Will AI replace bioinformaticians? The expertise that gets more valuable, not less

The dominant narrative about AI and scientific expertise is one of replacement. Algorithms are consuming the tasks that once required years of training, and the professionals who built careers on those tasks are running out of time to adapt. In bioinformatics the argument has particular force. If a large language model can draft an RNA-seq pipeline, annotate variants and produce a publication-ready interpretation of a differential expression analysis, what exactly is left?

Nearly everything that matters, argue Goh, Polster, Wong and Cvijovic (2026) in a perspective in npj Digital Medicine. Their framing is the most useful one currently in print, and this post takes it as a starting point before adding what we have learned building the tools in question.

Their central observation is uncomfortable and worth quoting closely. Without fundamental expertise, they write, AI does not democratise science. It merely democratises the appearance of science. The failure mode is not obvious nonsense. It is analyses that look rigorous, outputs that sound authoritative, and conclusions that cannot be evaluated by the people who act on them. That is a different and considerably more dangerous thing than replacement.

It helps to be specific about which AI is meant, because the expert requirements differ sharply. Goh and colleagues distinguish three categories.

That third category is not hypothetical. Swanson and colleagues (2025) built the Virtual Lab, in which a human researcher sets a high-level goal and a coordinated team of language model agents specialising in immunology, computational biology and machine learning designs and runs a pipeline incorporating ESM, AlphaFold-Multimer and Rosetta. It produced 92 experimentally validated SARS-CoV-2 nanobody candidates. The expertise required to build, validate and critically interpret such a system extends well beyond writing prompts.

The perspective organises the change into three transitions. They are worth reading as a description of where the value moves, not as a warning.

Expand

The three transitions described by Goh and colleagues, with what AI absorbs and what it leaves behind. Full citation in Sources.

From tool user to AI custodian

The first shift is a change in who is responsible for the output. Moving from using AI to being personally accountable for whether its outputs are correct, reproducible and biologically meaningful is what separates progress from a proliferation of confident noise that nobody is equipped to evaluate.

This is not soft advice. Benchmarking studies cited in the perspective show performance varying by 20 to 40 percentage points on identical biological tasks depending purely on how the prompt was constructed. The same model, same data, different phrasing. That is a methodological variable, and treating it as one has consequences: prompts become versioned assets that get tested across inputs and documented, not incantations pasted from a colleague. Two caveats sharpen the point. Chain-of-thought prompting can produce step-by-step reasoning that reads coherently and still lands on the wrong conclusion. And a prompt library that works today carries no guarantee after the foundation model is retrained, which makes continuous monitoring for silent degradation part of the job rather than an optional extra.

From pipeline executor to pioneer

The second shift is toward the problems where AI is most powerful and where, without expert guidance, it is most likely to be wrong in ways that are hard to detect. The GWAS pipeline illustrates it cleanly. AI can generate and debug the scaffolding for quality control, population stratification and results formatting: work that is conceptually routine but technically time-consuming. What it cannot do is decide which QC thresholds are appropriate for this cohort and platform, recognise that a cluster of unexpected associations near a known locus may reflect linkage disequilibrium rather than independent signal, or judge whether population structure correction is adequate given the ancestry composition of the samples. Failure to account for population structure has historically produced false-positive associations that entered the literature and required correction. Automating the stratification step without expert verification amplifies that risk at scale.

There is a useful test buried in this. If AI frees up your time and that time returns to the same routine work in a different form, the transition has not happened. The recovered time is the metric.

The most consequential questions in biomedicine sit at the intersection of several data modalities: genomics and clinical records, proteomic profiles and imaging, molecular signatures and longitudinal patient trajectories. Each has its own scale, resolution, noise structure and biological logic. This is both the hardest integration problem in the field and precisely where generic AI tools are least adequate.

The difficulty is not computational. Algorithms for multi-omics integration and cross-modal learning exist and are increasingly powerful. The issue is interpretive. When a joint model of genomic and clinical data identifies disease subtypes, the first question is not whether the clusters are statistically robust. It is whether they are biologically real, whether they align with clinical outcomes in an actionable way, and whether the features driving the separation are meaningful signals or technical artefacts.

Compounding this is a vocabulary gap that is rarely acknowledged. A “feature” in a machine learning pipeline and a “biomarker” in a clinical context may refer to the same measurement, but they carry different assumptions about how it was obtained, what validates it, and what authorises its use in a decision. Systems that are technically sound within one frame turn out to be clinically uninterpretable within another.

This is where we want to add something to the argument rather than restate it, because it is the part we have built against directly.

Goh and colleagues are pointed about current explainable AI. Attention-based explanations, SHAP values and feature importance scores can be unstable across runs, sensitive to implementation choices, and in some cases post-hoc rationalisations of model behaviour rather than genuine explanations. Their conclusion is not that explainability is futile but that its inadequacy is exactly what makes expert biological judgement indispensable: these methods are starting points for expert investigation, not endpoints.

We agree, and we would add a distinction that the critique implies but does not draw out. The instability is a property of post-hoc explanation specifically: train an opaque model, then ask a second method to guess why it did what it did. A model whose architecture encodes known biology does not have that problem in the same form, because there is no second method guessing. The structure is the explanation.

Expand

Post-hoc explanation is a hypothesis about a model. Interpretable structure is a claim about biology, which is the kind of claim an expert can actually test.

What this looked like in practice

In work carried out and validated with the Danish Cancer Institute, we took a question that has both problems at once: multi-modal, and dependent on an explanation an expert could challenge. A share of gastric adenocarcinomas are homologous-recombination deficient despite wild-type BRCA1/2. What drives the defect, and can it be targeted?

The starting point was a biologically informed neural network in the P-NET style (Elmarakeby et al., 2021), built from Reactome and taking methylation, transcriptomics, mutations and proteomics as input. Every node maps to a real biological entity: a gene, a protein, a pathway. That is the structural advantage. The model is interpretable by design rather than a black box rationalised after the fact, so when it separates high- from low-HRD cases, the features doing the separating are biological objects a domain expert can accept or reject on biological grounds.

It surfaced EZH2 dysfunction as the likely driver. The methodological point is what makes this relevant here. Test thousands of genes one at a time and apply Bonferroni correction and EZH2 drops below the line with the noise: the association is real but not strong enough to clear that bar alone. It is also invisible to transcriptomics by itself. The signal only holds up once methylation, mutations and proteomics are integrated alongside expression in a model that reasons over pathway structure rather than one gene at a time. Routine analysis would have buried a true vulnerability. Gastric cancer cell line HRD scores then correlated with sensitivity to tazemetostat, an FDA-approved EZH2 inhibitor, which turns an explanation into a testable therapeutic hypothesis.

A second problem follows immediately, and it is the one that usually goes unmentioned: an explainable network is still unreadable. Thousands of nodes and edges make it nearly impossible to see what matters by eye. So we layered a language model over the network, and the division of labour is the whole point. The network supplies interpretable structure. The model supplies interpretation at scale: it reads the full graph, surfaces the genes and pathways upregulated in HR-deficient patients, scores each uncertain node across more than 300 connected tools and databases including UniProt, AlphaFold, DepMap and ChEMBL for druggability, cellular location, essentiality and competing compounds, and runs the literature search that connects a target to a repurposable drug.

Note what the language model is not doing. It is not deciding that EZH2 matters, and it is not generating the explanation. It is reading a structure that was already biologically constrained and doing the volume of lookup a human could not. The expert still adjudicates. That is the shape Goh and colleagues argue for, and it is the shape we found works. The full case study has the method and the figures.

One of the least visible ways expert knowledge determines AI performance is through data. A model trained on data that is biased, mislabelled or biologically uninformative will produce biased, mislabelled or biologically uninformative outputs, regardless of architectural sophistication. With language models it will do so more confidently, in a form indistinguishable from correct output to anyone lacking the domain knowledge to evaluate it.

The perspective cites a finding from drug discovery that should give any tool buyer pause: a systematic review found that performance improvements attributed to model architecture frequently disappeared once training sets were curated to remove known data quality artefacts. Expert-driven data curation, not algorithmic sophistication, was the primary driver of the reported gains. The lesson generalises. The expert’s ability to recognise a batch effect, identify label noise, or flag that a molecular representation systematically disadvantages a class of compounds is not a secondary contribution to AI-driven research. It is frequently the primary one.

Which makes data work scientific work. Metadata standards, ontology alignment and provenance tracking are not administrative overhead; they are the foundation trustworthy AI is built on. Institutions that invest in this infrastructure consistently outperform those that do not, because they stop re-solving the same upstream problems on every new project. We have written separately about why identifier resolution is the bottleneck that most agentic pipelines quietly fail on.

The final shift is organisational, and it is the one with no automated component whatsoever. The full risk of AI adoption in biomedicine is not that individual analyses will be wrong, though some will be. It is that AI gets deployed at institutional scale, in clinical and research settings, by people who are not equipped to evaluate its outputs, in ways nobody has validated for the populations they will be used on.

The documented failures share a common cause: the absence of expert oversight at the point where training data, architecture and deployment context were defined. The most widely cited case is a commercial algorithm used to identify patients for additional care management, which systematically underestimated the health needs of Black patients relative to White patients with equivalent disease burden. The cause was a training objective that used healthcare costs as a proxy for health need, a design decision that encoded existing disparities into the model without any algorithmic error. Correcting it would have raised the share of Black patients receiving additional help from 17.7% to 46.5% (Obermeyer et al., 2019). Identifying that failure required clinical domain knowledge, statistical scrutiny and equity-aware evaluation the developers had not applied.

This is also becoming a legal obligation rather than a professional preference. The EU AI Act entered into force on 1 August 2024 and classifies AI systems used in medical devices and clinical decision support as high-risk, subject to conformity assessment, transparency requirements, human oversight obligations and post-market monitoring. The phased compliance dates for high-risk systems have been the subject of amendment during 2026, so check the current position before planning against a specific deadline. The direction is not in doubt: documented data governance, subgroup performance testing and human oversight are moving from good practice to requirement.

If the expert is the condition for AI producing real science, the useful question about any platform is not how autonomous it is. It is whether it lets an expert do their job.

This is the standard we hold ourselves to, and it is why Hydra is built around planning and validation rather than answer generation. It runs real bioinformatics analysis across roughly 100 preloaded scientific databases and more than 200 codified skills, and it validates every result rather than presenting a conclusion and inviting trust. The codified skills matter here specifically: they are where domain expertise is encoded so it applies on every run, rather than depending on whether the person at the keyboard happened to know that a given threshold was inappropriate for a given cohort.

That does not remove the expert. It is designed on the assumption that the expert is the point. What it removes is the part of their week spent writing code they have written before.

The bioinformatics profession is not becoming obsolete. It is being asked a clarifying question, and the perspective puts it better than we would: do you understand the science you are doing deeply enough to know when the machine is right, and when it only looks that way?

AI is genuinely capable of accelerating discovery. It finds patterns in data at a scale and speed no human analyst can match, surfaces candidate hypotheses from vast literature, and compresses the distance between raw data and interpretable results. But acceleration is not the same as validation. Transforming output into findings that are valid, reproducible, interpretable and clinically actionable requires expert judgement at every stage: in the design of the analysis, the curation of the data, the interpretation of the results, and the recognition of failure when it occurs.

Remove the expert and what remains is not science conducted more efficiently. It is the appearance of science, conducted at scale, by systems that cannot know when they are wrong. That risk exists precisely because AI is so capable of producing convincing outputs. Expertise is what ensures judicious use of AI to convert output into knowledge.

The test has already begun, and what is being examined is not whether you can use AI. It is whether you understand enough to be the expert behind it.