Learning Bioinformatics in the Age of AI
TL;DR — AI has not made bioinformatics easier to learn. It has moved the difficulty. Writing a script is now cheap; knowing whether the output is biologically true is expensive. Build your training around the second thing.
Why this post exists
Every few months someone announces that bioinformatics is about to be automated away. Meanwhile, the actual job listings keep growing, and the people doing the hiring keep saying the same thing: they need scientists who can judge whether a result is real.
That tension is the whole story of 2026. Models are extraordinarily capable. Model outputs are frequently wrong in ways that are invisible unless you know the biology and the statistics. If you are starting now — whether you’re a bachelor’s student who just discovered that pandas exists, or a PhD student staring down a Perturb-seq dataset — this is a genuinely good moment to enter the field. But the path looks different from what your supervisor walked.
Part 1 — What actually changed
Four shifts matter for how you should learn.
1. Structure prediction stopped being a bottleneck and became a starting point
AlphaFold2 solved the single-chain problem well enough that most people stopped thinking about it. AlphaFold3 extended this to complexes — proteins with ligands, nucleic acids, ions — and the open-source ecosystem caught up fast. Boltz-1, Boltz-2, BoltzGen and Chai-1 now give you AlphaFold3-class co-folding under permissive licenses, and Boltz-2 goes further by predicting binding affinity, not just geometry.
What this means for you: predicting a structure is a homework exercise now, not a thesis. The thesis is in what you do with 10,000 predicted structures, and in knowing that these models still fail badly on disordered regions and that confidence scores like pLDDT and pAE were designed to measure structural confidence — not functional relevance.
2. DNA and RNA got their own foundation models
Two 2026 Nature papers define the current frontier. AlphaGenome takes 1 megabase of sequence and predicts thousands of functional genomic tracks — expression, chromatin accessibility, histone marks, TF binding, splicing — at single-base resolution, and matched or beat the best available external models in 25 of 26 variant-effect evaluations. Evo 2 was trained on roughly 9 trillion base pairs spanning all domains of life with a 1-million-token context window, and predicts functional consequences of variation without task-specific fine-tuning.
The interesting downstream work is already appearing: probing these models’ internal representations to produce interpretable pathogenicity predictions, because a clinical geneticist working under ACMG/AMP guidelines needs categorized evidence, not an opaque score.
3. Agents arrived and are genuinely being used
Biomni, published in Science in July 2026, is a general-purpose biomedical agent that mines tools, databases and protocols from tens of thousands of papers, then composes workflows on the fly — no predefined templates. Its environment bundles roughly 150 specialized biomedical tools, 105 software packages and 59 databases. The prototype was reportedly in use by more than 10,000 labs before formal publication.
You should try it. You should also notice what it does not do: decide which question is worth asking, or tell you when the answer is biologically absurd.
4. The honest part: benchmarks keep humbling the hype
This is the section most “learn bioinformatics with AI” posts skip, and it’s the most important one for a student.
- A 2025 Nature Methods study compared five foundation models and two other deep learning models against deliberately simple baselines for predicting transcriptome changes after genetic perturbation. None of them beat the baselines.
- A separate benchmark found that the simplest possible baseline — taking the mean of the training examples — outperformed both scGPT and scFoundation on post-perturbation prediction.
- A 2026 preprint showed that careful normalization plus linear methods reached state-of-the-art or near-SOTA across standard single-cell benchmarks, and actually beat foundation models on out-of-distribution tasks involving novel cell types and organisms.
- VCBench (2026) evaluated five single-cell foundation models against pre-registered linear and nearest-neighbour baselines and reported that no foundation model publishes a complete cell-level training manifest — which makes data contamination undetectable to users.
None of this means the models are useless. It means benchmarking is a first-class skill, and “I ran the big model” is not a result. The single most employable habit you can build this year is reflexively asking: what’s the dumbest baseline that could work, and did I beat it?
Part 2 — The core that AI does not remove
Some things got cheaper. These did not.
| Skill | Why AI didn’t kill it |
|---|---|
| Statistics | Multiple testing, batch effects, confounding, and effect size vs. significance. An LLM will happily hand you a p-value from a design that can’t support it. |
| Experimental design | Power, replication, controls. If the data can’t answer the question, no model can. |
| Data wrangling reality | Real data is messy in specific, biological ways — sample swaps, adapter contamination, ambient RNA, reference build mismatches. |
| Reproducibility engineering | Git, environments (conda/mamba, pixi), containers (Docker/Apptainer), workflow managers (Nextflow + nf-core, or Snakemake). |
| Domain biology | Knowing that a “significant” gene set is dominated by ribosomal proteins because of a normalization artifact. |
| Scientific writing | Methods sections that someone else can actually execute. |
Notice that reproducibility engineering is more important now, not less. When an agent generates 400 lines of analysis code in four seconds, provenance is the only thing standing between you and an irreproducible thesis chapter.
Part 3 — A path by degree level
Bachelor's: build the floor
Months 0–6. Do not start with foundation models. Start with the thing they sit on top of.
- One language, properly. Python (with
numpy,pandas,matplotlib,scikit-learn) or R (tidyverse,Bioconductor). Pick based on your lab’s ecosystem. Learn one deeply before adding the other. - The command line and Git. Non-negotiable. Bash, ssh, file permissions,
grep/awk, and version control from day one. - Statistics for biologists. Hypothesis testing, regression, multiple testing correction, and — crucially — what a confidence interval actually means.
- One complete analysis, end to end. Download a public RNA-seq dataset from GEO or SRA. Do QC, alignment or pseudo-alignment, differential expression, enrichment. Write it up as if for a supervisor.
Project idea: reproduce the main figure of a published paper from raw data. When your numbers differ, find out why. That investigation is the learning.
Master's: specialize and industrialize
You have the floor. Now pick a vertical and learn to work at scale.
- Choose a domain: single-cell and spatial omics; population and statistical genetics; structural bioinformatics and protein design; metagenomics; clinical/translational genomics; or ML methods development.
- Learn workflow management for real. Take a pipeline you wrote as a pile of scripts and rewrite it in Nextflow. Then look at how nf-core does it and feel appropriately humbled.
- Learn HPC. Job schedulers, resource requests, GPU allocation. In Sweden this means getting an account through NAISS.
- Use the frontier models — critically. Run Boltz on a complex you care about. Query AlphaGenome for variants in a locus you know well. Compare an scFM embedding against a plain PCA/HVG baseline on your own data. Write down what you find, especially when the simple thing wins.
- Learn a bit of ML properly. Not just
model.fit()— cross-validation strategy, data leakage, distribution shift, calibration.
Project idea: build a small benchmark. Take three methods for one task, three datasets, pre-register your metrics, and report honestly. This is publishable at master’s level and it teaches you more than any course.
PhD: contribute something the models can't
Your comparative advantage is not implementation speed. It’s taste, depth and rigour.
- Own a biological question, not a method. Methods age in 18 months; questions last a career.
- Become the person who knows the failure modes of the tools your field depends on. That reputation compounds.
- Treat benchmarking and negative results as real contributions. The papers cited in Part 1 are among the most-read in their subfields precisely because someone did the unglamorous comparison.
- Engage with governance early. The EU AI Act is rolling out in stages through 2027, with obligations for general-purpose AI models already in force. If your work touches clinical data or decision support, logging, traceability and risk management need to be designed in, not bolted on.
- Build one thing others use. A package, a well-documented pipeline, a curated dataset. It’s the most durable career asset you can produce during a PhD.
Part 4 — How to actually use LLMs while learning
The failure mode is specific and common: you get the right answer and learn nothing. Some guardrails that work.
Do delegate: boilerplate plotting code, regex, file format conversion, API scaffolding, “explain this error message”, literature triage, first drafts of documentation, translating a method between languages.
Don’t delegate: deciding the statistical test, interpreting whether a result is biologically plausible, choosing what to compare against, or writing the reasoning in your methods section.
Three habits worth building:
- Predict first. Before you run the model’s code, write down what you expect the output to look like. Compare. Every mismatch is a lesson.
- Demand the baseline. Ask explicitly: “What’s the simplest method that could solve this, and how would I compare against it?” Then actually run it.
- Verify at the boundary. LLMs are strongest on well-documented, widely-used tools and weakest on version-specific behaviour, obscure file format edge cases, and recent library changes. Check anything that touches a version number against the actual docs.
A concrete verification loop for generated analysis code:
# 1. Does it run on a tiny subset first?
head -n 10000 big_file.fastq > test.fastq
# 2. Do intermediate dimensions match your expectation?
# (rows == cells? columns == genes? or did it silently transpose?)
# 3. Does a known-positive control behave correctly?
# e.g. do housekeeping genes stay flat, does the spike-in behave?
# 4. Is it pinned and captured?
git add . && git commit -m "analysis: DE with pinned env"
One more thing worth internalizing: a widely-cited 2025 MIT study found that the large majority of enterprise generative-AI pilots failed to deliver measurable impact — mostly not because models were weak, but because they sat disconnected from real workflows and data foundations. The same failure pattern shows up in individual research projects. A brilliant model attached to a badly-designed experiment produces confident nonsense faster.
Part 5 — Where to learn (with a Swedish accent 🇸🇪)
International and free:
- nf-core — community-curated Nextflow pipelines; reading their code teaches production bioinformatics.
- Bioconductor workflows and Galaxy for GUI-based learning.
- Rosalind for algorithmic fundamentals.
- The Carpentries lessons for shell, Git and programming foundations.
- bioRxiv preprint alerts in your subfield — in a field moving this fast, waiting for journal publication means being 12 months behind.
Part 6 — What the job market is actually asking for
Industry survey data from 2026 shows where AI has genuinely landed in life science R&D: literature and knowledge extraction is the most widely adopted application, followed by protein structure and property models, scientific reporting, and target identification. Around half of biotech organizations report faster time-to-target already. Growth areas over the next two years cluster around workflow automation, multimodal models, and deeper data integration.
Read that as a skills list. The people who benefit are the ones who can connect models to real workflows, validate the outputs, and explain both to a biologist and to a regulator.
As one bioinformatics services group put it in their 2026 outlook: AI proficiency is no longer optional, but the differentiator is deep domain expertise and scientific judgement — because AI can assist with analyses but cannot decide which questions to ask or validate whether results make biological sense.
A 12-week starter plan
If you want something concrete to begin tomorrow:
| Weeks | Focus | Deliverable |
|---|---|---|
| 1–2 | Shell, Git, project structure | A GitHub repo with a sane layout and a README |
| 3–4 | Python or R fundamentals + data wrangling | A cleaned, documented dataset from a public source |
| 5–6 | Statistics for genomics | A written analysis with justified test choices |
| 7–8 | One full omics workflow (e.g. RNA-seq) | Reproduced figure from a published paper |
| 9–10 | Environments, containers, Nextflow | Same analysis, now as a portable pipeline |
| 11 | Add an AI-era tool (Boltz, AlphaGenome, or an scFM) | Comparison against a simple baseline |
| 12 | Write-up | A blog post or report someone else can follow |
Twelve weeks won’t make you a bioinformatician. It will make you someone who can be trained into one — which, right now, is what supervisors and hiring managers are actually looking for.
References and further reading
- Avsec et al. (2026). Advancing regulatory variant effect prediction with AlphaGenome. Nature 649:1206–1218. doi:10.1038/s41586-025-10014-0
- Brixi et al. (2026). Genome modelling and design across all domains of life with Evo 2. Nature 652:1349–1361. doi:10.1038/s41586-026-10176-5
- Huang et al. (2026). Autonomous biomedical research with an artificial intelligence agent (Biomni). Science. doi:10.1126/science.adz4351
- Ahlmann-Eltze, Huber & Anders (2025). Deep-learning-based gene perturbation effect prediction does not yet outperform simple linear baselines. Nature Methods. doi:10.1038/s41592-025-02772-6
- Benchmarking foundation cell models for post-perturbation RNA-seq prediction. BMC Genomics (2025). doi:10.1186/s12864-025-11600-2
- VCBench: A Multi-Dimensional Benchmark for Single-Cell Foundation Models. bioRxiv (2026). link
- Guo et al. (2025). Foundation models in bioinformatics. National Science Review 12(4). doi:10.1093/nsr/nwaf028
- Fios Genomics (2026). Bioinformatics in 2026: What to Expect. link
- Benchling (2026). Biotech AI Report. link
- NBIS training catalogue: nbis.se/training
Leave a comment