AI x Bio
In the lab I’ve been working on developing new technologies to build around my core discovery engine. This involves very fast hypothesis generation, experiments, and iteration. LLMS like Gemini and ChatGPT have generally been worse than useless in assisting, other than helping to plan initial stabs. It’s shocking in practice how useless they are, and I use the most expensive, current models. New models keep getting released claiming “better than PhD scientist” benchmarks. There is no way the benchmarks are measuring scientific ability.
All over I see ridiculous amounts of money pouring into AI x Bio, with claims from leading figures saying things like “I think one day maybe we can cure all disease with the help of AI.”
I hear rumors of senior management in companies entirely delegating experimental design to AI which is, and I cannot stress this strongly enough, completely insane. I don’t want to come off as a hater as I do use AI a lot, and it is incredible for some things, but it is not good at science.
What do I mean?
An example: I was struggling with a difficult reaction. It worked well 60% of the time, and failed 40%, sometimes between different days, sometimes between samples on the same day. That’s not good enough to use when I need it at 95%+. I had run through five experiments assessing various hypotheses and ticking them off as irrelevant. I’d bounce the experiment and results off an LLM along with my hypothesis for what should come next. Whether it agrees or not appears to be basically meaningless in its predictive validity.
One hypothesis was that I had subtly varied whether I added reagent 1 into a frozen aliquot of reagent 2, versus a thawed aliquot of reagent 2, and that switching from thawed to frozen was impacting the reaction in idiosyncratic ways. ChatGPT agreed!
Going into frozen was way better, and the local concentration/degradation effects were definitely the explanation.
“Wait, ChatGPT!” I said. “My hypothesis was the opposite of the explanation you just gave.”
“Oh yes! My mistake! Going into thawed was way better and the local concentration/degradation effects were definitely the explanation.”
I test it out in the lab and it makes zero difference. Obviously, it is troubling that it makes a confident, equally compelling explanation in either direction. Relevant context, if searched well in the literature, surely exists to justify one position or dismiss the question.
Just today I found out some growth media I need was backordered. I asked ChatGPT to find me an alternative, compatible with two classic yeast genes, Met15, and Lys2. To use Met15, canonically you need to drop out both methionine and cysteine; as yeast can turn cysteine into methionine. I wasn’t sure if this was just legacy knowledge and I could get away with just dropping out Methionine, rather than both, so I asked ChatGPT.
“Wait ChatGPT, the paper you linked doesn’t show that at all, and it says I have to drop out both.”
“Ah yes, I made it up” says ChatGPT. If I had followed that advice I would have wasted about 2 weeks, and >$1,000.
I’ve had many examples like this. Even stupid ones, like I accidentally used SOC media (which has glucose) in a competent e.coli preparation, instead of SOB (no glucose equivalent). I said oops what do you think Gemini, is this fine? No, it is not fine. You should dump them and start over. The e coli are stressed out now, because of the sugar. They won’t work. And even if they do, you won’t be confident. I get annoyed, say: “you have no idea what you’re talking about Gemini.” I test it out; it’s fine. Transformation efficiency is unharmed. I go back to gloat. It retorts “Ok maybe they worked, but they’re still stressed out, you better do everything by the book or it won’t work.”
This tweet embodies the problem, in a funny sort of way.
Beyond that; every basic protocol request comes with five recommendations for slightly different ways to do it. “I’m going to transform a species of yeast I haven’t transformed before; can I do it the same way as I’ve done cerevisiae?” “No, I wouldn’t assume that, it could be more heat sensitive, need more DNA, here’s what I would do as a backup, and another backup, and another backup.”
I decide I’m gonna ignore all that and do it my way; works fine. If I had taken the advice, it would have turned an experiment that takes up ¼ of my day into my whole day.
Requests for a sequence are often plain wrong. It returns a sequence which looks plausible, but when you look into the original paper there is no sequence included. The LLM just made it up. If I had accepted that sequence; it would take weeks to months and thousands of dollars to realize. You have to really press the LLM to admit it made it up (and it’s unsatisfying when you get it to admit that it made it up because you know it’s saying that with the same confidence as its hallucination).
I have a dozen years at the bench under my belt, and I am the type of scientist who says: “there’s no way that could possibly matter” and ignores advice. That is how you become an efficient scientist. Anyone trying to learn biology in this way will be a bad scientist and turn every experiment into a big, complicated mess.
There’s a lot of hype around AI / bio, but man, it does not look at all justified to me. These systems are trained on existing information and seem confined to advanced forms of interpolation and pattern matching.
I’d posit it’s an inherent property of a LLM that it can’t be useful for anything which requires a new way of thinking about a problem. I’m not the first to make this point but LLMs have been trained on every scientific paper ever and have generated no major breakthroughs. Real biologists are trained on 0.1% of papers and make breakthroughs.
Idea generation in bio (and science at large) relies on forming a mental model of the system you are working on. What measurable change happens if my mental model is correct and I press my system this way? What else should happen? What happens if there’s a simpler explanation than my model? Where is the literature inconsistent and how do the inconsistencies affect the prevailing model? 1/12x data points I generated isn’t as expected, what possibilities for modifications to the model would make it fit? And how could I evaluate that?
There is no internal world to an LLM. It can rearrange and regurgitate but it doesn’t form an imaginary model of a system and probe it against all information. It’s a fundamentally different task than searching for and interpolating between pieces of information, which AI is spectacular at.
While holes like hallucination can be plugged (maybe?), I do not think the structure of knowledge and computation on it is an issue that can be overcome. While it might look like it’s thinking, it is not. It is the core difference between human intelligence and artificial intelligence.
Now, different bespoke models for certain tasks may be genuine advance. Alphafold 3, and new protein binder models are useful. But my guess is they do suffer the same limitations as the LLMs, in that they don’t create internal models of the fundamental problem, they just interpolate what looks closest to the answer based on what is available.
It’s rare to see examples of protein binder models coming up with binders to GPCRs that don’t already have one. To the best of my knowledge folding models still cannot touch directed evolution in producing high performance variants. Maybe these capacities exist internally to companies, maybe they pretend it does to try to capture more money. “Foundation models” are less effective than simple linear methods which have been around for decades. These papers are getting published in the most prestigious journals, and they just don’t work.
It all seems a little scammy.
The one area where AI is near miraculous for me is data analysis. Being able to describe a massive dataset, measured in 3 different ways, and in minutes get code which integrates/analyzes it, it is nothing short of amazing. It’s opened new experimental doors for me and makes being a biologist with high-throughput chops a lot more fun. Because of how fast I can iterate and the sheer output of AI code I can do the discovery work of a lab by myself. Of course, you have to build in internal controls to your experiment and make sure those look good because AI still does make some weird judgement calls.
It is decent at searching literature for relevant papers; but still misses obvious ones that you describe precisely.
On the flip side of my skepticism, with a ton of data there is even more opportunity for interpolation, and interpolation could be useful for uncovering hidden connections. It does feel like there is some opportunity here, but I imagine it’s going to look weirder than we think right now. As a coincidence, uncovering high throughput emergence of function of genes in ways that looks bizarre and inexplicable currently (what I do) may be a rich source for this. For now, I’ll build in the data collection, remaining skeptical. Happy and grateful to chat with any AI people who have a take for me.
Now; as a last note. The fact that I see the biggest problems in the only field that I am an expert in; does not escape me. I would guess the strengths and weaknesses are the same across fields, but I don’t know how that matters. Maybe your AI lawyer is making the exact same categorical mistakes.
Does that mean we are in a bubble whose simultaneous pop is going to violently tear down the economy and eviscerate 401ks?
Haha I sure hope not! I’ll leave that to the professionals.