DALI-Vec: DALI-Supervised Sequence Embeddings for High-Recall Protein Structural Retrieval

Structural alignment detects remote protein relationships but is too costly to apply exhaustively to million-scale databases. We introduce DALI-Vec, a sequence-only retriever adapted from TM-Vec using DALI-ranked training pairs and an embedding-preservation penalty. On 500 held-out AlphaFold Database (AFDB) queries searched against 20,000 random targets, DALI-Vec reduces the mean database fraction required for 95% recovery of recorded DALI hits ($Z\ge5$) from 29.45% to 22.55%, with improvements reproduced across three independent training seeds. On a separately assembled 55-domain SCOPe benchmark against 2.3 million AFDB targets, the same measure decreases from 30.6% to 23.2%. At a budget of 1,000 structural verifications per query, DALI-Vec recovers 19.6% more confirmed hits than TM-Vec; a 24-query timing audit shows a 1.13-fold increase in confirmed hits per DALI verification CPU-hour. Conformal filtering further reduces candidate retention on held-out queries drawn from the calibration distribution, although unchanged thresholds do not preserve the nominal missed-hit risk target when transferred to SCOPe. Together, these results support DALI-Vec as a sequence-only prefilter that improves candidate prioritization for database-scale structural search while retaining explicit structural alignment for verification.