Jun 25, 2026 12:53 PM
From Proteomic Discovery to the Clinic: A Conversation with Steve Williams from Alamar Biosciences
Date: June 2026
Participants: Steve Williams, MD, PhD (Chief Scientific Officer, Alamar Biosciences), Jason Amsbaugh, MBA (CEO, Samba Scientific), John Corliss (Co-Founder, Carolina Molecular)
Overview
For twenty-five years, the promise of precision medicine has been framed largely in genetic terms. But what about the diseases genetics can’t explain: the metabolic dysfunction, the cardiovascular risk, the neurodegeneration quietly accumulating in people whose genomes look unremarkable? We recently sat down with Steve Williams, Chief Scientific Officer at Alamar Biosciences, to discuss the “information gap” that genetics leaves behind, and why he has spent his career betting that proteins are where the answers reside. Steve’s vantage point is rare, spanning translational medicine at Pfizer, the discipline of machine-learning-driven proteomic discovery he built at SomaLogic, and his current work translating that science into the clinic. The conversation traces a single arc: how we move from extraordinary discovery to tests that everyday medicine can actually use.
Why Proteomics Fills the Information Gap
Jason Amsbaugh: Where do you like to start when you think about proteomics and its place in medicine?
Steve Williams: We’re all in the information business, whether we’re life science researchers, translational scientists, or biopharma. We’re all looking for the information that helps us make better decisions, whether in drug development, patient care, or understanding physiology. The question is simply: where is that information?
My honest view is that genetics has been a bit disappointing relative to the hype of twenty-five years ago, when people said every disease would be explained genetically. It hasn’t turned out that way. Most of the diseases that kill people aren’t explained by genetic predisposition. There’s some contribution, but more than half is something else, probably environmental. You have wonderful successes in cancer and in rare dominant variants, where you get unique signals from mutations. But here’s how I like to frame it. If I have a room of 100 people in front of me, maybe five or six have an actionable genetic variant, one might be pregnant, one or two might have cancer at that moment. That leaves more than 90 people for whom genetics has nothing to offer today, and many of them are carrying prediabetes, metabolic dysfunction, liver or kidney disease, even a predisposition to cancer that genetics doesn’t capture. That’s the information gap.
Jason Amsbaugh: And the multi-omics camp would say, measure everything at once.
Steve Williams: If you’re rich, sure. You can go searching for information in the microbiome, metabolome, lipidome, proteome, and transcriptome all at once. But if you had to take a bet on the single most likely place the information resides, I would bet on proteins. In fact, I have bet on proteins. We’re made of them, they sit downstream of both genetics and the environment, and they’re the targets of about 95% of all known drugs. So why not? You might ask why we didn’t start with proteins, and the answer is that we couldn’t measure them. There may be 20,000 proteins in circulation, and many times that with post-translational modifications. If proteins are like the internet, you need bandwidth and a search engine to interrogate them: how many you can measure at once, plus the high-powered computing and machine learning to make sense of it. Twenty-five years ago, neither existed. Over the last couple of decades, that’s what’s changed. Interrogating the proteome finally became possible.
Proteomic Discovery and the Missing "Off-Ramp"
Jason Amsbaugh: Of the proteome we can measure today, how many proteins actually have diagnostic utility?
Steve Williams: There are only about 217 FDA-approved protein measurements. Beyond that there are surely hundreds of laboratory-developed tests, but few make a wide impact on health. The reason is that the ability to measure tens, hundreds, or thousands of proteins at once is quite recent, and getting from “we have a technology that can measure these things” to something medically actionable is slow and difficult.
The godsend has been the large biobanks. If you go into a biobank where future outcomes are known and interrogate the proteome, with SomaScan, Olink, or some of the newer mass spec approaches that can now measure thousands of proteins, you can do two things. You can find individual proteins tied to individual phenotypes, which turns out to be relatively rare. Far more often, you find patterns. Proteins work in networks, and those networks relate to clinical outcomes. UK Biobank, ARIC, and EPIC have really enabled that kind of discovery.
The problem is what happens next. At SomaLogic, Larry Gold’s vision was always the “liquid health check,” delivering all those conditions at once on the big platform. SomaLogic took 17 of those multivariate tests all the way through to laboratory use. But it breaks the regulatory and reimbursement systems and the way physicians manage health. You can’t argue individual medical necessity for 17 tests spanning dementia to cardiovascular risk to how much alcohol someone drinks. The concierge physicians doing active prevention love it. But integrated health systems aren’t interested, partly because insurer turnover is two years or less, so there’s no payback window for a comprehensive check. So it’s stuck as a cash-pay product.
That’s the limitation of the big platforms. They’re wonderful at discovery, but there’s no off-ramp. If you want to play in today’s environment, where each prediction must be justified as individually necessary, it’s tough, because the platforms are too expensive and clunky to say “just take those six proteins and forget the other 10,994.” And to measure thousands of proteins, you accept some analytic compromises relative to those 217 well-characterized, FDA-approved ones.
The Multidimensional Sweet Spot for Clinical Proteomics
Jason Amsbaugh: How do you see that gap being filled?
Steve Williams: Medical practice today lives in unidimensional space. Roche, Siemens, Danaher: huge installed bases where your doctor checks a box for one measurement at a time. A clinician might look at several results and infer that your kidneys or your metabolic health are a problem, but that’s still reductionist, one-at-a-time inference. I don’t think that’s where the future of medicine is. Diseases are heterogeneous. Metabolic health, brain health, cancer surveillance aren’t single proteins or even single pathways. The future of diagnostics is inherently multidimensional. But those models don’t need to be thousands of proteins big. They’ll likely be modest and indication-specific.
That’s where a platform like Alamar comes in. I look at it as the off-ramp. It’s a multiplex big enough to catch and refine protein-network patterns, yet you can shrink those patterns down to small enough numbers to have a credible path to FDA approval. You can aggregate new single-plex discoveries, what I like to call “pathology sentinels,” at one end, and bring protein-network patterns from SomaScan, Olink, and mass spec at the other, and put both into that sweet spot in the middle, where the analytic performance is compatible with clinical use without being immediately squished down to a single-plex.
When I joined SomaLogic about eighteen years ago, the frontier was whether you could build protein-network models more informative than any combination of risk factors or any genetic risk score. It took a while, but the answer is yes. For most of the diseases that kill most people, those protein patterns are more powerful than the genetic ones, and you can shrink them down even when you needed large multiplex to discover them. Once we’d overcome that frontier, the exciting question became, “now what?” How do you get these into large-scale clinical trials and medical practice? That’s what drew me to the ARGOTM platform. It’s a distributable box that lives on a bench. Central labs can install one, but so can specialist neurology clinics. It’s genuinely easy. You load the machine and press the button, and under about 15 proteins you get the readout in the box, while for hundreds you get it by sequencing. You can even run dried blood spots to make it less invasive.
Designing Proteomic Tests for the User, Not the Scientist
Jason Amsbaugh: Are leading molecular pathologists ready to bring on protein expression profiling for clinical use, or does the market still need to develop?
Steve Williams: When you develop a clinical test, you have to start with the user: who are they, and what decision are they trying to reach? Underneath, you might be using a complex mathematical model built on dozens of proteins, but the output has to be something they can understand, often a binary yes or no. In the neuro world, that’s “this person would be read as amyloid-positive, or negative.” When we built tests for physicians on SomaScan, the outputs were things they could act on: does this person drink more than 14 units of alcohol a week? How likely are they to have a cardiovascular event in five years?
What we learned is that you have to flip the script. We began as academics asking “what information can we present?” Over time we realized you have to ask what the user wants and then build the model that delivers it. What goes on inside the box should be concealed. We once added the individual levels of 27 proteins to a report because users asked for it, and it caused enormous confusion. “Protein number 16 is outside the normal range, what does that mean?” And we’d have to say, we don’t know, because the machine learning built the model as a composite to predict a composite outcome. We took it back out. The closer you get to patient care, the more people want straightforward, decision-oriented outputs.
ADNI, Truth Standards, and Proteomic Validation
Jason Amsbaugh: You helped initiate ADNI. How did that come together?
Steve Williams: That came from my time at Pfizer heading the clinical technology group, when Pfizer had a major neurodegeneration program. Aricept was a symptom-only drug. People got an immediate boost, but the progression slope was unchanged, with the placebo and active curves staying parallel. The question was how you’d show you’d actually changed the slope of Alzheimer’s progression without running a cognitive study of tens of thousands of people over five or ten years. We thought the answer was probably imaging, so we pulled the world’s imaging experts together for a meeting in Georgetown to choose a technique: PET, MRI, CT, or something else.
By the end, the conclusion was that even within a single technique, the differences were as big as the differences between techniques. We couldn’t choose. And the reason we couldn’t was that there was no longitudinal, consistent cohort study where all of these had been applied head-to-head. That was a new question: how would you set up a study like that? Mike Weiner, one of the experts we’d invited, recognized that need. Beyond identifying the problem and providing some startup money at Pfizer, I didn’t do much more, but we helped identify the nature of the problem and a path to the solution. Coming back into the field nearly twenty years later, it’s remarkable how much ADNI helped. Those imaging truth standards let people ask, can I mimic that without imaging? That led to CSF, and then to plasma. The whole evolution, from no truth standard, to imaging, to CSF, to plasma, owes a great deal to the ADNI team.
Why Proteomic Models Stay Platform-Specific
Jason Amsbaugh: Are the multi-marker panels in your published papers platform-specific, or will they translate across platforms?
Steve Williams: At the moment they’re completely platform-specific, and people may not realize that. SomaScan’s output is in relative fluorescence units, not concentration-calibrated, but calibrated to predict a truth standard. So, the model you train and validate is completely platform-dependent. People assume that platforms claiming absolute quantitation solve this, that you could train a model on one platform and translate it anywhere because every feature is calibrated to concentration. It turns out that’s untrue, because none of the platforms calibrate concentrations using the same protein standards. Even within the phospho-tau world, comparing Fujirebio, Siemens, and us, all claiming absolute quantitation, you run the same samples and get different concentrations. Without universal calibrators for every protein feature, you can’t borrow a published model and run it on your own platform. There can be a shortcut. If someone has already identified the features, you may not need to retrain everything from scratch back out at UK Biobank, but you’ll still need to recalibrate each feature weight to your platform.
The Next Frontier: Democratized Proteomics
Jason Amsbaugh: Looking out five years, what’s the seminal achievement you’re working toward in clinical proteomics?
Steve Williams: I think the jury on multi-omics will come in. Within five years it’ll be obvious, and I believe proteomics will dominate, probably in combination with genetic markers. The two together will likely out-compete the others. Metabolomics and transcriptomics will have niche applications, but won’t dominate broadly.
The bigger shift is democratization. Today we have very large, expensive, finicky platforms that require blood to be shipped around the country, mostly by venous collection. I think we’ll see a much more distributed model, with proteomic setups in near-patient hospital labs, possibly using dried blood spots at scale. Automation and ease of use will enable that. You may see a whole new industry, an explosion in the number of people developing diagnostic models, and certainly a democratization of who can run multiplex tests and where they get done.
Conclusion: The Future of Clinical Proteomics
As Steve frames it, the story of modern proteomics is no longer about discovery, since that frontier has largely been crossed. The real work now is the “off-ramp,” carrying powerful, multidimensional protein signatures down from sprawling discovery platforms into tests that clinical practice can actually justify, run, and act on. That transition, from discovery-phase complexity to clinical-ready simplicity, is where the most meaningful innovation happens today.
At Carolina Molecular, we are proud to support cutting-edge technologies with partners like Alamar Biosciences, helping to translate high-level proteomic insights into standardized, actionable workflows. Whether the goal is closing the information gap that genetics leaves behind or bringing multiplex proteomics nearer to the patient, our mission remains consistent: getting the right treatment to the right patient at the right time. We invite you to join us as we continue to move these technologies forward.