A few years ago, building a reproducible pipeline from scratch — from a raw FASTQ file through alignment, from differential expression to network analysis — would take days. Today I can draft the same skeleton in minutes. This does not mean my work has shrunk; on the contrary, it means the most tedious part of it has disappeared and only the genuinely hard part remains. As a computational biologist working where multiomic approaches meet protein sequence–structure analysis, my career mostly revolves around one question: "Is the signal I'm seeing real, or is it a trick played on me by the data, the protocol, or the biological context?" AI has of course not yet resolved the factors that make me ask this question — but it has substantially increased the time I can spend asking it. Thanks to AI, I now spend my days not writing code, but framing the right question so that I write the right code.
The hidden trap of speed: every unspecified parameter
I think the most insidious side of this speed lies in a trap it brings with it. For every parameter you forget to specify, the model reaches for the default value and threshold most commonly used in the literature — and this far-from-negligible detail, when acted on unawares, can send your analyses somewhere else entirely. So I now set every parameter individually, according to the needs of the data in front of me and the hypothesis I'm testing; I choose a threshold, a normalization strategy, a clustering distance metric never because "this is how the literature uses it," but because "this dataset and this question require it."
The interesting part is that I even take AI's help with this very rigor. After completing the basic structure of an analysis, I ask it to tell me which parameter I forgot to specify — and it usually does find a point I genuinely missed. So the same tool can be both the source of the trap and a way to audit that trap. The difference, to me, comes from knowing well which role the tool is being used in. AI turns out to be an accelerator of what we ultimately intend to do — of our intent. When our purpose is clear and biologically well-grounded, it saves us days; when it remains ambiguous, the odds rise sharply that it hands us a wrong analysis produced fast and with a flawless appearance.
Layered data: disordered regions and batch effects
This ambiguity can become especially painful when working with layered data such as protein structure. Some protein regions, for instance, are intrinsically disordered by their very nature; they do not have a fixed, predictable three-dimensional structure. A structural modeling pipeline written without knowing this treats that region as if it were an evolutionarily conserved, regularly folded sequence, and produces a model that technically "works" but says nothing biologically. Unfortunately there is, for now, no way for AI to know this difference on its own; stating it explicitly and asking it to select the right approaches from the relevant models falls entirely to the domain expert, or to the person doing the analysis.
A similar dynamic holds for batch effects in sequencing data. Running a correction algorithm is easy; deciding whether correction is needed at all, or how aggressive it should be, is hard, and requires a deep understanding of the experimental design. Over-correct, and you can erase the very biological variation under study; under-correct, and it may turn out that the "discovery" was nothing more than the reflection of a technical variable. In the end, the person doing the analysis does not merely clean the data; they shape it according to what the hypothesis in hand is actually asking.
At system scale: the hidden confounders under networks
The need for direct, continuous human oversight of AI-run analyses is not confined to the single-gene or single-protein scale; it appears just as sharply in far larger, system-level analyses. In population-level constructs such as co-expression networks, for example, a module that at first glance looks independent and biologically meaningful may in fact be an indirect reflection of a different underlying biological state of the samples — say, which maturation or differentiation stage the cells are in. Left unaccounted for, this surfaces as a confounder that is easy to miss and that can completely change the result and its interpretation. AI may be fast and capable at producing such a statistical structure — but questioning what that structure biologically represents is, I believe, still entirely human work.
Indeed, when I use AI in my own work, I have to review every line of code I write, one by one; because even when I add warnings, the model sometimes makes assumptions on its own or fails to use the information I have at the right point. This can always happen, and I have to be aware of it. That is why, to me, this review step should not be an optional part of the workflow but an inseparable one.
When domain knowledge runs out: learning through agents
Where my domain knowledge falls short, I follow a different strategy. I create agents on the relevant topic, gather information from them, learn, verify it in the literature, evaluate it, and only then bring it into my work. This lets me greatly accelerate a research and learning process that would normally take weeks. But I want to stress again that the human factor — given where current models have arrived, at least in bioinformatics specifically — has to be there, always. As a result, I no longer write much code myself; instead, I audit hypotheses that have been encoded as code.
The bigger picture: auditing the data that trains tomorrow's models
I think one of the most valuable contributions of advances in AI may be that they make it easier to notice the gaps and errors in existing datasets — and in the models trained on those datasets. Our findings, which we are currently preparing for publication and which are at this stage an observation put forward by us, point to the fact that some important, widely used datasets in genomics require serious corrections. I present this as speculation for now; but what made reaching such a determination easier, I believe, is the speed of AI-assisted analysis — being able to interrogate the same data from far more angles in far less time.
The bigger picture I see here is this: AI is not only increasing our analysis speed, it is also directly helping us audit the quality of the data that future models will be built upon. Correcting a dataset flaw detected today becomes a self-feeding loop that lets tomorrow's models be trained better — and its effect grows with every passing day.
To sum up
If I have to gather my thoughts: in bioinformatics work, AI saves me days, even weeks; it accelerates my learning, and makes it easier to catch nuances that could otherwise be missed. But knowing why a region should be modeled differently, which biological state might be hiding beneath a pattern, and which parameter is truly meaningful for this dataset still remains a responsibility that is mine. What is left to us now is no longer trying to write a pipeline end to end, but making sure of the correctness and the reproducibility of the design.