Labs are burning grant money on AI predictions the model wasn’t sure about. DEGU, a new genomic AI from Cold Spring Harbor Laboratory, just solved this—by compressing 10 models into one that’s 10 times smaller while explaining its confidence. The catch? Only institutions with the infrastructure to deploy it can access the tool that would save everyone else the most money.
This matters now because precision medicine has stalled waiting for AI that clinicians can actually trust. When a model says “this variant might cause disease,” researchers run expensive wet-lab validation. If the AI wasn’t confident, that’s wasted resources. DEGU tells you when to trust the prediction.
DEGU does what 10 models couldn’t—with one-tenth the computing power
The technical breakthrough is counterintuitive: fewer models beat brute-force consensus. DEGU-trained neural networks inherit the predictive power of entire ensembles—those collections of 10+ models that vote on outcomes—but run as a single, dramatically more efficient system. Peter Koo’s lab at Cold Spring Harbor, working with former postdoc Jessica Zhou and graduate student Kaeli Rizzo, proved one distilled model can match ensemble accuracy across genomic tasks like STARR-seq and ATAC-seq.
The efficiency gains are real. DNNs trained using DEGU consume significantly less computational overhead than standard deep ensemble methods—the kind that require running predictions through multiple models and averaging results. For labs processing thousands of variant predictions, this translates to lower cloud computing bills and faster turnaround times.
But the real innovation isn’t speed. It’s uncertainty quantification. DEGU doesn’t just predict—it shows its work. When the model flags a genetic variant as potentially disease-causing, it assigns a confidence score. Low confidence? Don’t waste lab resources validating it. High confidence? Prioritize for experimental follow-up. This is the breakthrough that matters most for precision medicine applications, where AI predictions guide treatment decisions worth thousands of dollars per patient.
The confidence gap costs real money—and nobody’s tracking it
Here’s what the research team won’t tell you: Cold Spring Harbor Laboratory hasn’t released cost-per-analysis figures. No data on how much a single DEGU prediction costs versus traditional methods. No published false positive rates. No case studies documenting wasted lab resources from overconfident predictions.
We know genome sequencing dropped to roughly $200 per sample by 2025. We don’t know what analyzing those sequences with DEGU costs—or whether small labs can afford it.
And DEGU isn’t plug-and-play. Like other enterprise AI tools requiring integration with existing systems, DEGU needs to connect with genomic databases and clinical workflows before hospitals can deploy it. That requires IT infrastructure, bioinformatics expertise, and institutional buy-in—resources concentrated at well-funded research centers.
The accessibility problem no one’s talking about
DEGU is “working to improve efficiency and make it more accessible worldwide,” according to Cold Spring Harbor Laboratory. That phrasing does work: it confirms the tool isn’t accessible now. To whom? The lab won’t specify. At what cost? No public data exists. What infrastructure requirements? Unaddressed.
This creates a two-tier system. Well-funded institutions—the ones that could afford manual validation anyway—get computational efficiency and explainable predictions. Small labs and rare-disease researchers, who need trustworthy AI most because they can’t afford to chase false leads, are priced out.
Unlike consumer-facing healthcare AI tools, DEGU targets research institutions—a market where infrastructure gaps determine who benefits from algorithmic advances. The math is elegant. The distribution isn’t.
AI can now predict genetic disease and explain its reasoning. But the labs that need explainable AI most can’t access the tool, while well-funded institutions that could validate predictions manually get the efficiency gains. DEGU solves the trust problem. Who solves the access problem?
