Bioinformatics in the Age of AI

Framing the right biological question

Kevin opens with a piece Tommy wrote on bioinformatics versus the AI hype, which he calls the unsexy craft. Tommy says he uses Claude Code daily for his data analysis, and that the models are very good and improving day by day. The hard part of being a good bioinformatician, in his view, is framing the question well. That takes deep domain knowledge, and since a bioinformatician can rarely match the biologists they work with, collaboration and communication matter: learn the biology from them so you know which questions are most important.

His example comes from hormone receptors. Someone studying the estrogen receptor in breast cancer and the androgen receptor in prostate cancer might look at the binding of both. AI will happily write scripts for any comparison, he says, but without the biology you may end up asking a question that doesn’t matter in that context. He also gives AI enough context about what he is studying, because otherwise, he says, it goes everywhere.

Confident answers and amplified experience

Kevin notes that AI is rarely good at saying “I don’t know,” and points to AI evaluations in biology, such as those from Latch Bio. Tommy goes further: AI will answer your question no matter what, and the more dangerous part is that it can answer wrongly and confidently. Without deep enough knowledge of the domain, he says, you can easily be fooled by results that are wrong or hallucinated.

That makes AI a tool for advanced practitioners, Kevin suggests. Tommy agrees that AI amplifies experience. He spent fourteen years accumulating what he knows, much of it through mistakes, programs that wouldn’t run, and searching for answers himself. Difficulty, he says, is part of how people grow, and the danger now is taking answers from AI without processing them critically. Kevin describes the same gap from the other side: troubleshooting his water heater with AI, he kept going in circles until he called someone who knew water heaters.

Tommy brings it back to pipelines. Anyone can ask AI to write a pipeline that aligns reads and counts them for RNA sequencing data, but someone who has never done it may not know which parameters to question, whether the defaults fit their data, or how read length changes the choice. You don’t know what you don’t know, he says, so you don’t know what to ask.

Foundations and the command line

For junior researchers, Tommy emphasizes a solid foundation: Unix commands and comfort with the terminal, as much biology as possible, and a programming language such as R or Python. People say they can just ask AI, he notes, but when something breaks you have to work out why. Kevin recalls that Tommy made the same case for the command line when they spoke eighteen months earlier, and asks whether he still holds it.

He does. Tommy learned Unix commands about fourteen years ago and became really comfortable after three or four years. Today, he says, coding agents such as Claude Code are far more powerful when you work directly from the terminal, and understanding paths, the file system, and where to put files and markdown instructions lets him transfer old skills to the age of AI. Kevin describes the same with Git: knowing branching and version control lets him direct an agent like a small dev team. Tommy, who considers himself a data scientist rather than a software engineer, adds that you can’t ask for practices such as feature branches or unit tests if you don’t know they exist.

The hammer, the nail, and when to use agents

Kevin mentions his Episode 5 conversation with Brandon Galang, who argues that when code is cheap, scope is at a premium. Tommy seizes on the word problem. AI is a tool, a hammer, he says: find the nail first, and you may not even need the hammer. In his view, many companies want to slap AI on everything when that isn’t always the right approach. Kevin says he often tells partners that a deterministic, automated job already does what they need.

Tommy notes that large language models are probabilistic, predicting the next token from the previous ones, which makes their output hard to reproduce. Tasks that are already deterministic belong in pre-written scripts. For next-generation sequencing, preprocessing is routine: Nextflow or Snakemake pipelines take FASTQ files to counts or to variant calls in VCF files. He would bring in an agent where judgment is needed, for example to decide whether lower-quality data are still usable for a given experiment.

Downstream, he uses AI as a thinking partner: to summarize the literature on a target, to brainstorm whether an analysis makes sense, and to teach him some of the biology, always validating what it says with wet-lab biologists who know the target well. Kevin adds that agent skills can encode those caveats, and Tommy says a skill can also carry an R or Python template so the agent doesn’t wander and waste tokens.

Shared skills and reproducible analysis

Asked how teams share this work, Tommy says routine analysis used to be packaged as R or Python functions everyone could reuse. Now everyone develops their own agent skills, and sharing them through a common marketplace is still a work in progress. He adds that the same skill may give different results with a different language model. Kevin describes how quickly recent model releases have changed his own preferences.

That pace, Tommy says, complicates reproducible analysis: models evolve and old ones are deprecated, so an analysis that depends on one may become impossible to repeat exactly. His workaround is to ask AI to write the scripts, which he can commit to a GitHub repository, understand, tweak, and rerun. Kevin calls this a best practice he sees in regulated industries: build quickly with AI, then lock down the validated step so the same inputs give the same outputs.

Tommy gives an example. He is not a software engineer, he says, but with Claude Code he can build an R Shiny app in a couple of days, or a Python application with Streamlit. He writes standardized scripts to process the data, then wraps them in an app for wet-lab scientists, who upload the spreadsheet from their instrument and get results that stay reproducible because the app runs the same deterministic script.

Teaching bioinformatics with AI

Kevin asks how Tommy would teach bioinformatics today. Before AI, Tommy says, learners had no choice but to search forums and Google. Now AI can help people learn much faster if they use it as a tutor rather than a shortcut: to explain concepts, build a learning plan for R, Python, or Unix commands, and find resources to read. He still tells junior bioinformaticians to build foundations in statistics, Unix, and programming, and Kevin adds biological principles. It took each of them about ten years, Tommy says, and experience changes how you see the same problem.

Kevin suggests that students will have to be deliberate about getting reps in, for example by asking AI to assess their understanding. Tommy compares it to the gym: everyone knows the principle, but people avoid pain, which is why GLP-1 drugs are tempting. Kevin notes that muscle still has to be built, and that skills such as judgment and context awareness can atrophy without practice. Tommy says he forces himself to understand what AI gives him before passing results to collaborators, because AI is not accountable for wrong results. Kevin agrees: the person who invoked the tool owns the output.

The entry-level dilemma

Looking five years out, Tommy describes a dilemma. Experience matters, so many companies hire more senior people, and fewer entry-level jobs are posted. Yet every senior bioinformatician started as a junior one, and he says the next generation still needs those opportunities. He and Kevin were lucky, he says, to be trained before AI and to use it now to supercharge their work, and he doesn’t yet see a solution for people starting out.

Kevin describes efforts he is part of: an advisory council for his alma mater’s College of Science and Math, and the LIFT-VA program in Virginia, which aims to upskill young scientists. He describes one approach used there, credited to Dr. Tshaka Cunningham, the Episode 3 guest: each trainee treats a cohort project like their own drug discovery company preparing a proposal to the FDA, and must defend it almost like an oral exam. Tommy’s advice to young people is to keep an open mind, embrace AI rather than hide from it, and apply it to small, real projects, the way he learned bioinformatics by applying it during his PhD. The traits that predict success, he says, haven’t changed: consistency, persistence, and a willingness to learn.

Where AI can and can’t change drug development

On frontier AI companies moving into the life sciences, Tommy asks them to be specific about which step of drug development they want to change. He walks through the process: finding a target, often now with AI co-scientists that read the literature and analyze data; designing small molecules or larger molecules such as antibodies and antibody-drug conjugates; optimizing the compound; toxicology studies in animals; phase 1 to 3 clinical trials; and the medical writing for an FDA submission. AI can make the work faster, such as writing code or documents, but uncovering new biology with AI is the hardest part, in his view, because biological data are sparse.

Kevin mentions cloud labs and robotics, with Ginkgo Bioworks in Boston as one example, and a recently announced virtual biology initiative involving the Biohub that aims to generate more biological data. Tommy says more data is a good direction, but many virtual cell foundation models rely mainly on single-cell RNA sequencing, while a cell also has proteins, post-translational modifications, lipids, metabolites, and the crosstalk among them. Even public repositories such as GEO hold vast amounts of data, he says, but only a small number of samples fit a specific disease, treatment, and mutation. Kevin sees the same pattern in pathogen genomics, where SARS-CoV-2 is unusual for its millions of sequences.

Model systems are another limit. Animal models, Tommy says, don’t fully recapitulate human biology, and he cites aspirin as a drug that, by his account, is far more toxic in some animals than in humans. Under the conventional regulatory path, he says, a toxicity signal in mice or dogs can stop a program before it reaches clinical trials, which he sees as a regulatory problem AI can’t solve.

Asked where AI is most ripe for innovation, Tommy points to the bottleneck at the end: late-phase clinical trials, where he says value is created. He estimates that a phase 3 trial takes about three years from the first patient to the close of data, with challenges in finding, recruiting, and transporting patients. Shortening that, he argues, could mean another year of sales for a successful drug. He also sees potential in manufacturing and in medical writing for FDA submissions, while noting that the space is heavily regulated and AI scientists will need partners with domain knowledge. Kevin observes that this bottleneck is about operations and logistics rather than cutting-edge biological models.