The AI That Can Diagnose Genetic Birth Defects in Hours Instead of Months
Every year, millions of parents-to-be stare at ultrasound images, hoping for reassurance and bracing for uncertainty. For about 6% of all pregnancies worldwide, the scan reveals something unexpected: a structural anomaly in the developing fetus. A heart that formed differently. A spine that didn’t close. In the moments that follow, a grueling diagnostic odyssey begins — one that, even with the best genetic sequencing technology, can take months or years and often ends without answers. A new artificial intelligence system from Zhejiang University, described in a preprint posted this week, just made that journey dramatically shorter. Its name is DeepBD, and it might be the closest thing yet to a diagnostic safety net for the most vulnerable patients who will never speak for themselves.
Background & Context
Birth defects affect roughly one in every sixteen babies born, accounting for a disproportionate share of infant mortality and lifelong disability. When a structural anomaly is spotted on prenatal ultrasound — an enlarged heart chamber, a cleft in the neural tube, a missing kidney — clinicians face an agonizing question: what caused it? For a substantial subset of cases, the answer lies in the baby’s DNA. A single letter change in a crucial gene can derail the intricate choreography of embryonic development.
Modern exome and genome sequencing can read that DNA with breathtaking precision, identifying thousands of genetic variants in a single patient. But here’s the problem: finding the causal variant among those thousands is like searching for one misspelled word in a library of a hundred thousand books. The bottleneck has shifted from detection to interpretation. A clinician might sequence a fetus with a congenital heart defect, get back a list of candidate variants, and have no clear way to determine which one is responsible. The phenotype — what the baby actually looks like on ultrasound — is often incomplete because many developmental features haven’t emerged yet. Evidence is scattered across population databases, variant-effect predictors, gene-disease registries, and protein-structure models. Someone has to connect the dots.
What the Researchers Did
DeepBD, developed by computer scientists at Zhejiang University led by Haishuai Wang, is built around a deceptively simple insight: the task of diagnosing a genetic birth defect contains subtasks of fundamentally different character, and each should be assigned to the tool best suited for it.
The system has four layers. The first — LLM-assisted case structuring — takes raw clinical input (ultrasound reports, free-text descriptions, and sequencing data) and converts it into a machine-readable workspace. A language model maps phrases like “increased nuchal translucency” into standardized Human Phenotype Ontology terms, preserving the original text alongside so nothing is lost in translation.
The second layer is the heart of the system: a pretrained evidence engine that learns, from more than 18,000 real fetal and infant cases, how to weigh different kinds of evidence when ranking candidate variants. For each variant, the engine builds three parallel evidence streams. The first captures explicit clinical rules — population frequency, whether ClinVar lists it as pathogenic, whether the gene is known to cause disease. The second captures the variant’s intrinsic properties — what the surrounding DNA sequence looks like, what AlphaMissense and SpliceAI predict about its impact. The third builds a biological context graph linking the candidate gene to the patient’s specific phenotypes through cellular pathways, anatomical systems, and tissue expression data.
The third layer consists of specialist modules that can be called in when a candidate needs deeper investigation. If a variant looks borderline, DeepBD can query AlphaFold for protein-structure predictions or run Rosetta-based modeling to estimate how a mutation might physically destabilize a protein. These tools are computationally expensive, so DeepBD uses them sparingly, only for candidates that have already cleared the evidence-engine bar.
The final layer is a grounded diagnostic agent. This isn’t a free-form chatbot — it’s constrained. It receives the ranked candidate list with evidence traces and provenance records, and its job is to synthesize that information into a reviewable diagnostic hypothesis. It turns the evidence engine’s output into something a clinical geneticist can inspect, challenge, and act on.
What They Found
The results are striking. On a held-out benchmark of 549 solved cases — real fetuses and infants whose causal variants had been clinically confirmed — DeepBD placed the correct variant at rank one in 65.8% of cases. By rank three, that number jumped to 88.2%. By rank five, 91.2%. By rank ten, 92.9%.
To put those numbers in context, Exomiser — the current standard-bearer — achieved a rank-one recall of just 49%. The agentic system DeepRare hit 48.9%. Even the best general-purpose large language models, given the same candidate variants and asked to rerank them, topped out at 44.4%. DeepBD didn’t just outperform these systems; it widened the gap with every additional rank position.
The gains held across variant types: missense mutations, frameshifts that throw entire proteins out of alignment, splice-site disruptions, and truncating mutations. The system proved particularly valuable for the hardest cases — variants that lacked clear database signals and required synthesizing weaker evidence from phenotype match, pathway context, and protein-structure modeling.
Ablation experiments confirmed that no single evidence type carries the system alone. Removing rule evidence collapsed rank-one recall from 65.8% to 47.4%. Removing the graph-derived biological context dropped it to 49.4%. Removing structure-informed refinement dropped it to 52.1%. DeepBD’s advantage came from integrating all of them together. Perhaps most importantly, 117 cases at rank one were uniquely prioritized by DeepBD — cases that would have been missed by every existing system. These are the families who sequence their child’s genome and get back a report that says “no variants of clinical significance found.”
Why It Matters
A molecular diagnosis can change everything for a family facing a birth defect. It informs whether the condition will recur in future pregnancies, whether targeted treatments or clinical trials exist, and whether other family members should be tested. Without a diagnosis, families are left in limbo — repeating the same tests, seeing the same specialists, carrying the same uncertainty for years.
When a system can place the causal variant in the top three candidates for nearly 9 out of 10 cases, a clinical geneticist can focus their expertise on evaluating those three candidates rather than sifting through hundreds. That doesn’t just save time — it saves hope. And because DeepBD preserves evidence provenance — every ranking comes with traceable trails showing exactly which database contributed which score — a clinician can inspect the evidence, challenge it, and update it as new information becomes available. This is not a black box; it’s a structured reasoning system.
How It Could Change Our Lives
Imagine a 20-week prenatal ultrasound reveals a congenital diaphragmatic hernia — a hole in the diaphragm allowing abdominal organs to compress the developing lungs. Today, the parents might wait weeks for sequencing results, then months for clinical interpretation, and in many cases never receive a molecular diagnosis at all. With DeepBD integrated into clinical workflow, that timeline could shrink to days. The ultrasound findings are entered, sequencing data runs through the evidence engine, and within hours the clinical team has a ranked, evidence-supported candidate list.
This isn’t hypothetical. DeepBD was developed on a real in-hospital cohort of 18,622 fetal and infant cases — the kinds of cases that walk through the doors of a major medical center every day. It works with standard sequencing outputs, clinical notes, and imaging reports. No $10 million supercomputer required.
The Bigger Picture
DeepBD embodies a philosophy that is reshaping AI in medicine. Early systems tried to replace doctors; that approach failed because medicine is too messy and high-stakes for black boxes. The newer generation amplifies clinical judgment instead of attempting to replace it. DeepBD does what machines are good at — sifting millions of data points, weighing evidence systematically, maintaining provenance — and leaves what humans are good at — evaluating uncertainty, communicating with families — to the experts. This evidence-allocation principle is likely to become a template for AI across many areas of medicine.
Limitations & What’s Next
DeepBD needs external validation across different hospitals, sequencing pipelines, and ancestry groups before it can be deployed clinically. The benchmark evaluates retrospective cases — the next essential step is prospective validation measuring whether the system actually improves diagnostic yield and turnaround time in real time. The diagnostic agent is intentionally lightweight in this version, and its output has not been separately evaluated by domain experts. Still, the direction is unmistakable. DeepBD shows that the path forward for AI in genomic medicine is not more autonomous agents, but better evidence architectures — systems that know what they know, show their work, and put the final decision where it belongs: in the hands of the clinician and the family.
📄 Source: DeepBD: A Grounded Agentic Workflow for Variant Prioritization and Diagnosis of Genetic Birth Defects — Li et al., Zhejiang University, June 2026.