Bioinformatics lies at the intersection of biology and mathematics. There are typically two paths one finds themselves on to wind up in bioinformatics: a quantitative background, applying these skills to biology; or a biological background, applying these skills to identify and solve problems with a newly learned quantitative skillset. Bioinformatics is for those who appreciate a diverse skillset – it is unlikely any one formal undergraduate training covered all the skills needed to succeed in this field, and yet, since so few are equipped by default to work in this field, the demand for those who take on this challenge is very high.
The scope of bioinformatics is vast and growing as new tech generates more data. The job of a bioinformatician includes collecting data, which could come from high throughput sequencing, microarrays, medical images, clinical records, etc. Their job is also to apply some sort of mathematical preparation which may include normalization, subtraction of background noise, and removal of outliers. Often the job includes integrating this data from multiple sources, for example combining multiple omics technologies to consider multiple cellular layers in unison. Sometimes, the job involves developing and curating knowledge stores for your biological data or software – including support users and updated versions to keep up with rapidly changing technology. Finally, model generation and interpretation of this data is key to interface back to the biological and medical communities, often providing new interpretations of various disease or biological phenomena.
On the relationship between biology and other sciences
Bioinformatics belongs to a long tradition of borrowing quantitative methods, whether from physics, mathematics, and engineering, and giving back in return. D’Arcy Thompson crossed that boundary between biology and quantitative methods in On Growth and Form (1917), arguing that mathematics and physical laws were the key to understanding biological structure.1 Fisher, Haldane, and Wright meanwhile discovered that genetics and evolution could be fitted with mathematical foundations and understood as mathematical disciplines.2 After laying the foundations for modern computers, Alan Turing showed quantitatively that interacting chemicals diffusing through tissues can generate patterns in life, founding the mathematical study of Morphogenesis.3 Edwin Schrödinger (a famous theoretical Physicist and foundational thinker in quantum mechanics) proposed from physical principles that an “aperiodic crystal” could serve as a “code-script” for storing hereditary information in cells.4
Crossing the boundary in the opposite direction, many advances in engineering and computer science were taken from observations in the biological world. Norbert Weiner and others developed cybernetics, which studied of control, communication, and feedback systems in animals and machines, leading to much of modern control systems and robotics technology. He argued that some of the most productive science occurs in the “no-mans-land” between disciplines such as biology and mathematics.5 We are obliged to agree!
Finally, McCulloch and Pitts created a mathematical model of neurons in the brain that served as the ancestor to modern artificial neural networks. It is perfectly plausible to say that modern AI is biology inspired.6 Since this field has benefited so much crossing the boundary in one direction, we are also obliged to ask – how is biology now benefiting from developments in AI?
AI has changed the game
Due to the rapid development of AI-assisted coding, learning, and demand to apply the rich array of new AI developments on untouched biological and medical problems, the field of bioinformatics has changed. A plausible python-based RNA-seq analysis pipeline is now available with a few clicks and information about input data, which previously would have involved months of coding manually, or at least days of learning the ins and outs of existing packages. For traditional workflows, the skillset needed to implement these analyses has shifted even more away from memory of coding syntax and patience navigating package references, and much more to interpreting the data being analyzed. Coding copilots are printing presses for code, and they love to freestyle. In the hands of a careful reviewer with plenty of attention to test cases, these can be miraculous productivity and quality of work boosters. In the hands of the incautious or overwhelmed, they can be tools for churning out nonsense!
Bioinformatics workflow BC (Before Copilot)
My father is a big rig truck driver. When I was learning to drive, he was very clear: I had to learn how to do it the old fashioned way, by busting out the dusty road map and reading the street signs. Despite having plenty of GPS options that could easily get me from point A to point B, I had to learn it the old fashioned way because “one day the GPS won’t be available, and then what are you gonna do?”
Much of the same if you take a bioinformatics course from me! Although there is plenty of technology around to augment the learner’s experience, I think a healthy dose of practice using the old fashioned way is needed to handle the firehose of AI-generated assistance that’s available from getting from point A to point B. For what it’s worth, I have never found an occasion where I wasn’t able to use GPS and had to use the map to this day, so I am happy to concede those that take an AI-first approach to learning may have the right of it! However, quite a lot of the average bioinformatician’s work nowadays is proofreading and shaping the co-pilot generated syntax. This is very difficult if you’ve never had to generate the code yourself.
A day in the life of a bioinformatician BC (before copilot) usually involved setting your mind on a task and carrying out the (sometimes excruciating) work of figuring out how to actually get it to run. This included getting to know your data, preprocessing it, finding the right existing software, finding out the software you need doesn’t exist, finding out how to develop the software you need, finding out actually somebody did already develop it- it was just never cited so you never found out about it, and so on. Dependency installs, converting file formats, adapting existing scripts, and spending lots of time reading references, and lots more debugging your preliminary products was maybe 90% of what the job entailed. This long slog through days of working to break a brick wall only to be rewarded with another brick wall are often enough to turn off a large fraction of prospective bioinformaticians. But for many, that other 10% of the job, where one gets to focus on generating and executing on a new hypothesis, possibly making a new discovery or developing a method no one ever thought to previously, was just too rewarding to give up.
Take analyzing RNA-seq: your PI or colleague comes to you with raw sequences reads they’re interested in investigating to see what’s expressed differently in cancer versus normal tissue. Your job now is to know (or google) what quality-control (QC) software, aligners, reference genomes, coding language, and packages you’re going to need to get to the bottom of this. Then you need to iterate with your data owner, go over preliminary results, and figure out the best way to tell your owner enough is enough and to wrap the project already! This could take anywhere from months to years to get from start to finish. A remarkable amount of time is consumed in mechanics, but at the end you have pushed the frontier or biological understanding forward.
Bioinformatics workflow AD (AI-Dominated)
AI copilots have dramatically diminished the amount of time one has to spend addressing the tedious work above. A bioinformatician can now describe the computing environment, data, and goals of the analysis, and have a plausible pipeline to address their goals assembled in minutes. The difficult work increasingly revolves around asking if the right decisions are being made. Is this the right analysis for my problem? Are the assumptions all sound? Are there any logical jumps the pilot code has included that are incorrect? Are there batch effects, confounders, low-quality samples, or annotation issues to worry about? Most importantly, what does the biological output mean?
If we are unsure, how do we go about resolving it? We consult the literature of course! But the utility of this is even receding. As AI slop analyses are increasing, the literature is also increasingly flooded with AI-generated workflows which may seem plausible on the surface but are either contrived or inappropriate for the data used. That’s without even addressing researchers using it to draw (sometimes spurious) conclusions about their data.
On the optimistic side, since the arduous 90% has been reduced dramatically, so we potentially have the benefit of spending more time on the old 10% - hypothesis generation, discovery of new trends, rapidly carrying out secondary and tertiary analyses. Since much of the incidental work has been taken care of, we (hopefully) have more time for science. So what have we been doing with it?
Recent Successes
Quite a lot, as it turns out. In the past few years bioinformatics has predicted structures for more than 200 million proteins via AlphaFold7, uncovered roughly 119 million base pairs absent from previous reference assemblies via the Human Pangenome Project8, 100 million human cells have been catalogued for single-cell analysis across tissues, diseases, and developmental states by the Human Cell Atlas and related initiatives9, and thousands of patients with rare diseases have been given diagnoses via the 100,000 Genomes Project.10 AI Scientists are increasingly being explored to rapidly solve medical and biological problems that have long eluded human researchers.11 At the same time, the field of bioinformatics as a whole is shifting to large scale multi-source, multi-modal projects attempting to make the most comprehensive data-driven picture of patients (which has been the focus of much of my work).12-14 The overlap between big data generation and big data analysis has led to an explosion of new advances, with more to come in the years to follow.
No Place for Bioinformatics?
To round out the learning to drive analogy, it’s worth mentioning that self-driving cars may be in the process of replacing drivers on the road. Whether this means that manual driving will become obsolete or more of a parlor trick for small numbers in the future is an open question. For practical purposes, its likely there will be a period of transition if it happens at expanding the opportunities for automated driving in some places, and leaving the human in others. What about with bioinformatics?
I have no interest in stating the case that bioinformatics as a discipline cannot or should not be eliminated. Bioinformatics is for better or worse interdisciplinary, which means you can always flee back to one or more of those disciplines when necessary. I think what’s most likely is we will see an opportunity for rapid upskilling of existing biologists and steering of the quantitatively able into biological and medical problem-solving. What that means is less individuals will be bioinformaticians, and more individuals. It has been common in the last decade or so for biologists to learn a little coding to analyze their data, and a little goes a long way. For more daunting tasks, the local bioinformatician would be called in maybe in exchange for lower authorship on the work. Nowadays I think we will see a shift where an interested wet lab biologist can learn a lot of coding (or perhaps I should say produce a lot of code). Perhaps fewer people will have “bioinformatician” as their professional title, while more biologists will routinely do work that we would currently call bioinformatics. That does not mean all biologists will take an interest in this subject area, just that the number it’s easily accessible to will increase. In my experience lecturing intro to bioinformatics a good chunk of the room would like to upskill this way, but found it daunting or unappealing to spend the day digging through syntax and memory bugs. This is a bit of fat that can be efficiently cut by copilots, and usually for the better.
In this way, I think it is more important than ever to have human-to-human learning materials available for bioinformatics learning. I am not sure whether from a technical standpoint the discipline cannot be eliminated by coming technology, but I am fairly confident biologists will always want a good understanding of the mathematical and statistical principles underlying AI-assisted work, even if they can focus on the intricacies of software references, syntax, and machine code less.
For more human-to-human content, consider following on LinkedIn or Substack.
References:
1 Thomson, J. A. On Growth and Form. Nature 100, 21–22 (1917). https://doi.org/10.1038/100021a0
2 Charlesworth, B. & Charlesworth, D. Population genetics from 1966 to 2016. Heredity 118, 2–9 (2017).
3 Turing, A. M. The chemical basis of morphogenesis. Bulletin of mathematical biology 52, 153–197 (1990).
4 Schrödinger, E. What is life? The physical aspect of the living cell. (Cambridge University Press, 1944).
5 Wiener, N. & von Neumann, J. Cybernetics or Control and Communication in the Animal and the Machine. Physics Today 2, 33–34 (1949).
6 McCulloch, W. S. & Pitts, W. A logical calculus of the ideas immanent in nervous activity. The bulletin of mathematical biophysics 5, 115–133 (1943).
7 Stroe, O. AlphaFold predicts structure of almost every catalogued protein known to science, <https://www.ebi.ac.uk/about/news/technology-and-innovation/alphafold-200-million/> (2022).
8 Liao, W.-W. et al. A draft human pangenome reference. Nature 617, 312–324 (2023).
9 Institute, W. S. Human Cell Atlas, <https://www.sanger.ac.uk/collaboration/human-cell-atlas>
10 Walker, S. Finding answers for participants: 10 years of genome analysis, <https://www.genomicsengland.co.uk/blog/10-years-genome-analysis> (2023).
11 Gao, S. et al. Empowering biomedical discovery with AI agents. Cell 187, 6125–6151 (2024).
12 Project, A. AIDA, an AI diagnostic assistant for detecting lesions before gastric cancer, <
https://www.aidaeuproject.org/>
13 Higgins, K. et al. The Helicobacter pylori AI-clinician harnesses artificial intelligence to personalise H. pylori treatment recommendations. Nature communications 16, 6472 (2025).
14 Higgins, K., Gonzalez, G., Veselkov, D., Laponogov, I. & Veselkov, K. PPI-Net connects molecular protein interactions to functional processes in disease. arXiv preprint arXiv:2605.07838 (2026).
This is a very timely article in such an overwhelming time, for someone with a biology background with some coding experience (which of course is now being scaled by AI). I've always wanted to be a Bioinformatician with the curiosity to learn more about the math and programming for tools development. With the rapid growth of AI, where do you think the competitive advantage lies, both as an individual learning and in general?
This was a great read, and the framing of the shift from the 90 percent to the 10 percent is the clearest version of that argument I have seen.
The part that does the most work is where you point out that consulting the literature is itself getting less reliable, because the same copilots are filling it with plausible-looking workflows. That lands harder because of the successes you list just after it. AlphaFold, the pangenome reference, the Human Cell Atlas, the 100,000 Genomes cohort: every one of those is a curated, versioned, heavily vetted reference layer, and that curation is why anyone can build on them without re-deriving the whole thing.
The literature has never had an equivalent layer, which may be why it degrades first under the same pressure. And that makes your GPS argument extend further than coding. The reviewing instinct you get from having written the pipeline yourself is the same instinct that tells a sound paper from one that merely reads like one, and that one has no obvious substitute yet.