This is a test version of Biostars. For the public version, visit https://www.biostars.org.
Forum: Real-world adoption of AI/ML in bioinformatics workflows: Questions from a junior researcher

Hello everyone,

I am a final-year biotechnology student preparing to transition into a Master's program in computational biology. Currently, my daily skill set revolves around traditional bioinformatics tools, R, Python, and working within Linux/HPC (SLURM) environments.

While there is a massive amount of hype surrounding Artificial Intelligence and Machine Learning in biology right now, I want to understand its actual, day-to-day integration into real research pipelines before I dive into graduate studies.

I would love to hear from active researchers:

  1. Practical Adoption: Roughly what percentage of your routine computational workflows actually rely on AI/ML models compared to traditional algorithmic tools?
  2. Reliable Applications: In which specific sub-fields or computational steps (e.g., structural biology, noise reduction, data filtering) are you finding AI tools to be genuinely useful and reliable in production right now?
  3. Advice for Next Steps: For someone comfortable with basic programming and HPC environments, what specific AI/ML frameworks or foundational concepts would you strongly recommend focusing on?

Thank you for sharing your real-world experiences and guidance!

bioinformatics machine-learning career discussion artificial-intelligence
• 192 views • link
updated by i.sudbery i.sudbery • written by oguzhanisilay • oguzhanisilay

2 answers

I think it depends on what you mean by AI/ML. I like to separate out general AI/ML from narrow AI/ML.

By general AI, I mean LLM and agentic based systems, such as Claude, Gemimi, etc that people use for coding, literature reviews etc. By narrow AI/ML, I mean models specifically trained to solve a specific scientific problem. Famous examples include AlphaFold, ESM2, etc.

For general AI/ML, I say most people are now using some sort of AI help with their coding. Much rarer is actually getting help from AI in thinking about science, although its not unheard of. I would not trust anyone actually using it to do the analysis for them.

For narrow AI/ML, I'd say that my work is using an ML algo in routine pipelines maybe 10% of the time, although I'd like it to be more. In terms of what tools/algos we use, we've used SpliceAI a fair bit (which is a CNN), AlphaGenome a little (a transformer based model). Most of the time we'd use our own models though. The majority of the time simple models do what we need: Lasso regression or Gradient Boosting. Most of the time we are interested in these models for the feature selection ability - we want to know which features are important for a classification or regression we are studying.

As for where to get started - I recommend Andrew Ngs courses on Coursea: The Machine Learning Specialisation and Deep Learning specialisation.

• 0 views • link
written by i.sudbery i.sudbery
  1. 30-50% depending on area
  2. structural bio, nanopore seq work, all sorts of genomic models may come good some day. Maybe variant calling in future, alignment not so much so far.
  3. Pytorch seems ubiquitous. But get on github and follow lots of bioinformaticians, awesome lists and gpu topics in bioinfo!

Also AI for coding seems like it is having a bigger impact than tools using AI at the moment - standards are being lifted at long last in bioinfo coding.

• 0 views • link
written by colindaven colindaven

Log in to answer this question.

More posts like this