Jonathan D. Grinstein, PhD, North American Editor of Inside Precision Medicine, hosts a new series called Behind the Breakthroughs that features the people shaping the future of medicine. With each episode, Jonathan gives listeners access to his guests’ motivational tales and visions for this emerging, game-changing field.
Matthew Rabinowitz switched from engineering and computational research to medicine after a breakthrough on the Human Genome Project. He realized that telecommunications, aerospace, and machine learning technologies could help him understand human biology. His shift in focus was influenced not only by scientific interest but also by personal loss, including the deaths of family members affected by genetic conditions. These experiences convinced him that current diagnostic methods were inadequate, especially for patients and families in critical situations.
After founding Natera, Rabinowitz and his team developed Panorama Prenatal Test, a noninvasive prenatal test. This technology uses DNA variant analysis, Bayesian statistical methods, and machine learning to detect genetic conditions in cell-free fetal DNA in maternal blood samples. It increased accuracy, accessibility, and reduced invasive procedures. Myome, his new project, uses whole genome sequencing to find rare diseases. Myome uses AI models to assess cancer and cardiovascular disease risks using genomic and clinical data to improve early detection and prevention.
In this episode, Rabinowitz discusses regulatory constraints, fragmented data systems, and difficulties translating complex genetic information into clinical decisions. In the long run, he wants to create blood test diagnostics that can predict health and allow proactive medical intervention.
Rabinowitz uses several technical engineering and computational concepts, including:
- Packet Switching: a method of dividing data into smaller units that are transmitted independently and reassembled at their destination
- Transformer Model: a type of artificial intelligence system that processes entire datasets simultaneously to identify relationships between elements, widely used in modern AI systems such as GPT
- Gradient Descent: an iterative method used in machine learning to minimize error by adjusting model parameters
This interview has been edited for length and clarity.
IPM: What originally drew you into applying engineering and machine learning to genetics and clinical diagnostics?
Rabinowitz: There was all this incredible work happening in the early 2000s around the Human Genome Project, along with applications of signal processing and machine learning, which is what I focused on during my electrical engineering training.
There were really three catalysts for me.
One was in 2003, when my sister gave birth to a child with Down syndrome at one of the top hospitals in the country, and they didn’t know until he was born. I spent six days flying around trying to help. They went through one procedure after another and after six days, the baby died from complications. It was absolutely horrific.
Second, I couldn’t believe that we had all these advanced technologies in our phones, laptops, and spacecraft, but they hadn’t made their way into clinical diagnostics. At that point, I felt I had to apply my background in signal processing and early machine learning to these problems.
The third reason was about 15 years ago, I lost a child due to a genetic condition, an absolutely devastating experience. After going through that, I felt there was a path I needed to follow.
The engineer in me took over. It felt like a problem I had to solve. It was like being struck twice: unrelated events, but the same kind of tragedy.
That’s when we used a pregnancy sample to apply for NIH funding to improve prenatal testing. We got that grant, then several others, and ultimately built Panorama, which has transformed pregnancy care globally.
From there, one thing led to another. Now, through Myome and companies like Natera, we’re working on projects that could save the U.S. healthcare system around $200 billion per year. It’s been a very personal mission.
IPM: What are you seeing today with whole genome analysis that feels fundamentally new or different?
Rabinowitz: We’re now diagnosing conditions with whole genome analysis that simply weren’t detectable before. Myome has largely led the charge.
When I look at these case studies today, I get the same feeling I had 20 years ago. How were we not able to see this before?
We’ve spent a lot of time extracting signal from noise so you don’t need multiple sequential tests. You can start with the whole genome and layer analyses. This includes SNPs, CNVs, difficult deletions, tandem repeats, mitochondrial DNA, and methylation.
One example: an eight-year-old with developmental delay, autism, and hypotonia had already undergone exome sequencing with no findings. We identified a subtle deletion about one kilobase involving a single exon too large for short-read breakpoints and too small for coverage changes. That finding completely changed the child’s life.
Another example: a man in his mid-20s with dystonia, convulsions, and vomiting had undergone standard neuromuscular panels. They missed a tandem repeat very difficult to detect with short-read or exome sequencing. We developed new statistical methods and identified the breakpoints, which changed his life.
More broadly, rare disease costs in the U.S. are about $1 trillion annually with ~47 physicians involved over a 4–7 year diagnostic journey and massive lost productivity. The fact that we can now catch these cases is remarkable.
On the pregnancy side the belief was that the issue was solvable, but the technologies were limited. People at the time used shotgun sequencing and looked at DNA quantity. We instead analyzed SNPs between individuals.
We built a massively multiplexed PCR system ~20,000 primers in a single reaction. The challenge is noise, cross reactions, and primer dimers. We developed a machine learning optimization so every primer is tuned relative to every other, standardizing thermodynamics across the system.
From there, we built a statistical framework integrating across trillions of hypotheses, crossover events, noise, and fetal fraction. When it converges, you see a clear maximum likelihood peak that tells you what’s happening. If not, you know something is wrong.
This allowed us to detect things others couldn’t: very high sensitivity for aneuploidy and structural variants like microdeletions.
We could detect triploidy and vanishing twins, determine zygosity, and even de novo mutations, which are more than five times as common as Down syndrome. It was a completely new approach combining passion, engineering, and statistics.
IPM: Why can’t we have one universal test that does everything?
Rabinowitz: The short answer is that there’s so much more we can extract from each sample, especially with AI.
These transformer models trained to predict the next word require learning enormous context. With gradient descent, backpropagation, and large datasets, the performance is extraordinary. We didn’t fully appreciate the significance early on but today the possibilities are enormous.
That said, you can’t have one universal test. First, sample context matters. In pregnancy, you’re analyzing fetal cell-free DNA very different from adult disease testing. Second, it’s not just blood. There are many analytes. Beyond DNA, you need methylation, RNA, proteins. Most diseases require a multi-analyte approach. Third, regulation. You need to validate each test rigorously. You can’t validate everything at once across the genome. We also have variants of unknown significance. If you look for everything, interpretation becomes a problem.
So you have to focus your inquiry and ensure results are validated and actionable. That said, from a single blood draw, we can already do an incredible amount.
IPM: How has cfDNA and noninvasive testing evolved with AI?
Rabinowitz: Around 2017–2018, Natera began applying deep learning to diagnostics. [It was] one of the first large-scale uses in genetic testing. We had used neural networks earlier (e.g., in HIV mutation analysis) but this was different.
We applied convolutional neural networks to detect microdeletions in low fractions of cell-free DNA. A key example is 22q11.2 deletion syndrome. This occurs in about one in every 1,500 to 2,000 pregnancies. It’s more common than many screened conditions. Early detection allows intervention at birth. We initially used classical statistics, but after generating millions of samples, we trained deep learning models. The AI learned noise patterns and edge cases better than we could model, like Kasparov versus Deep Blue.
In a study of ~20,000 patients, we saw 100% sensitivity for larger deletions and ~83% for smaller ones, with a specificity of 99.95%. That translated to a positive predictive value (PPV) of ~53%, compared to 3–5% clinicians are used to.
Despite this, adoption has been slow due to reimbursement and guidelines, which is frustrating, because many children still miss early diagnosis.
IPM: Looking forward, how is AI transforming broader healthcare and genomics?
Rabinowitz: Today, we’re combining whole genome sequencing with AI and clinical data to predict disease risk far more accurately than even five years ago.
Across ~30 major diseases we can now predict susceptibility at a transformative level. If applied broadly, for example to people over 45, we could save over $200 billion annually by catching diseases earlier. Many interventions are simple like diet and lifestyle. Even small improvements matter. Every 1% increase in sensitivity can mean ~$7 billion in savings.
We’re also predicting neoantigens for personalized cancer vaccines, training on real patient outcomes—something that wasn’t possible before. And we’re building foundational genomic models, like language models, that learn the structure of the genome itself.
So across diagnostics, treatment, and prevention, AI is fundamentally transforming the field.
IPM: You mentioned earlier that you underestimated neural networks. What changed your perspective?
Rabinowitz: Around 2005, we were applying machine learning to genetics. We weren’t wrong, but we underestimated neural networks. We worked on HIV drug resistance. predicting which mutations respond to which therapies.
We used lasso regression, support vector machines, [and] carefully constrained models. Neural networks didn’t perform as well, which is what we expected. Our mindset was to control complexity to avoid overfitting. What we didn’t anticipate was massive data, stochastic training, and compute power, which allowed neural networks to escape local minima and scale.
In 2010, I had a patent on training neural networks with memory. It lapsed because Stanford didn’t maintain it. That was right before Google Brain scaled these approaches. The lesson is to stay open-minded. Technology can open possibilities you don’t see coming.
IPM: How do you see the future of diagnostics evolving from a single blood draw?
Rabinowitz: We’re moving toward a world where a single blood draw can tell us an enormous amount. Historically, progress was slow: blood cell counting in the 1800s, automation in the mid-1900s, cell-free DNA in the 1990s. Since then, progress has been explosive. From one sample, we can identify incidental findings (e.g., rare diseases, pharmacogenomics, and predictive risk) across many conditions. We can also detect cancer noninvasively through circulating DNA.
The capabilities are remarkable, but the genome is complex—three billion bases, with interactions that require enormous data to model.
IPM: What challenges remain in making these technologies widely usable?
Rabinowitz: Two main challenges. First, data, standardizing and aggregating clinical data across institutions, is historically very fragmented. AI is helping, but more coordination is needed.
Second, education. As we generate more information, explaining it to doctors and patients becomes harder. What’s known, what’s uncertain, and what action to take. At Myome and Natera, we invest heavily in genetic counseling. But across the field, there’s a tendency to simplify by withholding information. That won’t scale. We need better ways to communicate complexity responsibly.
IPM: How important is diversity and multi-ethnic data in building accurate models?
Rabinowitz: It’s absolutely critical and still underserved. Many models were trained on homogeneous populations, limiting accuracy. We’ve focused on building multi-ethnic models using diverse datasets and functional genomics, but we still need more data from underrepresented populations.
For example, in cardiovascular disease, we built a multi-ethnic model using large datasets.
We were able to reclassify ~50% of patients in the intermediate-risk category, identifying who is truly high risk (>20%) versus low risk (<5%). That improved decision-making significantly with over 10% improvement in classification. When followed over time, outcomes matched predictions closely.
This has huge implications for individuals and for healthcare systems. By identifying risk earlier and intervening, often with simple lifestyle changes, we can reduce costs and improve outcomes at scale. That’s why building diverse, high-quality datasets is so important. It’s one of the most powerful ways to improve healthcare globally.
The post Matthew Rabinowitz: Engineering a New Era of Diagnosis appeared first on Inside Precision Medicine.
