In clinical practice and in genetics labs there's a bottleneck you don't see on TV: you interpret millions of genetic variations for a single patient and, many times, there's no clear diagnosis. Sound exhausting? It is — and Microsoft Research together with Drexel and the Broad Institute studied where AI can actually help. (microsoft.com)

Why whole-genome sequencing matters

Whole-genome sequencing is a powerful tool to find variants that cause rare diseases. But the process is not just running a machine: it means filtering and prioritizing more than a million variants, gathering evidence from papers, databases and annotations, and building a clinically useful narrative.

Typical turnaround is 3 to 12 weeks, and still less than 50% of cases get an initial diagnosis. That leaves many files unresolved and patients waiting for answers. (microsoft.com)

Less than 50% of initial cases receive a diagnosis; reanalysis can help, but it is costly in time. (microsoft.com)

The study: co-designing an AI assistant for geneticists

What did the researchers focus on? First, they interviewed 17 genetics professionals to understand their workflows, tools and pain points. Then they ran co-design sessions and built a prototype generative assistant that was tested and refined with those same experts.

The goal: identify concrete tasks where AI adds value without replacing human clinical judgement. (microsoft.com)

Tasks experts asked the AI to do

• Flag unresolved cases that should be re-analyzed when new scientific evidence appears. This would help reduce the backlog of unreviewed files.
• Gather and synthesize information about genes and variants from scientific literature and databases, presenting it in a way that’s usable for clinical analysis.

Both tasks aim to save time in the so-called "sensemaking": searching, filtering and synthesizing data to build diagnostic hypotheses. (microsoft.com)

Important design considerations (what they learned)

The study's findings are not only technical: they're human. Three design ideas emerge as essential:

• Distribute sensemaking: AI can create shareable artefacts that make teamwork easier, but those artefacts must show what the AI generated and what humans edited to keep transparency and trust. (microsoft.com)

• Support re-sensemaking: analyses aren't one-off events. AI can turn static data into dynamic artefacts that record prior reasoning and flag what changed, so whoever picks up a case later understands why a conclusion was reached. (microsoft.com)

• Combine multiple modalities: integrating text, images (for example, representations of gene structure), locations and other data types helps build a richer view of the case. Multimodal models can synthesize this evidence into clearer formats. (microsoft.com)

Risks and limits: not magic, well-designed assistance

The researchers warn of two things you should keep in mind: AI must solve real, clearly defined problems (many systems fail because they attack the wrong problem), and professionals must be able to verify and correct AI outputs. In medicine, responsibility and traceability are non-negotiable; that's why the prototype was designed around human verification and collaboration. (microsoft.com)

And what's next?

The study concludes that generative AI has real potential to increase diagnostic yield and reduce time to diagnosis, but more research is needed: real-time trials, task-based evaluations and controlled deployments with clinical staff to measure real workflow impact.

Microsoft Research and its collaborators invite the community to continue the conversation and review the full publication. (microsoft.com)

You can read the full academic paper in ACM here and a related preprint on bioRxiv here. (microsoft.com)