Attma

Attma is a research agent that turns a biology question into a reproducible, checked workflow. A researcher types or dictates a question and attaches data (a VCF, an .h5ad, a counts table, a PDF report). One agent on Meta's Muse models plans the work, runs each step with real tools (Human Phenotype Ontology ranking of about 12,900 diseases, Ensembl VEP with gnomAD, ClinVar, AlphaMissense and CADD, UniProt and AlphaFold, PubMed, g:Profiler, STRING, and Python in an isolated per-chat sandbox), shows every step with its source, and answers with a citation for each claim. The problem it solves is trust. In rare-disease and genomics research, an AI summary is only useful if a scientist can check it. Attma's trust gate makes the model propose and deterministic code decide: every number in a report is re-read from the tool output it cites, a second Muse Spark pass checks the claims, a biosecurity reviewer screens the request, every model-written script and the final report, and nothing is final until a person approves it. Reloading a chat shows the same workflow. On 200 solved, published cases (GA4GH phenopacket-store), with the HPO annotations drawn from those same papers removed to prevent leakage, Attma ranks the causal variant first among 20 candidates in 50.5% of cases and in the top 5 in 63%; with variant evidence alone, top-1 is 0%. Method, tuning/test split and full results are in the repository. Attma is research software: it ranks hypotheses for qualified people and never gives a diagnosis.