Why is AI for Biology so difficult? (Thoughts from NeurIPS 2025)
Published:
I never doubt the ability of AI and the AI scientists to revolutionize the field. The most well-known example is AlphaFold, which genuinely changes the way that we form hypotheses in protein science.
But what else? Any more examples where AI is integrated into basic biological discovery?
To be fair, there are also medical imaging (for example, MRI and PET-CT) processing tools, based on computer-vision-inspired deep learning models, mostly built for clinical applications. Such tools are pretty lucrative, especially if you know how much a scan costs in the US.
And more recently, we have the single-cell foundation model/LLM agent. Personally, I have no idea what is happening under the hood. My fictional pet parrot says they are more like a joke or a way to get easy money, especially when NIH and NSF loses funding.
There are probably more AI tools for clinicians, like electronic health records or diagnostic agents for patients, but I won’t count them as basic biology – they are closer to healthcare.
This begs the question – why are there so few tools for biological applications? The answer is a few hurdles that AI has to clear to enter biology.
A well-defined problem. Biologists like to make big statements. For example, “I want to understand protein interactions within a cell”. That’s not a well-defined problem. First, what cell? In what environment? What protein? At which state? By interaction, do you mean binding or downstream effects like binding? This statement is probably as hard as a computer scientist declaring, “I want to understand learning and why it works”. An open question is no good for computer scientists, while most biological problems are open and too broad for AI to make sense.
A good enough collection of datasets for training. Suppose we have a specific enough problem. For example, I want to “find the rate at which membrane transporter protein X transports some drug Y into cells at 38°C”. I need training and testing datasets. That could not be a few data points (<10) gathered from a couple of papers done by 50 postdocs/PhDs spanning 30 years using 20 different technqiues. You need a systematic database that is internally comparable and allows training of a complicated model. Like the one and only PDB (Protein Data Bank, a protein structure database).
Profits. Let’s be honest, money makes science. An expanding field implies money is flowing into it. No one invested in Mendel’s pea experiment or Darwin’s theory back in the 19th century, but the Right Reverend was supported by his church and Charles was born into money (Google the Darwin-Wedgwood family). Biology has gotten even more expensive – I’m literally spending thousands of dollars for an unassuming vial of liquid. If rich people can’t be enticed (by the likes of aging/cancer research), the research must be profitable, e.g., protein drug design or MRI analysis software (that costs a lot!).
A mature pipeline for validation. Science is for discovery. Engineering comes later. When/if you discover something through your silicon-based model (read: dry lab), you need carbon-based validation (read: wet lab). If you discover a new mechanism for cells to communicate, you have to start ordering petri dishes. From my experience, I say you should expect years of wet lab work. But after one year, your AI model would be prehistorical and you need to sit down and have coffee for four more years to finish validation and get your results published, if lucky. If not, you would have to pat yourself on the shoulder for proving the model wrong with the slight cost of a few years and a few hundred grands.