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

Structural alignment can reveal protein relationships that are difficult to recover from sequence alone, but repeatedly applying alignment methods such as DALI or TM-Align to databases containing millions of proteins is computationally expensive. We introduce DALI-Vec, a sequence-only retriever obtained by fine-tuning TM-Vec to reproduce DALI-derived similarity rankings from a minimal dataset of DALI alignments. DALI-Vec is trained with a pairwise ranking objective that favors DALI-consistent candidate ordering, together with a preservation penalty that limits distortion of the pretrained embedding space. On held-out AFDB ranking pairs, DALI supervision induces a graded change in embedding: stronger DALI matches show progressively larger increases in cosine similarity, while pairs without a reported DALI hit move apart. In matched comparisons on 400 SCOPe domains, compared against TM-Vec, DALI-Vec increases superfamily sensitivity from 0.656 to 0.686, while a broader-training variant increases fold sensitivity from 0.119 to 0.146. Across 73 SCOPe queries searched against 2.3 million clustered AFDB targets, DALI-Vec improves recovery of recorded DALI-positive targets and reduces the candidate depth needed for high-recall screening. These retrieval gains translate directly to coarse-to-fine structural search. Under a fixed budget of 1,000 structural-verification requests, the DALI-Vec-prefiltered workflow recovers approximately 21% more confirmed $Z\ge5$ hits than the TM-Vec-prefiltered workflow. Together, these results establish DALI-Vec as a sequence-based prefilter for efficient, DALI-quality protein retrieval at database scale.