require full 3D structural data, which is unavailable for most proteins. Thus, there is a need for sequence-based methods capable of inferring structural similarity efficiently and at scale.

resultWe present Rprot-Vec (Rapid Protein Vector), a deep learning model that predicts protein structural similarity and performs homology detection using only primary sequence data. The model integrates bidirectional GRU and multi-scale CNN layers with ProtT5-based encoding, enabling accurate and fast similarity estimation. Rprot-Vec achieves a 65.3% accurate similarity prediction rate in the homologous region (TM-score > 0.8), with an average prediction error of 0.0561 across all TM-score intervals. Despite having only 41% of the parameters of TM-vec, Rprot-Vec outperforms both public and locally trained TM-vec baselines in all tested settings. Additionally, we constructed and released three curated training datasets (CATH_TM_score_S/M/L), supporting further research in this area.

conclusionRprot-Vec offers a fast and lightweight solution for sequence-based structural similarity prediction. It can be applied in protein homology detection, structure-function inference, drug repurposing, and other downstream biological tasks. Its open-source availability and released datasets facilitate broader adoption and further development by the research community.