This is a test version of Biostars. For the public version, visit https://www.biostars.org.
Forum: AI for bioinformatics pipeline development. Is it a good idea?

AI can create a bioinformatics pipeline that runs smoothly. But can you find out if it is correct?

In the last year AI completely changed the way of programming. The development of tools that required weeks of hard work in the past can now be generated within hours. Variant calling? Differential expression analysis? Methylation calling? The AI will generate perfect code and a running pipeline. That is impressive. No discussion about this. But after running a new pipeline for the first time, the (for me) more interesting question is: Does the successful execution mean that the results are correct? Were the best tools for this task selected? Are the parameters set correctly? What about edge cases?

For me, bioinformatics is not only about writing clean code. That can be done by any programmer (or AI). The challenging part is always the connection of the designed workflow and its parameters with the scientific question you have. And keep in mind that, especially in research, a completely new composition of selected tools and parameters might be the best choice and not a composition that was previously used a lot (and thus learned by an AI). There is a huge amount of expertise, experience, knowledge and intuition involved when an expert designs bioinformatics pipelines.

I think that AI changed the complete world of computer science and also bioinformatics. It is great in generating pipelines. But I also think that expert validation becomes more and more important.

How is your feeling or even better your experience? Have you tested AI for bioinformatics pipeline development? How did it perform? Did you double-check the output? Were you impressed by the output?

bioinformatics ngs pipelinedevelopment artificialintelligence
• 532 views • link
updated by zabiullama • zabiullama • written by David Langenberger David Langenberger

Novice users often make the mistake of assuming that because a tool/pipeline "did not fail/exit" and produced a result, so that result "must be correct". This is something that needs to be addressed as a part of training. GIGO principal - Garbage in Garbage out is always applicable, especially when a string of tools is chained together.

One additional downstream issue is going to be documentation of such "pipeline" use in published papers. If authors don't make the code available and do not fess up (and simply include generic descriptions of tools) then it would be impossible to validate the results.

• 0 views • link
written by GenoMax GenoMax

It think it was always a problem that too many biologists have had to analyze their own data, because no bioinformaticians were available. Until now, they were forced to train themselves in how to program, be active in forums and attend training courses. But now, AI gives them the feeling that this is not needed any more. They just need to tell the AI what they want and they get a result with full interpretation... ready for publication. And then, the AI will use these results in the future for training.

• 0 views • link
written by David Langenberger David Langenberger

3 answers

Nice post and I agree with most of it !

Does the successful execution mean that the results are correct?

IMO, it is also true when a physical person is also writing code.

One can set up a nice workflow architecture, trying to anticipate all possible corner edges and still miss some obvious or not so obvious situations. That is what the issue section on GitHub is for, to allow thousands of people to bug report what a single brain has done.

As long as one has a systematic mindset to beta test and foresee where the pitfalls could be, there would not be more issue than when they would code it themselves using a keyboard.

The key part and hardest part is to preshot what the AI will do wrong.

Before, the bioinformatician had to think "what can the user do to break the application" and how can we prevent it. Now, another problem on top of the previous one is emerging, "what the AI can put in the application code that could generate an unanticipated behavior".

• 0 views • link
written by Bastien Hervé Bastien Hervé

What I see more and more is the situation where biologists with no background in computer science or bioinformatics use AI agents to analyze their data. That is alarming in my view, because they trust AI agents too much.

But I completely agree with you. The bioinformatics community is pretty aware of these issues and found strategies to overcome them. And they will find proper ways to double-check AI-generated code.

• 0 views • link
written by David Langenberger David Langenberger

Good one! While I echo the comments, as educators and professors, the onus is on us to ensure we develop and mentor biologists who can code, and above all with integrity and ethics.

This is what I call it as a hippocratic oath so that there is enough room for these people who vie for publications, not just copy and paste but understand and reinstantiate the code.

Yesterday at ISMB in DC, we had several policy frameworks that we were discussing in "Education" track.

We at Bioclues.org have been promulgating these ideas and ensure the school children are introduced to these formative ethics so that the next-generation will be in safe hands.

We should also ensure and train them where and how AI can go wrong!

Bottom-line: use natural intelligence Prash

• 0 views • link
written by Prash • Prash

Hi David, this post captures exactly the stress that I've been designing around for the past few months. I am building a multi-agent pipeline for NGS workflows(MSc, Environmental microbiology background). The Ai runs the workflow, but the entire architecture is built around what you just described: human approval gates at critical decision points. After QC, after ASV generation and before final report - the orchestrator pauses, summarises findings in plain English on the researcher's phones, and awaits approval before continuing.

The idea is that Ai does what it's good at, i.e., running tools reliably, formatting outputs, flagging anomalies, etc. While the researcher does what they are good at.

Would genuinely value your perspective on where you see the risk highest- is it in tool selection, parameter choice or downstream interpretation? Happy to share the architecture if you are curious.

• 0 views • link
written by zabiullama • zabiullama

Hi zabiullama,

thanks for sharing your project. I think that you are approaching this topic with exactly the right mindset. A human-in-the-loop framework with clear approval gates might be a very good way to combine the speed of AI with scientific validation.

For me, the highest risks are the following:

  1. Parameter selection: I think this is the highest risk, because wrong parameters are often very difficult to recognize. The pipeline will finish successful, the report will look good and no error message will appear. But the results can still be completely wrong. AI is very good in selecting standard parameters or default settings. But especially NGS datasets from fields like environmental microbiology are often not standard. A default cutoff might filter away important low-abundance taxa or generate artificial ASVs. The user will probably not recognize this, because the complete pipeline is running smooth.
  2. Downstream interpretation: For me this has probably the highest impact. AI is extremely good in generating a plausible biological story. Even when the statistical signal is weak, confounded or actually not existing. This becomes especially dangerous when the researcher does not understand the statistical assumptions or the previous processing steps. The interpretation might sound very professional and confident, but this does not mean that it is scientificly correct.
  3. Tool selection: I would classify this as a moderate risk. LLMs normally know the popular tools like FASTQC, DADA2 or QIIME2. But they will often select the tool that was used a lot in the training data and not necessarily the tool that is best for the specific scientific question. Especially for edge cases or completely new research questions, the most commonly used tool combination might not be the best one.

I really like your multi-agent approach and especially the idea of stopping at critical milestones like QC or ASV generation. I think this is very important.

One thing I would probably add: The plain-English summaries should not only show the final metrics. They should also explain why a specific tool and every important parameter was selected. What is the biological and technical reason for this decision? What would change when another parameter is used? Which results might be filtered away?

Otherwise the user might approve a step only because the summary looks plausible.

I think that expert validation will become more and more important, especially when AI generated pipelines become more complex and more easy to produce.

Best of luck with your MSc project!

• 0 views • link
written by David Langenberger David Langenberger

Hi David, thank you for the response. This is exactly the kind of feedback I was hoping for.

The way you mapped the risks reshaped my sense of where the approvals must be toughest. The point about pipelines running cleanly and being silently wrong is one I've been designing around without articulating that clearly, and the suggestion about explaining the reasoning behind each parameter choice, not just the metrics to be implemented in the design. Approval without reasoning is a theater, as you mentioned already.

On the interpretation risk, this is the one that I was afraid of the most when I read your reply, and it forced me to redesign how the AI behaves after the phyloseq object is generated. would rather continue that specific part of the conversation privately if you are open to it. There are architectural details I am working through that I would like to think out loud with someone who understands the validation problem the way you do.

Would you be open to connecting on LinkedIn? genuinely value your perspective and would like to keep in touch as this develops.

Thanks again for the thoughtful reply.

• 0 views • link
written by zabiullama • zabiullama

Log in to answer this question.

More posts like this