Microsoft designs AI assistant to diagnose rare genetic diseases

Rare diseases affect up to half a billion people worldwide, and finding a definitive diagnosis often turns into an agonizing multi-year odyssey.

While whole genome sequencing offers the best shot at answers, doctors and analysts must sift through more than 1 million genetic variants using bioinformatics tools like seqr. The process takes anywhere from 3 to 12 weeks, yet fewer than 50% of initial cases actually yield a diagnosis.

When a case goes unsolved, it gets shelved. New biomedical discoveries emerge every single day that could potentially solve those cold cases, but genetic analysts simply do not have the bandwidth to constantly re-examine backlogged files.

Now, Microsoft Research—collaborating with Drexel University and the Broad Institute—has published a new study in the ACM Transactions on Interactive Intelligent Systems journal, detailing how generative AI can be designed to break this bottleneck.

The real reason healthcare AI fails

The researchers highlighted a striking statistic right away: roughly 87% of AI systems never reach actual deployment simply because they solve the wrong problems.

Instead of building another flashy model in isolation, the team conducted in-depth interviews with 17 genetics professionals to map out their exact workflow hurdles. They pinned down three major pain points.

First is sheer information overload. Analysts spend countless hours synthesizing variant and gene data across scattered databases and literature, making the work exhausting and prone to human oversight.

Second is clunky collaboration. Sharing findings across multidisciplinary teams still relies on outdated, siloed methods that slow down decision-making.

Third is the reanalysis dilemma. Analysts know new papers hold answers for shelved patients, but there has been no systematic way to identify which dormant cases stand to benefit from which newly published findings.

An AI research partner, not an automated doctor

Working directly with genetic analysts through co-design sessions, the team built a prototype assistant zeroed in on two core tasks.

The first is automated reanalysis flagging. The assistant tracks newly published scientific papers and proactively alerts analysts when fresh evidence matches an unsolved case sitting in their backlog.

The second is biomedical literature synthesis. The AI pulls relevant insights about specific genes and variants from massive piles of scientific papers, serving them up in an accessible, digestible format so analysts do not have to read dozens of PDFs from scratch.

Crucially, the prototype includes collaborative sensemaking features. Because geneticists cautioned against blind trust in AI outputs, the interface lets specialists collaboratively edit, annotate, and verify AI-generated summaries, making corrections visible across the entire team.

Why it matters

We think this project stands out because Microsoft resisted the urge to build an autonomous diagnostic engine, choosing instead to tackle the unglamorous, time-sucking administrative friction of genomic medicine.

For genetic counselors and analysts, this is an immediate quality-of-life upgrade. If an AI assistant can quietly cross-reference cold cases against PubMed and flag a forgotten patient the moment a new variant paper drops, that directly shortens diagnostic odysseys that currently drag on for years.

Across the AI industry, this approach serves as a much-needed playbook for overcoming the infamous 87% failure rate. Meeting domain experts inside their existing software environments—like seqr—beats trying to overhaul clinical workflows every single time.

The catch, of course, is precision. Rare disease genomics has zero tolerance for hallucinations. If an assistant surfaces too many false leads, alert fatigue will cause doctors to ignore it entirely. The real test will be whether Microsoft can tune the system's filtering tightly enough to deliver high-confidence discoveries without adding more noise to already strained desks.