AI Protein Design Needs Literature-Backed Triage to Bridge Generation-Testing Gap

AI-driven lab-in-the-loop protein and antibody design can generate up to 30,000 candidates per lead molecule in a single round, yet only a few hundred are ever synthesised and tested. A documented Genentech and Prescient Design study across four rounds and four clinical targets showed roughly 45 variants assayed per lead molecule against a generation ceiling of 30,000 — a gap of nearly three orders of magnitude. Current ranking systems rely solely on model-internal signals such as predicted structure quality, binding affinity, and model confidence, which answer fundamentally different questions and can inadvertently deprioritise well-supported candidates. The proposed solution is a 'literature-in-the-loop' triage layer that retrieves published structural, sequence, and affinity data to rerank candidates before wet-lab slots are allocated. This layer would not eliminate novel designs but would enrich the assay queue with evidence-backed candidates while reserving dedicated capacity for genuinely novel architectures.