AI in bioinformatics earns its place in a few specific jobs: calling variants from sequencing reads, finding patterns in datasets too large for manual inspection, pulling evidence together from curated databases and the literature, and speeding up routine analysis steps. It does not yet replace the bioinformatician. The two frontier models tested on one benchmark answered open-ended analysis questions correctly at most 21% of the time, 44% of genomic studies in one review had leakage-inflated metrics, and structure prediction accuracy varies. Near-perfect accuracy is almost never a realistic expectation. Before a model’s output can be trusted, its data needs to be good and traceable, and a person needs to review it.

Where AI in bioinformatics already outperforms conventional tools

Variant calling is one clear case. A 2025 review in Frontiers in Bioinformatics reports that AI-based callers have outperformed conventional methods in key metrics, and that they handle repetitive and highly variable genomic regions where rule-based tools struggle. The same review gives DeepVariant SNP F1 scores of 99.9% on Illumina and PacBio HiFi data, and notes that it produces filtered variants directly, with no separate refinement step after calling. Clair3 is described as faster than the other leading callers and stronger at lower coverage, which is traditionally more prone to errors.

The headline numbers come with limits that matter in practice. DeepVariant’s InDel accuracy on Oxford Nanopore data falls to 76.8% because ONT reads have higher base-calling error rates. The models reviewed were all trained on Genome in a Bottle samples, which the authors describe as a small and limited dataset. They also say that limited interpretability raises questions about reproducibility and user confidence, especially in regulatory and diagnostic workflows. A 99.9% F1 on a well-characterised benchmark sample is not a promise about your cohort, your sequencing platform or your coverage depth.

Evidence synthesis is a second place AI helps. BioLunar, a research framework published on arXiv in 2024, chains gene enrichment, queries against curated knowledge bases such as CIViC and OncoKB, PubMed retrieval and LLM interpretation into one workflow, and compiles the findings into a report. Its value is in the time-consuming manual step after the initial NGS analysis, which the authors describe as error-prone. The Royal Society’s report on science in the age of AI puts the general case more broadly: AI techniques help find patterns and relationships in large datasets that would otherwise be too hard to see, and generative tools can speed up routine tasks such as processing unstructured data. For the wider view across biotech, our overview of AI in life sciences and where pattern recognition earns its keep covers more of the field.

Can an LLM agent actually run a bioinformatics analysis?

Not on its own, and not yet. BixBench, a benchmark published in 2025, gives agents an empty Jupyter notebook, real input data and open-answer questions drawn from more than 60 real-world analysis scenarios. In the open-answer setting, Claude 3.5 Sonnet scored 21% and GPT-4o scored 15%. Given the option to refuse, both models performed close to random. The authors conclude that a lot of development is still needed before anything like an autonomous bioinformatician exists.

BioAgents, a multi-agent system built on small fine-tuned models with retrieval-augmented generation, shows where the line falls today. On conceptual genomics questions it performed comparably to human experts at every difficulty level. On code generation it matched experts only on easy tasks and lost accuracy and completeness on medium and hard ones. The authors also observed diminishing returns from iterative self-refinement, where repeated rounds of revision could make outputs worse. One finding is more encouraging. Even when its answers were weaker, the system could name what it was missing, such as tool versions, reference genome details and compute resources, which the human experts found hard to do.

Our reading of this evidence: an LLM is a useful colleague for explaining which aligner fits an RNA-seq design or for drafting a first pipeline, and an unreliable one for producing multi-step workflow code that runs correctly without review. BioLunar’s authors are similarly cautious about their own system. They describe its workflow as hypothesis generation, fully supervised by humans and not meant for direct clinical use, and they note that hallucinations are possible. Our article on large language models in biotech and life sciences looks at the language-model side in more depth.

Structure prediction calls for the same caution. One published analysis notes that AlphaFold models only 36% of residues in the human proteome with very high confidence, against 73% for E. coli. It also notes that the models do not yet account for ligands, covalent modifications or environmental factors. The authors credit such methods with revolutionising structural biology while stressing that their accuracy varies.

Imbalance and leakage make genomics ML models look better than they are

A model can report excellent numbers and still mislead, and small, imbalanced genomic datasets invite this. A review of ML metrics in genetics and genomics gives the textbook case. In a dataset of 100 people with 10 affected, a model that predicts everyone is healthy scores 90% accuracy. Genomic datasets are often imbalanced in exactly this way, because control samples are easier to collect than samples from people with the condition, particularly for rare diseases.

Leakage is another common cause of exaggerated metrics. The same review cites work finding that 44% of the genomic studies examined had inflated metrics from data leakage during feature selection, with an average AUROC increase of 0.18. Unlike overfitting, this leakage inflates every metric equally, so care is needed both during training and when reading the scores.

Pitfall	How it shows up	What to check
Class imbalance	High accuracy driven by the majority class	Report MCC alongside accuracy; MCC uses all four confusion-matrix cells and copes better with imbalance
Leakage during feature selection	Every metric inflated together	Split the data before any feature selection, and keep the test set out of it
Overfitting through many trials	One of many models scores well on the test set by chance	Treat the scores as valid only for the data the model was trained and tested on
Population bias	Model built on one ancestry group performs worse on others	Report metrics separately by ancestry and sex where that data exists
AUROC used alone	Ranks models but says nothing about the deployed decision boundary	Choose and validate a decision boundary before judging deployed performance

This is why the review recommends choosing several suitable metrics based on the dataset and the research question.

Getting data ready for AI goes beyond FAIR

The FAIR principles apply to the algorithms, tools and workflows that produce data as well as to the data itself. FAIR is high level by design. Its authors state that the principles are not a standard or specification and do not prescribe any technology. Its reusability principle already asks for detailed provenance and clear usage licences.

The NIH Bridge2AI Standards Working Group requires FAIR only as a Level 0 baseline for AI-readiness. It defines seven dimensions of AI-readiness: FAIRness, provenance, characterisation, ethics, pre-model explainability, sustainability and computability. It treats deep provenance as a separate requirement because FAIR does not capture the complex derivation histories behind biomedical AI datasets. In practice that means any data shown to a model can be traced back to its unmodified source, such as raw NGS reads, imaging or clinical records, through every intermediate software and computation step. The group also notes that preparing AI-ready data takes more effort than capturing measurements for statistical analysis.

A short diagnostic before any model work:

• Can every training sample be traced to its raw source and processing steps?
• Has quality control been applied and documented?
• Is the dataset described in enough depth (distributions, biological assumptions, known biases) to predict how it will shape the model?
• Is the class balance known, and is the evaluation metric chosen to suit it?
• Are there processes that can generate more labelled data if the first model needs it?
• Does the workflow route uncertain outputs to a person, and will the team actually staff that review?

The Royal Society report puts the underlying point plainly. Incomplete, incorrect or unrepresentative data produces misleading outcomes, and data curators are essential to maintaining quality.

Making AI-assisted analysis reproducible and auditable

The nf-core authors note that fully reproducible pipelines remain notoriously hard to build across systems and institutions, primarily because of differences in hardware, operating systems and software versions. AI components make this harder, because model weights, inference settings and compilation behaviour join the list of things that can drift.

The sources describe practical fixes already in use. BioWorkbench captures provenance automatically, covering both execution performance and domain results, and reproduces its computational environment through a Docker container. BixBench runs every evaluation inside a versioned, pre-built container so that results do not depend on how software was installed. BioLunar’s reports state the analysis context, dates and software versions at the top. Even AlphaFold 3’s documentation notes that its compilation cache makes repeated runs reproducible, as well as faster.

Regulators have started to set expectations here. The FDA’s guidance sets out a risk-based framework for sponsors to establish the credibility of an AI model for a defined context of use, meaning how the model is used to answer a specific question. For a team whose analysis may feed a drug or biologic submission, that framing is useful even at the research stage. Write down what question the model answers, and keep the evidence that its output can be trusted for that question.

Lab automation standards and where automation stops paying off

Instrument data that arrives with its context attached is easier to feed into analytics. The Laboratory Automation Plug and Play framework builds on SiLA 2, which defines how instruments interface with automation software, and on LADS, a lab-specific layer over OPC UA whose first specification was released in November 2023. For data, the Allotrope Foundation’s format links raw data, results, metadata and provenance, and the ASTM AnIML standard gives a vendor-neutral XML format for analytical results from techniques such as chromatography and mass spectrometry. Instrument output in these formats is much closer to model-ready than vendor-specific files that need a bespoke parser.

Physical automation is a different calculation, with barriers specific to academic labs. A 2020 review in Frontiers in Bioengineering and Biotechnology describes an “automation gap” in academic labs, caused by short-term funding, high protocol variability and a culture that invests in people over equipment. It cites a text-mining study finding that 89% of published life-science articles included a manual protocol with an automated alternative. Potential is not payoff, though. Frequent protocol changes push labs towards bespoke systems with their own cost, and small variations in labware can make rigid handling equipment useless.

Automation also changes the shape of errors. A machine that dispenses a reagent wrongly can repeat the mistake across thousands of samples before anyone notices, and machine-to-machine variability can be harder to spot than manual variation. The review treats careful maintenance, calibration and quality control as essential. Our view is that this mirrors the AI case: both need someone checking the output. For applied examples beyond the lab bench, see our roundup of AI applications in biotechnology.

Sizing compute for AI bioinformatics workloads

Published figures give concrete reference points for sizing compute. The AlphaFold 3 repository supports inference on a single NVIDIA A100 with 80 GB of memory. It reports that setup as at least 2× more efficient in GPU-seconds than the paper’s 16 × 40 GB A100 configuration, across all token sizes. Compilation dominates the runtime of a single fold job. On a 152-residue chain, a warm compilation cache cut a second run from 143 s to 32 s. Disk speed influences the data pipeline’s genetic search speed, so storage matters too.

In one single-cell study, GPU memory grew far faster than runtime. A profiling study of a GPU-accelerated single-cell workflow measured peak GPU memory rising roughly linearly from 11.5 GB at 100k cells to 102.5 GB at 1M cells, about 10.1 GB per 100k cells. Over the same range, runtime grew only from 64.1 s to 151.8 s. That run used an NVIDIA H200 NVL with 141 GB. A 48 GB or 80 GB card would hit its limit well before 1M cells.

For variant calling, the review notes that DeepVariant and DeepTrio rely heavily on GPU or HPC resources, and that their requirements grow significantly with long-read data. It points to DNAscope and Clair3 as more resource-efficient options for labs without HPC or cloud access. Cloud computing is a viable alternative but raises data security, cost scaling and compliance concerns, particularly in clinical settings. Plain parallelisation still matters too. BioWorkbench reports roughly 4× and 5× speedups over sequential Python from running samtools- and GATK-based workflows in parallel with Swift.

Choosing where to start

Task	AI fit today	What must be in place first
Short-read germline variant calling	Strong; AI callers outperform conventional tools on key metrics	GPU or efficient caller choice, validation on your own platform and coverage
Long-read InDel calling	Mixed; depends on tool and coverage	Adequate coverage, platform-specific validation
Literature and knowledge-base synthesis	Useful for generating hypotheses	Human supervision, cited sources, versioned reports
Conceptual pipeline guidance	Comparable to experts in one study	Expert review of the recommendation
Generating complex workflow code	Weak on medium and hard tasks	Bioinformatician review and testing before any run
Predictive models from genomic data	Possible, often over-reported	Leakage-free splits, imbalance-aware metrics, provenance
Protein structure hypotheses	Valuable, with confidence varying by region	Confidence-aware interpretation and experimental follow-up

The question we see as decisive is organisational rather than technical: who owns the review step. BioLunar’s workflow is fully human supervised, and the FDA framework asks sponsors to establish model credibility for its use. A pipeline that is fast but has no one assigned to judge what the model got wrong has only moved the error downstream, where it costs more to find.