Is retrieval all you need? Assessment and emergence of novelty in protein structure generation 1. The paper argues that “full-chain novelty” (no close whole-chain hit in a structure database) is not sufficient evidence for new fold invention, because multi-domain proteins are modular and can look novel as whole chains while reusing known domains. 2. It introduces Domain Retrieval Rate (DRR): decompose each generated backbone into domains (Merizo), then retrieve each domain against CATH S40; a generated backbone counts as “retrieved” if any domain has a match above a threshold. This explicitly tests novelty at the domain scale rather than the chain scale. 3. Across 8 backbone generators (diffusion + flow-matching families, plus two binder-oriented systems run in target-free ablations), DRR is high under an aligned-length TM convention: at TM≥0.5, DRR ranges ~80.2% to 98.2%, meaning most generated chains contain at least one locally alignable known domain-sized unit. 4. In the same unified evaluation, conventional full-chain retrieval (query-normalized TM, qTM≥0.5) is much lower for most unconditional generators (0.0% to 9.6%), while DRR stays high—showing a strong protocol-dependent gap between “domain reuse” and “whole-chain novelty.” 5. The paper diagnoses a key reason: query-normalized TM has a length ceiling when searching long chains against a domain database (CATH S40 median ~132 aa). For long queries, qTM≥0.5 can be unattainable regardless of structural content if reference entries are too short. 6. To calibrate what “retrieval alone” can produce, it proposes RetFold, a zero-training baseline that retrieves CATH domains (using an ensemble of ESM-2 embeddings, Foldseek 3Di, and ProstT5), assembles 3 fragments (core + two extensions), then refines junctions with geometry-based helix-linker optimization. It runs on CPU at ~1.98 s per backbone (reported ~110× faster than their measured RFdiffusion run). 7. RetFold is recombination by construction, yet under the conventional full-chain qTM≥0.5 protocol it is retrieved only 20.0%, while its DRR is 96.0%. This demonstrates that low full-chain retrievability can occur even when the chain is explicitly built from known domains. 8. The paper then calibrates three TM-score conventions using (i) RetFold as a positive control and (ii) a leave-one-topology-out negative control on CATH S40. A key result: aligned-length TM (alnTM) has very high false positives for “fold novelty” at τ=0.5 (it retrieves 90.04% even after removing the query’s topology), so it mainly evidences local alignability rather than complete fold reuse. 9. A stricter reference-normalized criterion (target-normalized TM, tTM) becomes more specific only at higher thresholds: at tTM≥0.7 the leave-one-topology-out false positive rate drops to 5.38%, while RetFold remains retrieved at 100%. Under this calibrated view, unconditional generators largely concentrate in the ambiguous 0.5–0.7 range and drop sharply at 0.7. 10. Beyond retrieval metrics, RetFold highlights a separate designability issue: AlphaFold3 confidence drops at assembled junctions (inserted helix-linkers have much lower pLDDT than retrieved fragments; inter-fragment PAE is high). Whole-chain average pLDDT can mask these localized failures, so novelty and designability should be reported separately with matched baselines. 📜Paper: arxiv.org/abs/2608.10598 #ProteinDesign #ComputationalBiology #ProteinStructure #DiffusionModels #FlowMatching #CATH #Foldseek #AlphaFold3 #ProteinBackboneGeneration #Benchmarking