This is thoughtful piece. I couldn’t agree more. Biology has never lacked hypotheses, but has lacked efficient ways to distinguish signal from noise. AI’s greatest contribution may indeed be improving the quality of learning and decision-making, rather than serving as a magic wand for drug discovery.
Really great summary and captures a lot of the power, but also the drawbacks of AI as it is (emphasis on “as is”).
There’s no magic barrier for AI to start solving more complex problems. If mechanism → drug is what we can get by throwing compute at a formal system, then we should stop only throwing compute at formal systems and hope models find answers outside of these systems. Brute force statistical pattern matching won’t get you to true understanding. And (un)fortunately, biology demands deep understanding, and not simply getting to a pre-defined answer quickly.
The proposed solution is the right play within the current paradigm, and I genuinely hope it yields results, ad infinitum. But something in me wants an elegant solution, from the ground up.
Really resonant piece, and "right locks, not better keys" is indeed the crux of it. The part I'd broaden is that the diseases with the least mechanistic understanding tend to be the most human-specific ones, brain and behavior above all. That's exactly where our data is thinnest, our animal models weakest, and our constraints tightest. It's the hard problem of psychiatry.
Where I'd push a little further is on where the priors for the "right locks" come from when measurement is this expensive. At Sensorium we work in psychiatry and neurology, and we keep coming back to the fact that evolution and human use have already run a massive, if messy, screen for us. Natural products and the ethnopharmacological record encode bioactivity that was selected in living systems and, in a lot of cases, observed directly in humans. That's not a magic wand either. It's super noisy, confounded with placebo, and it still needs exactly the deep phenotyping and causal work Daphne is describing to turn a compound into a mechanism. But it's real human-relevant signal in the domains where virtual cells and rodent models struggle most.
So I read this less as "AI vs. biology" and more as a call to be honest about what our objective function actually is. Whether the prior comes from perturbation atlases, from clinical data, or from a molecule people have used for a thousand years, the hard part is the same. You have to connect it to a mechanism whose modification changes disease in a patient. The tools that win will be the ones pointed at that problem, not the ones with the cleanest benchmark.
Daphne, you rightly point out we are missing "systems-level temporal interactions". We have to stop treating biology like a software problem and start treating the wet-lab like a continuous-time data center. This means upgrading the physical hardware to actually read the continuous temporal derivative of the living system
Causal mechanism discovery is the right approach to the current bottlenecks. However, the tools to do so need to be carefully selected.
LLMs (and AI agents) are great at filling knowledge gaps in our current understanding, connecting the dots we have already measured, interpreted, understood and documented. Yet, they can’t jump and propose truly novel hypothesis. So in using them as our discovery engine we risk to be confined to what we know.
To make true causal discoveries about human physiology and extend beyond our current knowledge boundaries we need: I) data (as clearly stated in the article); II) human creativity, curiosity and intelligence equipped with III) causal AI to analyze the data at scale. Those three things combined is IMO the best way forward in the pursue of the causal graph of human physiology - at least a subset thereof.
I'd put more weight on the modality point than the target one. Every real step change in what we could drug came from a new modality, biologics, then siRNA, then editing, and none of those arrived by designing a better molecule with the tools we already had. Inventing a modality is a wet lab problem, and that's where the least AI effort is pointed right now.
Thought-provoking read. Drug discovery has never had magic wands only better tools. AI can accelerate research and improve decision-making, but scientific rigor, experimental validation, and patient-focused development remain the real drivers of success.
Are we looking at targets or monetary reward? As VCs you’re skewed.
Rare diseases have a smaller market but rare disease victims lifetime medical casts are 20 to 30x that of the normal population. This category also has a minimal biological testing component ( orphan drug status) equaling a fraction of time and money. If you want to help humans, save medicare and use AI/ genetics, fund rare disease abatement.
Great piece. Lantern Pharmaceuticals is the company to watch in this space with their RADR AI platform targeting improvements in all these areas specific to oncology today. They are spinning off their AI tool, withZeta as a separate entity now so they can expand their realm of expertise over time and unlock the value of the platform they built for their own drug discovery.
Rare Intriguing yet unhyped perspective on AI in human disease. One corollary to Daphne's points - especially as pertaining to biomarkers for rapid detection of drug efficacy - is use of mechanistic AI models for identification of biomarkers for early detection. Treating or curing disease in the subclinical phase is safer and more effective, but only blunt early detection tools exist for the majority of diseases.
This is the distinction I wish more AI drug-discovery conversations made. Finding targets or generating candidates faster can improve throughput without changing the biological reasons most drugs fail once they reach people. I’d be curious what evidence would convince you that AI is actually improving success rates, rather than just moving more candidates through the pipeline.
Daphne… this is nicely argued and I of course agree with you that nobody would turn down a complete causal model of human disease. But three of the important successes of the last decade, two Mendelian and one in cancer, didn't arrive that way. Each came from a much simpler question: why is this patient doing better than they should be?
In SMA, the question was why humans carry a near-copy of the missing gene that is almost, but not quite, functional. The problem shifts from replacing a gene, to persuading a spare copy the patient already has to splice properly, which an oligonucleotide can do.
In sickle cell, the question was why some people with the same mutation are barely ill. Janet Watson noticed the seed of the answer in 1948 - not switching off fetal haemoglobin. That turns fixing the adult gene in every cell into releasing a brake on a gene that already works.
In colorectal cancer, an early anti-PD-1 trial produced a single durable complete response, in a disease everyone had written off for immunotherapy. Diaz and Vogelstein asked what was different about him rather than regressing him away as statistical noise. He was mismatch-repair deficient, so his tumour carried a massive mutational burden. The predictor of response turned out not to be the organ, and the test was one standard Immunohistochemistry pathology was already running.
None of these were low-hanging fruits and none of them built a model of the cell. They built one causal chain from a human variant to a human outcome, narrow enough for the talented team to attack. The perturbation data was already collected and came with clinical outcomes attached. No huge new data generation required
There are counterexamples, KRAS being the obvious one, where the unlock was pure chemistry. But the cases above leave me wondering whether part of the key is simply refusing to treat the outlier in a failed trial as noise, which is a good deal cheaper than the alternative. Parker
Drug to patient is low for origin of failure? Don’t most clinical trials fail because of operational issues like recruiting enough patients or choosing the right endpoints? Iirc something like 60%…
Daphne is absolutely correct and hits the nail on its head! Most of the work in drug discovery is finding the molecule that is supposed to cure the symptoms, without understanding what the cause is. E.g. ROCHE agreed to pay $2.8 Billion for a future medication (Zelebesiran) based on RNAi that will silence a gene that helps generate Angiotensin II. This is without understanding if knocking out this Angiotensin will help to reduce mortality and morbidity of CVD patients. Just a hint: hypotension is as dangerous as Hypertension.
Instead of extinguishing the fire they fight the smoke. This is done in many medications because of lack of understanding the cause of the disease.
This is thoughtful piece. I couldn’t agree more. Biology has never lacked hypotheses, but has lacked efficient ways to distinguish signal from noise. AI’s greatest contribution may indeed be improving the quality of learning and decision-making, rather than serving as a magic wand for drug discovery.
Really great summary and captures a lot of the power, but also the drawbacks of AI as it is (emphasis on “as is”).
There’s no magic barrier for AI to start solving more complex problems. If mechanism → drug is what we can get by throwing compute at a formal system, then we should stop only throwing compute at formal systems and hope models find answers outside of these systems. Brute force statistical pattern matching won’t get you to true understanding. And (un)fortunately, biology demands deep understanding, and not simply getting to a pre-defined answer quickly.
The proposed solution is the right play within the current paradigm, and I genuinely hope it yields results, ad infinitum. But something in me wants an elegant solution, from the ground up.
Really resonant piece, and "right locks, not better keys" is indeed the crux of it. The part I'd broaden is that the diseases with the least mechanistic understanding tend to be the most human-specific ones, brain and behavior above all. That's exactly where our data is thinnest, our animal models weakest, and our constraints tightest. It's the hard problem of psychiatry.
Where I'd push a little further is on where the priors for the "right locks" come from when measurement is this expensive. At Sensorium we work in psychiatry and neurology, and we keep coming back to the fact that evolution and human use have already run a massive, if messy, screen for us. Natural products and the ethnopharmacological record encode bioactivity that was selected in living systems and, in a lot of cases, observed directly in humans. That's not a magic wand either. It's super noisy, confounded with placebo, and it still needs exactly the deep phenotyping and causal work Daphne is describing to turn a compound into a mechanism. But it's real human-relevant signal in the domains where virtual cells and rodent models struggle most.
So I read this less as "AI vs. biology" and more as a call to be honest about what our objective function actually is. Whether the prior comes from perturbation atlases, from clinical data, or from a molecule people have used for a thousand years, the hard part is the same. You have to connect it to a mechanism whose modification changes disease in a patient. The tools that win will be the ones pointed at that problem, not the ones with the cleanest benchmark.
Daphne, you rightly point out we are missing "systems-level temporal interactions". We have to stop treating biology like a software problem and start treating the wet-lab like a continuous-time data center. This means upgrading the physical hardware to actually read the continuous temporal derivative of the living system
Causal mechanism discovery is the right approach to the current bottlenecks. However, the tools to do so need to be carefully selected.
LLMs (and AI agents) are great at filling knowledge gaps in our current understanding, connecting the dots we have already measured, interpreted, understood and documented. Yet, they can’t jump and propose truly novel hypothesis. So in using them as our discovery engine we risk to be confined to what we know.
To make true causal discoveries about human physiology and extend beyond our current knowledge boundaries we need: I) data (as clearly stated in the article); II) human creativity, curiosity and intelligence equipped with III) causal AI to analyze the data at scale. Those three things combined is IMO the best way forward in the pursue of the causal graph of human physiology - at least a subset thereof.
I'd put more weight on the modality point than the target one. Every real step change in what we could drug came from a new modality, biologics, then siRNA, then editing, and none of those arrived by designing a better molecule with the tools we already had. Inventing a modality is a wet lab problem, and that's where the least AI effort is pointed right now.
Thought-provoking read. Drug discovery has never had magic wands only better tools. AI can accelerate research and improve decision-making, but scientific rigor, experimental validation, and patient-focused development remain the real drivers of success.
Are we looking at targets or monetary reward? As VCs you’re skewed.
Rare diseases have a smaller market but rare disease victims lifetime medical casts are 20 to 30x that of the normal population. This category also has a minimal biological testing component ( orphan drug status) equaling a fraction of time and money. If you want to help humans, save medicare and use AI/ genetics, fund rare disease abatement.
Great piece. Lantern Pharmaceuticals is the company to watch in this space with their RADR AI platform targeting improvements in all these areas specific to oncology today. They are spinning off their AI tool, withZeta as a separate entity now so they can expand their realm of expertise over time and unlock the value of the platform they built for their own drug discovery.
50% $$ in the Ads world is lost or waste
but how many percentage $$ get lost or waste in the bio R&D world?
Rare Intriguing yet unhyped perspective on AI in human disease. One corollary to Daphne's points - especially as pertaining to biomarkers for rapid detection of drug efficacy - is use of mechanistic AI models for identification of biomarkers for early detection. Treating or curing disease in the subclinical phase is safer and more effective, but only blunt early detection tools exist for the majority of diseases.
This is the distinction I wish more AI drug-discovery conversations made. Finding targets or generating candidates faster can improve throughput without changing the biological reasons most drugs fail once they reach people. I’d be curious what evidence would convince you that AI is actually improving success rates, rather than just moving more candidates through the pipeline.
Daphne… this is nicely argued and I of course agree with you that nobody would turn down a complete causal model of human disease. But three of the important successes of the last decade, two Mendelian and one in cancer, didn't arrive that way. Each came from a much simpler question: why is this patient doing better than they should be?
In SMA, the question was why humans carry a near-copy of the missing gene that is almost, but not quite, functional. The problem shifts from replacing a gene, to persuading a spare copy the patient already has to splice properly, which an oligonucleotide can do.
In sickle cell, the question was why some people with the same mutation are barely ill. Janet Watson noticed the seed of the answer in 1948 - not switching off fetal haemoglobin. That turns fixing the adult gene in every cell into releasing a brake on a gene that already works.
In colorectal cancer, an early anti-PD-1 trial produced a single durable complete response, in a disease everyone had written off for immunotherapy. Diaz and Vogelstein asked what was different about him rather than regressing him away as statistical noise. He was mismatch-repair deficient, so his tumour carried a massive mutational burden. The predictor of response turned out not to be the organ, and the test was one standard Immunohistochemistry pathology was already running.
None of these were low-hanging fruits and none of them built a model of the cell. They built one causal chain from a human variant to a human outcome, narrow enough for the talented team to attack. The perturbation data was already collected and came with clinical outcomes attached. No huge new data generation required
There are counterexamples, KRAS being the obvious one, where the unlock was pure chemistry. But the cases above leave me wondering whether part of the key is simply refusing to treat the outlier in a failed trial as noise, which is a good deal cheaper than the alternative. Parker
Drug to patient is low for origin of failure? Don’t most clinical trials fail because of operational issues like recruiting enough patients or choosing the right endpoints? Iirc something like 60%…
One thing is certain - human researchers have been and are far too slow in developing new drugs and medical cures on their own.
AI should be required on every team and it should be working 24 x 7, when humans are sleeping , eating and otherwise living their lives.
Daphne is absolutely correct and hits the nail on its head! Most of the work in drug discovery is finding the molecule that is supposed to cure the symptoms, without understanding what the cause is. E.g. ROCHE agreed to pay $2.8 Billion for a future medication (Zelebesiran) based on RNAi that will silence a gene that helps generate Angiotensin II. This is without understanding if knocking out this Angiotensin will help to reduce mortality and morbidity of CVD patients. Just a hint: hypotension is as dangerous as Hypertension.
Instead of extinguishing the fire they fight the smoke. This is done in many medications because of lack of understanding the cause of the disease.