Fewer Rungs on the Ladder

Automating the analyst's work doesn't turn analysts into principal scientists

A few days into my AI Dry July, I finally read this new perspective paper arguing that AI will not replace bioinformaticians, but will move them up to higher value work: experimental design, interpretation, governance, deciding what to do next.

Wen Bin Goh, W., Polster, A., Wong, L. et al. Rethinking bioinformatics expertise in the era of artificial intelligence. npj Digit. Med. 9, 398 (2026). DOI: 10.1038/s41746-026-02777-1. https://www.nature.com/articles/s41746-026-02777-1.

My friend Tommy Tang, director of bioinformatics at AstraZeneca, wrote positively on his newsletter about the paper and what it means for bioinformaticians. Go read his post if you want a contrasting perspective to what you’re about to read here.

It’s a fine paper, and I agree with much of it. I think the paper misses something though. “The profession survives” and “your job survives” are kind of different things. AI can move analysts up the value chain but the upper tier was always small, and I don’t think AI is necessarily going to expand it.

When I ran a bioinformatics core here at UVA many years ago, and later when I led large teams across several companies in the national security sector, the org chart had two broad tiers, and the compensation surveys (Radford and the like) drew a similar line. It’s been a minute, so I can’t remember the exact terminology, but we had two named labor categories. There were bioinformatics analysts, often recent grads or master’s degree holders, writing and debugging code, running analyses, reporting up. And there were bioinformatics scientists, usually PhDs, who spent more time on design, interpretation, and client communication, and far less time in the terminal.

The two overlapped heavily. An analyst is a scientist, and a scientist still does analysis. And there was no caste system on my team. In fact I only remembered who was classified as analyst and who as scientist when we put proposals and cost volumes together for federal agencies.

But the scientist band paid substantially more, partly because there were fewer of those roles. One senior scientist could direct several analysts.

This paper’s argument seems to assume that when AI absorbs the analyst work, those analysts move up into the scientist tier and become the experts directing the AI. That tier was always smaller. The pyramid does not widen at the top because you automated the bottom. If AI now writes the code, debugs it, and runs the first-pass analysis that 10 analysts used to handle, you don’t suddenly need 10 more principal scientists. You need about the same number of senior people you already had, each overseeing more AIs and fewer juniors.

I think the paper skips another problem too. The analyst job was where people learned to become scientists. You misspecified enough DESeq2 design formulae, and after several discussions with me or other senior scientists on the team, you’d develop a feel for when an output is wrong or some contrast doesn’t make sense. Automate the apprenticeship¹ and you cut off the path that produced the experts the paper calls irreplaceable.

Will demand expand enough to take up the slack? That I don’t know. Cheaper bioinformatics could mean more experiments get the full treatment and more projects become feasible. Even then, the composition shifts toward the senior tier and the entry rungs thin out.

The field being essential and the field employing as many people are different claims. “AI cannot judge biological meaning, so bioinformaticians remain necessary” tells you the profession survives. It says nothing about how many it employs, or who they are.

⁂

Now that I’m back in the front of a classroom, my job is clearer than it was years ago. The apprenticeship that used to happen on the job, the small corrections that taught an analyst when a result smelled wrong or whatever, these are the things industry isn’t going to pay for. If that kind of apprenticeship happens anywhere now, it’s got to happen here, deliberately, before a student enters the workforce.

So that changes what I teach. I used to teach programming to biologists in my Biological Data Science with R course (bdsr.stephenturner.us). This fall I’m flipping that around and I’ll be teaching genomics fundamentals to data scientists. I’ll be spending less time (or possibly none at all?) on the syntax that the models already write, and much more time on the things it can’t judge for itself, like whether a batch effect is passing for biology or whether a contrast even makes sense or what about the study population is causing that weird off-diagonal Q-Q plot of p-values from a GWAS.²

The students who learn that are (I hope!) the ones who end up in the small senior tier, the experts behind the AI rather than the people it replaced. There as few of those senior-level seats as there always were, and AI isn’t expanding them. What I can try to do now is make sure the people I send out are the ones who get them.

1. I’m back in higher ed now, and training and apprenticeship is what we do here. But in industry it absolutely is not. When I was a consultant we were competing with other firms on both expertise and price, and if I could bring the same level of expertise and undercut another firm who took pride in training young scientists but cost 2.5x more, there was no question on the path we would take. Because there was no question who our clients would select to award.

2. I’d absolutely love to take my data science students on a “field trip” to a real bio lab, have them open up a frosty -80C freezer to find hundreds of unlabeled tubes falling out of boxes with handwriting they can’t read (guilty). You know, the kind of stuff that happens in the real world that ends up in the count matrix that shows up in your QC, things that most data scientists have never considered. If you’re reading this at UVA and you’d be up for a visit this fall, you know how to reach me.