DANAE: Evidence-Guided Scientific Discovery Across Biological Scales

AI-assisted scientific discovery requires connecting interpretable hypotheses to candidate generation while distinguishing computational evidence from validated biological function. We introduce DANAE, a modular cross-scale architecture that links ecological modeling, genomic generation, orthogonal evaluation, and evidence-based decisions through explicit evaluation contracts and traceable records. Building on symbolic ecological discovery in Chilean copper tailings and metal-associated genomic candidate generation, DANAE extends these capabilities with broader environmental data, controlled comparisons, and an executed integration workflow. Evaluation on Tara Oceans covers 27 annotation-selected metal-associated ortholog groups, 180 DNA samples from 86 stations, and 129 paired DNA–RNA observations from 68 stations. Under station-held-out evaluation, environmental prediction of DNA reduces normalized mean absolute error by 25.5% relative to a training-mean baseline, while combining DNA and environmental inputs improves RNA prediction by 3.6% over environment alone. Operator and evidence-signal ablations characterize component contributions and the coverage cost of computational prioritization. A five-seed, 900-output experiment evaluates genomic conditioning under matched sampling budgets without establishing a clear conditioning advantage. A separate integration run links 15 ecological contexts to 90 newly generated candidates, of which 72 pass a coding proxy and 15 pass both coding and general-homology checks. In a controlled fault-injection benchmark, explicit decision contracts prevent unsupported promotion across 1,800 perturbed cases while retaining all 65 eligible clean records. These results demonstrate an evaluated workflow connecting ecological hypotheses, genomic candidates, and reproducible decisions, with portability from tailings to marine data through refitting. Computational priority remains distinct from experimentally validated function.