Outstanding article! The quote which got me was the following: "We should absolutely use predictive models to do everything you can to maximize PoS and rank assets, prioritize programs, estimate rNPV, and decide which shots are worth taking." I would think that biomarkers, proteins, and DNA/RNA would be the three "pillars" to build biobank models around future in-silico drug development when measuring potential efficacy/safety data (non-scientist view, could be wrong?). One question, though - why do the models have to be "proprietary"? Wouldn't there be a meaningful societal benefit to such models being open-source with a permissive/dual licensing framework (?) that allows drug companies to utilize the predictive models to develop effacious drugs at lower costs? Again, kudos to you and your team - wishing you continued success in your entrepreneurial journey!
Flawed, misleading and a very long essay – persuasive in structure but fragile in epistemology.
This read is akin to many LLMs outputs; it sounds authoritative, exciting, and highly convincing but once scrutinized by domain expertise, the blind spots become pronounced and one realizes the essay is flawed.
While the read is tightly argued, data-rich, aspirational and optimistic, there are many reasons to think that many of the premises discussed here do not reflect the real world of biopharma R&D.
The main problem with this read is the misleading premise that suggests that preclinical drug candidates are drugs, “In an era of drug abundance, discovery will likely become less of a constraint than development”. It gets worse “This trend began well before AI. Over the past decade, the industry has become much better at generating plausible therapeutic assets.” Then, even worser “AI will accelerate this shift. Improvements in screening, structural biology…, and now AI are increasing the number of plausible therapeutic assets that can be generated.”
Just to clarify, preclinical drug candidates are NOT drugs. They are complex scientific laboratory endeavors that promise potential clinically meaningful benefits to human patients. Unfortunately, the majority of those endeavors, actually 90% of preclinical drug candidates, fail clinical development. Most of those endeavors never had the necessary quality to be developed into drugs. They were NEVER drugs! The contribution of failure in clinical development is as follows: lack of clinical efficacy (40-50%), unmanageable toxicity (~30%), and poor pharmacokinetics (PK) (~10–15%). Clinical operational issues like trial design flaws and recruitment & retention challenges have a lower impact (~5%). Thus, one simple reason to explain why the number of discovered drug candidates is larger than the number of approved drugs is because you can’t approve low quality or inappropriate drug candidates. Certainly, the FDA learned its lesson after the sulfanilamide disaster in the late 1930s.
AI is not new to biopharma R&D. It didn’t get embedded into R&D workflows in 2022 when ChatGPT was introduced or even in 2012 when AlexNet became popular for winning the ImageNet competition. AI (Symbolic AI, ML, ANN, +) has been integrated in biopharma R&D over many decades ago. Unfortunately, new “AI experts” and influencers have a narrowed view of AI as just being GenAI.
I disagree with the notion that clinical development is the bottleneck and discovery has been “solved”, or at least will be solved mainly thanks to AI. All the contrary, discovery has not been solved and it continues being the bottleneck in biopharma R&D.
The fundamental problem with AI in drug discovery is the misleading promotional marketing and hype. “AI will accelerate this shift. Improvements in screening…, and now AI are increasing the number of plausible therapeutic assets that can be generated.” Here, the focus is on speed and volume, not on value to patients. We not only need faster new treatment but also better new treatments. Our knowledge of human biology, physiology and medicine is maximally imprecise. And that’s the same knowledge that AI tools –including multi-modal foundational models– are being fed with.
In 2005, Harvard trained physician-scientist/mathematician Stanford professor John Ioannidis published one of the most influential, downloaded and controversial papers in the world: "Why Most Published Research Findings Are False". In 2025, 20 years later, John revised the title of his seminal paper. Now it is "Why Most Published Research Findings Are STILL False". If we hold ourselves to high standards of ethics and integrity, we know that there are three types of human data generators: the diligent, the sloppy and the fraudulent. And claimed research findings may be simply measurements of confirmation bias.
Gen AI is just making things worse...when not used responsibly. On the one hand, many AI models are being trained with useless big data. On the other hand, many AI models are generating, at scale, linguistically rich text & voice and visually impressive images. The content seems plausible but it's factually false.
The success of Isomorphic Lab's AI model AlphaFold2 (2024 Nobel Prize in Chemistry), in predicting (~90% accuracy) the 3D structure for monomeric proteins, was not only credited to superb domain expertise and the innovative transformer AI model. Publicly available (Protein Data Bank) high-quality "gold" standard X-Ray crystallography data were used to train AlphaFold2.
Here is the rub. Biopharma R&D is inherently inefficient (low responder rates). Not only due to the lengthy time to discover and develop a new drug and the high attrition rate in clinical development but also due to the low benefit that many patients get from many of those new drugs. On average, most approved drugs provide a clinical benefit to 30% to 60% of the patients they are prescribed for, though this number can be much lower or higher depending on the indication. Typical Patient "Success" (Responder Rate) for Oncology is 10%-30% and for infectious Disease is 80%–90%. So, if AI is only going to be used to deliver faster new medicines that only work for 10%-30% of patients, that’s a poor aspiration for AI and it should not be called a “major scientific achievement”.
The essay's framework assumes that the scarce resource is clinical execution, not individual biological understanding. If we don't yet understand why drugs work differently in different patients (genomics, epigenomics, microbiome, etc.), then simply running more trials faster does not solve the efficacy problem, it may just produce more failures faster.
Drug discovery and early development are the bottlenecks, not clinical development. A flawed therapeutic hypothesis –no matter how many drug candidates per target– will not survive the real-world pressure of clinical development. Small molecules often suffer from low bioavailability and poor target specificity and selectivity properties. Biologics like antibodies are highly specific but are not selective, leading to on-target off-organ toxicity. Oncology is a tough one. After several rounds of therapeutic interventions with promising FDA-approved drugs, patients get into remission, then they relapsed, and finally they become refractory to any therapy. Sooner or later, drug resistance mechanisms emerge in the clinic.
Biopharma needs to organize its translational teams to capture the clinical feedback loop back into discovery in real time (clinical response data → computational model → next-generation scaffold). Interdisciplinary research (AI, Quantum Mechanics, Nanotechnology and Biotechnology), not just AI, is increasingly playing a major role in support of drug product development and CMC by addressing formulation, bioavailability, and targeted delivery. Thus, offering differentiated approaches to de-risk pipeline drug candidates. I believe the next frontier is when responsible AI (not only GenAI or hyped AI) can predict not just molecular properties, but clinical translation risk: which optimized drug candidates will actually show higher efficacy/lower risk in specific patient subpopulations, and which will struggle against drug resistance mechanisms you have not yet seen in the clinic.
Now and in the future the saying “we discovered a high quality and differentiated drug” will continue to be a strong basis for durable company value. Again, I am not talking about a drug candidate or an approved drug that only works in a few patients.
Preclinical assets and discovery-stage companies will NOT face valuation pressure if they focus on value instead of AI-powered volume. Specifically, companies anywhere in the world must build the most irreplaceable pipeline of scientific discoveries including FICs, BICs, and BID assets. So, that global markets cannot find comparable assets anywhere else in the world. Currently, Chinese companies are better positioned to deliver that thanks to the convergence of scientific, technological, regulatory, cultural and demographic factors. For instance, Hengrui Pharma has a culture deeply rooted in two pillars. 1st, scientific rigor with 24 Class 1 marketed innovative drugs – the highest tier of pharmaceutical registration anywhere in the world. 2nd, disciplined long-term vision rather than just quick upfront/milestone payments. On top of that, Hengrui has the patient datasets and the clinical & regulatory velocity that Western companies can only dream about.
By the way, the essay notes the "rise of China as a drug discovery and development engine". This is a sharp observation, but it sits awkwardly with its overall framework. Above I provided more context. If discovery commoditizes globally but clinical development remains bottlenecked by Western regulatory requirements, the value chain may fragment in ways that favor neither Formation Bio nor its competitors.
The section on “Predictive model” is not evolution, it's marketing periodization. As explained above, to improve health outcomes now, we need to capture the clinical feedback loop back into discovery in real time now. Therapeutic hypothesis validation and clinical translational medicine are where the true value and risk lie. The companies that truly transform biopharma will challenge the assumptions that present the frontiers as uneven.
The FDA's skepticism of in silico clinical trials is not merely bureaucratic inertia; it reflects a century of learning that biological complexity outstrips computational model complexity.
The rest of the essay is well thought-out marketing branding.
*This replica to Uneven Frontiers was not created, enhanced or checked using AI
Interesting perspective. Success in healthcare and biotech increasingly depends on turning insights into execution. Innovation alone isn’t enough; adoption, evidence, and long-term impact are what create real value.
Outstanding article! The quote which got me was the following: "We should absolutely use predictive models to do everything you can to maximize PoS and rank assets, prioritize programs, estimate rNPV, and decide which shots are worth taking." I would think that biomarkers, proteins, and DNA/RNA would be the three "pillars" to build biobank models around future in-silico drug development when measuring potential efficacy/safety data (non-scientist view, could be wrong?). One question, though - why do the models have to be "proprietary"? Wouldn't there be a meaningful societal benefit to such models being open-source with a permissive/dual licensing framework (?) that allows drug companies to utilize the predictive models to develop effacious drugs at lower costs? Again, kudos to you and your team - wishing you continued success in your entrepreneurial journey!
Kudos to you Benjamine , great post
Replica to Uneven Frontiers @ a16z, 20Jun26:
Flawed, misleading and a very long essay – persuasive in structure but fragile in epistemology.
This read is akin to many LLMs outputs; it sounds authoritative, exciting, and highly convincing but once scrutinized by domain expertise, the blind spots become pronounced and one realizes the essay is flawed.
While the read is tightly argued, data-rich, aspirational and optimistic, there are many reasons to think that many of the premises discussed here do not reflect the real world of biopharma R&D.
The main problem with this read is the misleading premise that suggests that preclinical drug candidates are drugs, “In an era of drug abundance, discovery will likely become less of a constraint than development”. It gets worse “This trend began well before AI. Over the past decade, the industry has become much better at generating plausible therapeutic assets.” Then, even worser “AI will accelerate this shift. Improvements in screening, structural biology…, and now AI are increasing the number of plausible therapeutic assets that can be generated.”
Just to clarify, preclinical drug candidates are NOT drugs. They are complex scientific laboratory endeavors that promise potential clinically meaningful benefits to human patients. Unfortunately, the majority of those endeavors, actually 90% of preclinical drug candidates, fail clinical development. Most of those endeavors never had the necessary quality to be developed into drugs. They were NEVER drugs! The contribution of failure in clinical development is as follows: lack of clinical efficacy (40-50%), unmanageable toxicity (~30%), and poor pharmacokinetics (PK) (~10–15%). Clinical operational issues like trial design flaws and recruitment & retention challenges have a lower impact (~5%). Thus, one simple reason to explain why the number of discovered drug candidates is larger than the number of approved drugs is because you can’t approve low quality or inappropriate drug candidates. Certainly, the FDA learned its lesson after the sulfanilamide disaster in the late 1930s.
AI is not new to biopharma R&D. It didn’t get embedded into R&D workflows in 2022 when ChatGPT was introduced or even in 2012 when AlexNet became popular for winning the ImageNet competition. AI (Symbolic AI, ML, ANN, +) has been integrated in biopharma R&D over many decades ago. Unfortunately, new “AI experts” and influencers have a narrowed view of AI as just being GenAI.
I disagree with the notion that clinical development is the bottleneck and discovery has been “solved”, or at least will be solved mainly thanks to AI. All the contrary, discovery has not been solved and it continues being the bottleneck in biopharma R&D.
The fundamental problem with AI in drug discovery is the misleading promotional marketing and hype. “AI will accelerate this shift. Improvements in screening…, and now AI are increasing the number of plausible therapeutic assets that can be generated.” Here, the focus is on speed and volume, not on value to patients. We not only need faster new treatment but also better new treatments. Our knowledge of human biology, physiology and medicine is maximally imprecise. And that’s the same knowledge that AI tools –including multi-modal foundational models– are being fed with.
In 2005, Harvard trained physician-scientist/mathematician Stanford professor John Ioannidis published one of the most influential, downloaded and controversial papers in the world: "Why Most Published Research Findings Are False". In 2025, 20 years later, John revised the title of his seminal paper. Now it is "Why Most Published Research Findings Are STILL False". If we hold ourselves to high standards of ethics and integrity, we know that there are three types of human data generators: the diligent, the sloppy and the fraudulent. And claimed research findings may be simply measurements of confirmation bias.
Gen AI is just making things worse...when not used responsibly. On the one hand, many AI models are being trained with useless big data. On the other hand, many AI models are generating, at scale, linguistically rich text & voice and visually impressive images. The content seems plausible but it's factually false.
The success of Isomorphic Lab's AI model AlphaFold2 (2024 Nobel Prize in Chemistry), in predicting (~90% accuracy) the 3D structure for monomeric proteins, was not only credited to superb domain expertise and the innovative transformer AI model. Publicly available (Protein Data Bank) high-quality "gold" standard X-Ray crystallography data were used to train AlphaFold2.
Here is the rub. Biopharma R&D is inherently inefficient (low responder rates). Not only due to the lengthy time to discover and develop a new drug and the high attrition rate in clinical development but also due to the low benefit that many patients get from many of those new drugs. On average, most approved drugs provide a clinical benefit to 30% to 60% of the patients they are prescribed for, though this number can be much lower or higher depending on the indication. Typical Patient "Success" (Responder Rate) for Oncology is 10%-30% and for infectious Disease is 80%–90%. So, if AI is only going to be used to deliver faster new medicines that only work for 10%-30% of patients, that’s a poor aspiration for AI and it should not be called a “major scientific achievement”.
The essay's framework assumes that the scarce resource is clinical execution, not individual biological understanding. If we don't yet understand why drugs work differently in different patients (genomics, epigenomics, microbiome, etc.), then simply running more trials faster does not solve the efficacy problem, it may just produce more failures faster.
Drug discovery and early development are the bottlenecks, not clinical development. A flawed therapeutic hypothesis –no matter how many drug candidates per target– will not survive the real-world pressure of clinical development. Small molecules often suffer from low bioavailability and poor target specificity and selectivity properties. Biologics like antibodies are highly specific but are not selective, leading to on-target off-organ toxicity. Oncology is a tough one. After several rounds of therapeutic interventions with promising FDA-approved drugs, patients get into remission, then they relapsed, and finally they become refractory to any therapy. Sooner or later, drug resistance mechanisms emerge in the clinic.
Biopharma needs to organize its translational teams to capture the clinical feedback loop back into discovery in real time (clinical response data → computational model → next-generation scaffold). Interdisciplinary research (AI, Quantum Mechanics, Nanotechnology and Biotechnology), not just AI, is increasingly playing a major role in support of drug product development and CMC by addressing formulation, bioavailability, and targeted delivery. Thus, offering differentiated approaches to de-risk pipeline drug candidates. I believe the next frontier is when responsible AI (not only GenAI or hyped AI) can predict not just molecular properties, but clinical translation risk: which optimized drug candidates will actually show higher efficacy/lower risk in specific patient subpopulations, and which will struggle against drug resistance mechanisms you have not yet seen in the clinic.
Now and in the future the saying “we discovered a high quality and differentiated drug” will continue to be a strong basis for durable company value. Again, I am not talking about a drug candidate or an approved drug that only works in a few patients.
Preclinical assets and discovery-stage companies will NOT face valuation pressure if they focus on value instead of AI-powered volume. Specifically, companies anywhere in the world must build the most irreplaceable pipeline of scientific discoveries including FICs, BICs, and BID assets. So, that global markets cannot find comparable assets anywhere else in the world. Currently, Chinese companies are better positioned to deliver that thanks to the convergence of scientific, technological, regulatory, cultural and demographic factors. For instance, Hengrui Pharma has a culture deeply rooted in two pillars. 1st, scientific rigor with 24 Class 1 marketed innovative drugs – the highest tier of pharmaceutical registration anywhere in the world. 2nd, disciplined long-term vision rather than just quick upfront/milestone payments. On top of that, Hengrui has the patient datasets and the clinical & regulatory velocity that Western companies can only dream about.
By the way, the essay notes the "rise of China as a drug discovery and development engine". This is a sharp observation, but it sits awkwardly with its overall framework. Above I provided more context. If discovery commoditizes globally but clinical development remains bottlenecked by Western regulatory requirements, the value chain may fragment in ways that favor neither Formation Bio nor its competitors.
The section on “Predictive model” is not evolution, it's marketing periodization. As explained above, to improve health outcomes now, we need to capture the clinical feedback loop back into discovery in real time now. Therapeutic hypothesis validation and clinical translational medicine are where the true value and risk lie. The companies that truly transform biopharma will challenge the assumptions that present the frontiers as uneven.
The FDA's skepticism of in silico clinical trials is not merely bureaucratic inertia; it reflects a century of learning that biological complexity outstrips computational model complexity.
The rest of the essay is well thought-out marketing branding.
*This replica to Uneven Frontiers was not created, enhanced or checked using AI
Interesting perspective. Success in healthcare and biotech increasingly depends on turning insights into execution. Innovation alone isn’t enough; adoption, evidence, and long-term impact are what create real value.