Largely agree with the thesis here, worth noting that transcription factors (intrinsically disordered, historically considered “undruggable”) may fall under the same category of GPCRs where we really are limited by drug candidates and AI could help. Also many isolated cases, consider suzetrigine, breakthrough pain medication where a large part of the difficulty was the excruciatingly difficult med-chem optimization. Actually now that I think about it, what percent of the proteome is currently considered druggable? There’s a plausible world where we have many hypotheses and possible drugs, but they’re systematically concentrated around kinase inhibitors or whatever, so there is tons of low hanging fruit in the biology sense that better AI drug design could help access
I still think that getting the right dosage clinical effect etc will involve in-human experimentation. Look at GLP-1 agonists, which took time to perfect for obesity and there were iterations between clinical trials and peptide chemistry improvements.
Yeah this is why I largely agree! Just worth noting that one driver of clinical trial costs is weak effect sizes; a me-too statin with a number needed to treat of 1000 will cost a ton to test, whereas a clinical trial for penicillin mostly consists of observing “hey my patients are all better instead of immediately dead”. So it’s possible that opening up whole new classes of drugs will cut down on costs that way.
The problem with generative AI is that it often gets simple things wrong. It can generate a picture of a dog with five legs. The geometry of a drug molecule is very critical! What if the generative model generated a target molecule with the wrong number of geometric features like the number of atoms in a molecule ? Not good 😊
Sure. But what makes Ruxandra's critique here potent is that it holds if the models do exactly what we want. Critiquing where the models are today usually gets dismissed because they get much better every 6 months. The value of this argument is that it just skips over that problem and gets to the bigger one, which is if models are *perfect* speed-to-discovery might 1) stay bad or 2) get worse if we don't fix regulation.
I think you could engineer around that. Could either run a script to validate the output of the AI, or give it a molecule validation tool to check it's own output. I'm not a chemist, but some googling suggests RDkit might be able to do that validation. Correct me if I'm wrong tho.
I was under the impression that Big Pharma has enough new potential drugs to last decades, and the main bottlenecks are the clinical trials you reference. Does anyone know if this is true?
However, one could make the argument that if we were better at predicting which drugs would work before testing them in humans, this would become less of a bottleneck.
Worth reading a recent piece by Kim Branson (SVP AI GSK) entitled “Right for the wrong reasons)-available via Linked-In on the inherent limitations of AI in biology. Derek Lowe has also been writing about this very subject for some time.
The utility of AI in drug discovery gets all the accolades (and the investment $) but I’m as interested in AIs use in clinical development and its potential use in both speeding up trials and patient selection.
But not everything is in conflict. More new hypotheses almost certainly *is* a good thing, and is more likely to lead to more “better” hypotheses, not fewer.
Again, you are conflating what is more/most important with what is helpful.
You have no evidence that having more hypotheses is causal for having fewer good/better hypotheses. And again, no logic for why it would be true.
You have no counterfactual re: your claim that more hypotheses is unhelpful, let alone counterproductive
We are in complete agreement - or at least I agree with Scannell that the number of hypotheses does not come close to explaining the decrease in drug productivity, and that regulation / government bureaucracy is likely the biggest culprit.
My assumption was that the increasing cost of drug discovery primarily due to reducing the low-hanging fruit. Eg. you can't discover tylenol today because tylenol has already been discovered
> However, existing AI systems are limited because they are not trained on the kinds of causal, dynamic, multi-scale biological data that would be required to replace or approximate in-human testing.
This is a key observation to understand the value of AI in many fields of science. We want to take data and theory from one scale (or domain) and apply it to another scale (or domain) to obtain new insights.
It's also important to realize that available multi-scale scientific data are often sparse and relatively limited in quantity.
Combining these two points, it's clear that the recent LLM paradigm is not a natural fit for these problems. LLMs rely on massive data (e.g. the entirety of human writing/literature + the web) and were originally undertaken as a research project. AI researchers wanted to see how much deep neural networks can learn, so they used architectures like transformers, which scale well to larger datasets, but do not have strong inductive biases (their underlying assumption is that any piece of data can relate to any other piece of data, and the nature of those relationships can be learned from the data itself). When data are limited and sparse, this scheme will not work well.
The more general idea of "AI" can still add value here. But it will require returning to inductive bias as a tool to supplement limited data with domain knowledge (as was done through science and machine learning up until the recent paradigm shift). Adapting these techniques to retain the benefits of the newer unsupervised pretraining methods is currently a burgeoning effort at best. (And, of course, more data will still always be better, hence the importance of unlocking clinical trial data.)
Agree, I am still waiting for the truly groundbreaking AI results. So far, I have seen mostly highly sophisticated algorithms that help improve things and make them more efficient, but they do not actually create truly novel insights.
I generally agree with your take here. People are way overindexing on what AI can do in biology. AI is a great tool and for certain problems it's the best approach we have, but it's nowhere near the silver bullet that can just magically generate answers to all our biological questions.
Check out this company’s tech. #HYFT patented and embedded into their software.
A couple of comments about your remarks
- MindWalk’s LensAI platform includes an immunogenicity screening module that flags risk of anti-drug antibodies (ADA) very early, before expensive in vivo or clinical testing.
- This lets them triage out risky candidates or guide sequence modifications (engineering) to make molecules safer before going to animal or human testing, reducing clinical risk, and know if they are “human friendly”.
- The LensAI platform (powered by HYFT®) can predict, from sequence alone (without needing structural data), where antibodies or immune recognition might bind — “epitope mapping.
- This broad generalizability means they can apply their predictions even to novel proteins or therapeutic targets, not just ones they’ve seen before (untrained targets).
- MindWalk combines multi-omics data (sequence, structure, functional data, literature) into a unified graph, which helps their AI “reason” in biologically meaningful ways.
- Their HYFT® technology extracts universal “fingerprints” (patterns) from biological sequences that are more meaningful than naive alignment-based methods. These fingerprints help them assess similarity to human proteins (“humanness”), immunogenic risk, and more.
- They have a platform for in silico humanization: modifying non-human sequences (e.g., mouse antibodies) to make them more “human-like” and reduce immunogenicity risk, while retaining functionality.
- Their method screens against the entire human proteome fast (on the scale of under a minute per candidate) to assess how “human” a candidate is, which is better than only partial or shallow assessments.
As a researcher working on a paper on generative AI for drug discovery I am, obviously, biased, but: Recent methodological advances, such as geometric inductive bias and score-based diffusion, have improved the quality of the candidates. More advances, such as more principled iteration of candidate generation and molecular dynamics simulation, could, I believe, improve this further.
But, I completely agree with you that the primary sticking points are in the real world validation.
Indeed, at my department a professor related a story: he had a PhD student with whom it was agreed they'd develop some AI drug discovery scheme that would be validated by a collaborating Biology department with a wet lab (so not even human experiments, just in vitro validation). It ultimately took so much time for all of the necessary materials to be acquired and procedures to be followed that the student was not able to get a paper accepted in time to get through a mid-doctoral evaluation and had to quit the PhD.
Since then the professor only supervises purely computational doctoral projects.
I’ve thought about the following scenario: What if AI started making suggestions like: “Feed this specific poison to a thousand frogs. Most will die, but breed the ones that survive, and repeat for 30 or 40 generations. Eventually they might evolve a compound that could help treat this disease.”
As outlined in my message to you, the translational gap problem, as described here is fundamental to achieving better, more productive drug discovery. However, that is only part of the problem in applying AI to drug discovery. There is a need to revisit and alter the reductionist drug discovery paradigm and fill the data gap for drug discovery to deliver safe and effective drugs for serious unmet medical conditions. AI can have an impact here if it addresses (together with human models), both the translational gap as well as paradigm-related problems.
When you begin to actually think what the body needs to be healthy… it’s not poisonous and useless drugs but nutrition. You will shift to a real healthy life. AI it’s programmed by the same companies that want to keep you sick for profits.
Largely agree with the thesis here, worth noting that transcription factors (intrinsically disordered, historically considered “undruggable”) may fall under the same category of GPCRs where we really are limited by drug candidates and AI could help. Also many isolated cases, consider suzetrigine, breakthrough pain medication where a large part of the difficulty was the excruciatingly difficult med-chem optimization. Actually now that I think about it, what percent of the proteome is currently considered druggable? There’s a plausible world where we have many hypotheses and possible drugs, but they’re systematically concentrated around kinase inhibitors or whatever, so there is tons of low hanging fruit in the biology sense that better AI drug design could help access
I still think that getting the right dosage clinical effect etc will involve in-human experimentation. Look at GLP-1 agonists, which took time to perfect for obesity and there were iterations between clinical trials and peptide chemistry improvements.
100%
Yeah this is why I largely agree! Just worth noting that one driver of clinical trial costs is weak effect sizes; a me-too statin with a number needed to treat of 1000 will cost a ton to test, whereas a clinical trial for penicillin mostly consists of observing “hey my patients are all better instead of immediately dead”. So it’s possible that opening up whole new classes of drugs will cut down on costs that way.
The problem with generative AI is that it often gets simple things wrong. It can generate a picture of a dog with five legs. The geometry of a drug molecule is very critical! What if the generative model generated a target molecule with the wrong number of geometric features like the number of atoms in a molecule ? Not good 😊
Well new molecules could be checked by humans I guess!
Sure. But what makes Ruxandra's critique here potent is that it holds if the models do exactly what we want. Critiquing where the models are today usually gets dismissed because they get much better every 6 months. The value of this argument is that it just skips over that problem and gets to the bigger one, which is if models are *perfect* speed-to-discovery might 1) stay bad or 2) get worse if we don't fix regulation.
I think you could engineer around that. Could either run a script to validate the output of the AI, or give it a molecule validation tool to check it's own output. I'm not a chemist, but some googling suggests RDkit might be able to do that validation. Correct me if I'm wrong tho.
I was under the impression that Big Pharma has enough new potential drugs to last decades, and the main bottlenecks are the clinical trials you reference. Does anyone know if this is true?
Yes, a large bottleneck is clinical testing.
However, one could make the argument that if we were better at predicting which drugs would work before testing them in humans, this would become less of a bottleneck.
Worth reading a recent piece by Kim Branson (SVP AI GSK) entitled “Right for the wrong reasons)-available via Linked-In on the inherent limitations of AI in biology. Derek Lowe has also been writing about this very subject for some time.
Thank you! I have been following Derek for a while. Will check out the Kim post.
The utility of AI in drug discovery gets all the accolades (and the investment $) but I’m as interested in AIs use in clinical development and its potential use in both speeding up trials and patient selection.
I agree, I would like to see discussed more as well.
“We do not need more hypotheses, we need better ones”
Sorry, but this is just a false assertion.
We need both.
Just because the former does not *guarantee* the latter, you have no evidence - and precious little logic - that the former will reduce the latter.
P.S. But if you wanna argue, as Scannell does, that the bigger issue is all the regulation, you will get no disagreement from me.
There's no point in generating new hypotheses if we can't test them.
And obvsly I say we need to generate better ones and that's good.
But not everything is in conflict. More new hypotheses almost certainly *is* a good thing, and is more likely to lead to more “better” hypotheses, not fewer.
I mean, no. We've had more scientists and more papers and they haven't led to better hypotheses. We've literally ran the experiment.
Again, you are conflating what is more/most important with what is helpful.
You have no evidence that having more hypotheses is causal for having fewer good/better hypotheses. And again, no logic for why it would be true.
You have no counterfactual re: your claim that more hypotheses is unhelpful, let alone counterproductive
We are in complete agreement - or at least I agree with Scannell that the number of hypotheses does not come close to explaining the decrease in drug productivity, and that regulation / government bureaucracy is likely the biggest culprit.
Do you always need the whole human for drug trials?
ahah what do you have in mind here?
Testing on individual organs or organoids.
My assumption was that the increasing cost of drug discovery primarily due to reducing the low-hanging fruit. Eg. you can't discover tylenol today because tylenol has already been discovered
> However, existing AI systems are limited because they are not trained on the kinds of causal, dynamic, multi-scale biological data that would be required to replace or approximate in-human testing.
This is a key observation to understand the value of AI in many fields of science. We want to take data and theory from one scale (or domain) and apply it to another scale (or domain) to obtain new insights.
It's also important to realize that available multi-scale scientific data are often sparse and relatively limited in quantity.
Combining these two points, it's clear that the recent LLM paradigm is not a natural fit for these problems. LLMs rely on massive data (e.g. the entirety of human writing/literature + the web) and were originally undertaken as a research project. AI researchers wanted to see how much deep neural networks can learn, so they used architectures like transformers, which scale well to larger datasets, but do not have strong inductive biases (their underlying assumption is that any piece of data can relate to any other piece of data, and the nature of those relationships can be learned from the data itself). When data are limited and sparse, this scheme will not work well.
The more general idea of "AI" can still add value here. But it will require returning to inductive bias as a tool to supplement limited data with domain knowledge (as was done through science and machine learning up until the recent paradigm shift). Adapting these techniques to retain the benefits of the newer unsupervised pretraining methods is currently a burgeoning effort at best. (And, of course, more data will still always be better, hence the importance of unlocking clinical trial data.)
Agree, I am still waiting for the truly groundbreaking AI results. So far, I have seen mostly highly sophisticated algorithms that help improve things and make them more efficient, but they do not actually create truly novel insights.
I generally agree with your take here. People are way overindexing on what AI can do in biology. AI is a great tool and for certain problems it's the best approach we have, but it's nowhere near the silver bullet that can just magically generate answers to all our biological questions.
See also: https://blog.genesmindsmachines.com/p/we-still-cant-predict-much-of-anything
amen to that!
Check out this company’s tech. #HYFT patented and embedded into their software.
A couple of comments about your remarks
- MindWalk’s LensAI platform includes an immunogenicity screening module that flags risk of anti-drug antibodies (ADA) very early, before expensive in vivo or clinical testing.
- This lets them triage out risky candidates or guide sequence modifications (engineering) to make molecules safer before going to animal or human testing, reducing clinical risk, and know if they are “human friendly”.
- The LensAI platform (powered by HYFT®) can predict, from sequence alone (without needing structural data), where antibodies or immune recognition might bind — “epitope mapping.
- This broad generalizability means they can apply their predictions even to novel proteins or therapeutic targets, not just ones they’ve seen before (untrained targets).
- MindWalk combines multi-omics data (sequence, structure, functional data, literature) into a unified graph, which helps their AI “reason” in biologically meaningful ways.
- Their HYFT® technology extracts universal “fingerprints” (patterns) from biological sequences that are more meaningful than naive alignment-based methods. These fingerprints help them assess similarity to human proteins (“humanness”), immunogenic risk, and more.
- They have a platform for in silico humanization: modifying non-human sequences (e.g., mouse antibodies) to make them more “human-like” and reduce immunogenicity risk, while retaining functionality.
- Their method screens against the entire human proteome fast (on the scale of under a minute per candidate) to assess how “human” a candidate is, which is better than only partial or shallow assessments.
Hope it helps
As a researcher working on a paper on generative AI for drug discovery I am, obviously, biased, but: Recent methodological advances, such as geometric inductive bias and score-based diffusion, have improved the quality of the candidates. More advances, such as more principled iteration of candidate generation and molecular dynamics simulation, could, I believe, improve this further.
But, I completely agree with you that the primary sticking points are in the real world validation.
Indeed, at my department a professor related a story: he had a PhD student with whom it was agreed they'd develop some AI drug discovery scheme that would be validated by a collaborating Biology department with a wet lab (so not even human experiments, just in vitro validation). It ultimately took so much time for all of the necessary materials to be acquired and procedures to be followed that the student was not able to get a paper accepted in time to get through a mid-doctoral evaluation and had to quit the PhD.
Since then the professor only supervises purely computational doctoral projects.
I’ve thought about the following scenario: What if AI started making suggestions like: “Feed this specific poison to a thousand frogs. Most will die, but breed the ones that survive, and repeat for 30 or 40 generations. Eventually they might evolve a compound that could help treat this disease.”
Couldn't agree more. That distinction betwen hypothesis quantity and quality, and the call for better predictive validity, is incredibly insightful.
As outlined in my message to you, the translational gap problem, as described here is fundamental to achieving better, more productive drug discovery. However, that is only part of the problem in applying AI to drug discovery. There is a need to revisit and alter the reductionist drug discovery paradigm and fill the data gap for drug discovery to deliver safe and effective drugs for serious unmet medical conditions. AI can have an impact here if it addresses (together with human models), both the translational gap as well as paradigm-related problems.
When you begin to actually think what the body needs to be healthy… it’s not poisonous and useless drugs but nutrition. You will shift to a real healthy life. AI it’s programmed by the same companies that want to keep you sick for profits.