OpenAntigens: a structure-aware database for antigen construct design across the human cell-surface and secreted proteome

Key results

1. ShownThe 2026-07-23 release covers 5328 human targets and generates 55 305 construct suggestions, which are hypotheses rather than expression-success predictions.

   5328 targets retained from 6364 Swiss-Prot candidates (1036 excluded). Topology buckets: 1959 multipass, 1725 secreted, 1507 single-pass, 137 GPI-anchored. All 5328 reports built with no error records. · Target universe and release snapshot; Construct recommendation workflow; Release build and target processing; Figure legends

2. ShownEvidence coverage is high for predicted structure, disease link and cross-reactivity screens, and lower for family context and experimental structures.

   Of 5328 targets: AlphaFold models for 5107, disease link for 5189, cross-reactivity screen for 5167, family context for 4113, experimental PDB structure for 2114. Local BLAST gave 148 722 similarity hits across 5167 targets. · Homology, paralog and cross-reactivity context; Topology, structure and domain processing; Figure legends

3. SuggestedOn CD45 (PTPRC), the lenient pLDDT setting trims the low-confidence N-terminal segment, and PAE-based strict splitting yields three domain-sized constructs, two matching annotated fibronectin type-III domains.

   Ectodomain 26-577; lenient construct 224-577; strict constructs 226-389, 390-481, 482-577. Strict constructs 2 and 3 align with UniProt FN3 domains 391-483 and 484-576. A cut needs mean inter-block PAE of at least 12 and at least 4 above intra-block PAE. · Construct recommendation workflow; Construct generation and annotation; Figure legends

4. ShownMouse and cynomolgus monkey equivalents of each human design region are mapped for most targets, with identity and coverage reported for each construct.

   HCOP-based ortholog sequences resolved for mouse in 5004 of 5328 targets and for cynomolgus monkey in 4593; human design regions projected by global alignment. Coverage counts only; mapping accuracy is not benchmarked. · Homology, paralog and cross-reactivity context; Ortholog, paralog and cross-reactivity analysis

5. SuggestedSwapping the predictor changes the CD45 constructs under the same default thresholds, so pLDDT and PAE cutoffs would need recalibration for any new structure predictor.

   ESMFold2 gave mean pLDDT 66.7 versus 34.4 for AlphaFold in the UniProt-disordered N-terminal region, so the lenient setting kept it. Lower PAE between the two C-terminal regions stopped the strict setting from separating them. One protein, one prediction per method. · Discussion; CD45 comparison with ESMFold2; Figure legends