My business partner Jonathan recently sat me down with a recorder and asked what I think about data. I told him it’s a great word. It’s got two A’s and every other person pronounces the first one differently.
He obviously was trying to get me to expound on what constitutes data and how it shapes our reality. I’ve spent three decades in genetics and bioinformatics staring at biological data, and here’s the short version: most everyone on the planet now has access to the same AI models, so the model is no longer the advantage. The advantage is the data you feed it and the person standing between the data and the machine.
It’s true in a genomics lab, and it’s just as true in a fifty-person company wondering whether AI can do anything useful. The lessons transfer better than you’d expect.
Garbage in, garbage out — but who decides what’s garbage?
The oldest rule in science is garbage in, garbage out, and AI has not repealed it. If you have bad data, you’re going to get bad results. That’s just how it runs.
The harder question is how you tell good data from bad among so many sources and so many opinions. My answer is the one nobody wants to hear, because it can’t be automated away. It’s domain knowledge. The person with the AI determines whether the source data is good or not. A domain expert doing quality control on the inputs understands whether an analysis is going to work before anyone touches a model.
I want to be fair to the models here — they’re astonishing when it comes to actual data analysis. But a model has no way to know that your sensor was miscalibrated last March, that one of your spreadsheets mixes two definitions of “revenue,” or that a public dataset was collected under conditions that don’t apply to your question. You know that, or you don’t. The quality of what comes out is set by the judgment of what goes in, and judgment is still a human specialty.
Your data speaks in dialects
In biology, context isn’t a nice-to-have. It’s everything. The biggest differences in the biological information you get are driven by tissue type and not necessarily disease. That isn’t an opinion. It’s the first thing you see in public atlas sources that measured gene activity across dozens of human tissues from hundreds of donors — where tissue identity dominates the variation and disease effects sit well below it. A muscle cell is a completely different creature from a nerve cell even though they contain exactly the same DNA. Tissue types produce different signals because they do different jobs in different environments.
The way I think about it, every tissue type is its own language. They all have access to the same words (~20,000 protein coding genes in humans) but each one uses a particular set of words, with different tones and different meanings. You can imagine a gene “saying” something in the liver can make a different statement than the same gene saying it in the brain. If you analyze biological data without respecting which dialect it’s speaking, you’ll produce confident nonsense.
Businesses have the same problem wearing different clothes. Your sales data, your support tickets, and your operations logs are all derived from the same company, but they speak different dialects of it. The word “customer” means something subtly different in each system, and a model fed all three as if they were one language will average away exactly the distinctions that mattered. Treat each data source as its own language first. Translate carefully. Then combine.
Signal isn’t volume, it’s who’s talking to whom
“Signal” is one of the most overused words in machine learning. In the world of bioinformatics we translate it into levels: a gene is more active or less active, a protein goes up or down in concentration. And that is precisely where I think people fail.
The failure looks like this. You imagine a person saying a word at a normal volume when they’re healthy and screaming the same word when they’re sick. Measuring the volume difference is useful, and it’s what most analyses stop at. But real signal goes beyond what any individual gene is saying. It’s who they’re talking to and the different sentences they create together. If a gene talks to certain partners in one condition and refuses to talk to those same partners in another, that change in the conversation is a huge event — and it’s invisible if all you measure is volume. A lot of what people call signal is just loudness. The relationships are the message.
Measuring levels is one comparison per gene. Measuring relationships is one per pair, and for twenty thousand genes that’s two hundred million conversations to audit. The problem is real. So is the payoff.
There’s a second half to this, and it’s my favorite thing about biology: genes have history. We have terabytes of accumulated observations, sitting in public databases, of how these genes have behaved across thousands of experiments. Why wouldn’t you use all of that to figure out how they’re going to talk to each other in the present or in the future? How do genes interact in someone with metastatic cancer versus someone with a benign tumor? The conversation patterns are in there, waiting, if you’re willing to go listen.
You don’t need perfect data. You need many good sources.
When COVID-19 first emerged, there was essentially no data on SARS-CoV-2. Everyone was waiting for the perfect dataset to appear. We didn’t wait. We went looking for a relevant stand-in — mouse models of SARS-CoV-1 infection, which the research community had been studying for years — and we pulled together what the public domain already had.
People are afraid of that public data, and I understand why. It was generated on older technology. There’s a whole divide between microarrays and RNA-seq: RNA-seq is more sensitive, microarrays arguably give a steadier baseline. But here’s the key that unlocked everything for us. Different platforms will never agree on the absolute level of a gene — but they agree remarkably well on direction. Up is up, and down is down, no matter how you measure it. That isn’t folklore. It’s the central finding of the MicroArray Quality Control project — a large FDA-led effort that ran identical samples across every major measurement platform to see what held up and what didn’t. Direction of change reproduced. Absolute intensity did not. You’d never trust a technology that couldn’t tell you whether something went up or down. So if you normalize on direction, you can align study after study on one scale and combine them.
That’s what we did: fifteen independent studies, seventy-four comparisons of infected versus uninfected mouse lung, aggregated with a “wisdom of crowds” approach so each study’s quirks got diluted by the others. Out of that came temporal meta-signatures of infection and a set of candidate biomarkers — genes like CXCL10 and BEX2, and a circulatory hormone called adrenomedullin. By the standard critique, this shouldn’t have worked. Mouse models of human inflammatory disease have a famously poor translation record, and the field has argued about it in print for a decade.
And here’s the part that still makes me proud: when human COVID-19 data finally became available, the same biomarkers we’d found in the mouse data turned up there too. Adrenomedullin was maximally elevated in the elderly patients who died. The crowd of imperfect mouse studies had told us the truth about a brand-new human disease.
We weren’t the only ones to land there. Adrenomedullin’s stable fragment, MR-proADM, has since been independently validated in the clinical literature as a severity and mortality marker in sepsis and in COVID-19 — a protein the clinical world reached from the opposite direction and found already waiting. If you’re curious, the full study is open access: Crowdsourcing temporal transcriptomic coronavirus host infection data.
I draw two lessons from that experience, and neither is really about coronaviruses.
The first lesson is one James Surowiecki put best in his 2004 book The Wisdom of Crowds: trying to find smart people won’t lead you astray, but trying to find the smartest person will. It’s not about finding the one best study. It’s about finding a lot of really good studies and combining them. Second: the data you already have, plus the data the world has already published, is usually enough to start. Companies stake their Series A on exclusive proprietary datasets, and sometimes that’s justified. But far more often, the honest answer is that you’re sitting on usable data right now and haven’t had anyone listen to it properly.
Trust the AI as far as you can confirm it
Do I trust AI in my own work? I trust it exactly as far as I can check its results, and not one step further.
In my day-to-day analysis — code, data wrangling, the work where I can check the answer myself — the AI is right the large majority of the time. That’s extraordinary, and I’ll take it. But if you have a situation where somebody’s life depends on the answer, you absolutely must have a human in there for the fraction where it doesn’t work. That’s not a hedge or a disclaimer. Human-in-the-loop is the design, not a limitation of the design.
Two practices follow from that. The first is to give the AI room. What we’ve found is that if you leave the analysis open-ended — rather than forcing it into one predetermined statistical test or pipeline — it will offer you different ways to look at the data, and that flexibility is where the value lives. The second is to treat it like what it is: the best research assistant I’ve ever had, at pennies on the dollar. Let it gather, let it synthesize, let it summarize fifty papers while you drink your tea. But when it tells you something surprising — something that would change a decision — go to the original source and confirm it yourself. Don’t depend on the AI to tell you exactly what a source says. Verify the things that matter.
I’ll add a personal note, because it still amazes me. Before AI came on board, I spent 80% of my time just getting data into a format I could analyze. Eighty percent. That bottleneck is simply gone now, and the time goes where it always should have gone: thinking about what the results mean.
What I tell the next generation
Young scientists ask me what they should do in a world where a model can draft their analysis in seconds. My answer surprises them: embrace AI — and specifically, embrace an agnostic AI. There’s no doubt in my mind that computers process information and build datasets better than I ever could. I’ve watched it happen, I’m okay with it, and fighting it is a waste of a career.
Then I tell them where to put their hours, because that’s the real question now. Not on the mechanics. Data analytics isn’t about staring at spreadsheets all day. It’s not about the numbers — it’s what the numbers mean. Learn the biology of the system you’re studying deeply enough that you can tell a plausible answer from a merely fluent one. Learn to ask a good question, because the model will answer exactly what you asked and not what you meant. And learn where your tools break: push the model, ask it the same thing five different ways, find the conditions under which it fails. That last one isn’t in any course catalog, and it’s the part that compounds. A generation that can run a pipeline is not scarce. A generation that knows which pipeline was the wrong one is.
What this means if you’re the one buying AI
You don’t need to care about bioinformatics. But if you’re evaluating an AI project for your business, the lessons from thirty years of listening to genes translate directly:
- Ask who is judging the data. If nobody with deep knowledge of your business is doing quality control on the inputs, you’re buying garbage-in with a confident interface.
- Ask whether your data’s dialects are being respected. Combining sources without translating their contexts produces averages, not insight.
- Ask what “signal” means in the proposal. If it’s just metrics going up and down, you’re missing the relationships, which is where the real story lives.
- Ask what happens when the AI is wrong. It will be. You want a human in the loop by design, and a habit of verifying the outputs that matter at the source.
Whether the system underneath your data is a lung or a ledger, the principles hold: respect context, combine many good sources, keep the human in charge of judgment, and never confuse the volume of a number with its meaning. AI is the best listener you’ll ever hire. But it can only hear what your data actually says, and if the data is lying, it will repeat the lie fluently, at scale, in a confident voice.
This post grew out of a recorded conversation with my business partner, Jonathan Rapisarda — the same conversation, in fact, where I learned that “data” having two A’s is not considered a complete answer. The recording was mine; the opinions are mine; the commas were supervised.