BGI Genomics Open-Sources OneGenome, Enabling AI to Trace a Path from DNA Variants to Rare Disease Clues

The challenge of rare diseases often lies not in whether DNA can be read, but in how to identify, among tens of thousands of variants, the one truly associated with a patient’s symptoms. OneGenome, jointly developed by the BGI Research Institute of Life Sciences and Zhejiang Lab, seeks to connect this time-consuming interpretation work into a single AI workflow: first understanding genetic sequences, then linking patient phenotypes, disease knowledge, and clinical literature to generate diagnostic clues for review by professionals.

At the core of OneGenome is Genos, a human genomic foundation model, augmented with the text comprehension and reasoning capabilities of a large language model. The former identifies the potential biological functions of DNA sequences and variants, while the latter organizes unstructured clinical descriptions, compares them with disease literature, and explains the relationships between candidate variants and symptoms. The system can also be extended to treatment and medication information, although such outputs are better viewed as guides to the literature rather than direct prescribing recommendations.

The Genos model and its software resources had previously been made public. Zhejiang Lab disclosed that the latest model was trained on a corpus built from 636 high-quality, telomere-to-telomere human genomes and can process sequences of up to approximately 1 million bases at single-base resolution. The significance of this long-sequence capability is that the model does not have to examine only a short stretch of DNA near a variant and may also be able to capture more distant regulatory relationships. However, associations learned by the model from sequences are still not equivalent to proven disease-causing mechanisms.

The development team said OneGenome outperformed several general-purpose language models and traditional genomic models in tests of rare disease diagnosis and medication guidance. However, currently available public information does not fully disclose the number of test cases, data-splitting methods, comparison benchmarks, or error rates across different populations, and no independently peer-reviewed clinical validation has been reported. Without these data, it is difficult to determine whether the model has memorized existing cases or whether it can handle patients in real-world outpatient settings who have incomplete phenotypes, variable testing quality, or multiple coexisting variants.

Open sourcing allows research teams to inspect the code, reproduce tests, and adapt the model using local data. It can also help rare disease centers with fewer resources gain access to analytical tools. In July 2026, OneGenome received an application case award in the International Telecommunication Union’s “Open-Source AI for Global Impact” category; BGI also showcased downstream agent tools for rare disease screening and precision oncology. However, technology awards demonstrate an application concept and public value; they cannot replace medical device review or evidence of diagnostic performance.

The next hurdle is bringing impressive benchmark results into clinical practice: conducting prospective comparisons across different hospitals, publicly reporting missed diagnoses, misclassifications, and population bias, and ensuring that literature cited by the model is traceable and that patients’ genetic data are properly protected. Even if the system can rapidly narrow the range of candidates, final interpretation must still incorporate family history, inheritance patterns, functional experiments, and review by clinical experts. Whether OneGenome can truly shorten the years-long “diagnostic journey” faced by patients will be determined by these real-world conditions.