Agree with the direction, and the hardest part of that shift is rarely the org chart - it is the operating model underneath it. If you move from a fully integrated company to a network of partners, CDMOs and platform vendors, you no longer need armies of doers but a thin layer of people who can specify, contract and hold others accountable: MSAT, tech transfer, vendor QA, CMC leads, programme managers with real decision rights. Those are exactly the profiles that are scarcest in Europe right now, so the future pharma company is bottlenecked on orchestration talent rather than on capital. We mapped how life sciences companies redesign their operating model for scale here: https://www.sire-search.com/how-life-sciences-companies-redesign-operating-models-for-scale/
Interesting perspective. I usually hear an opinion that Europe is a cradle of research and talent, but is lacking capital for scale-up and growth phases of biotech ventures. Anyway, the practical question is how to make Europe more competitive when it comes to global biotech/pharma positioning... Thanks for the links with some insights, also feel free to share insights directly here in the chat.
A key data source that (IMHO) would really help are details of negative results, especially from the late preclinical and all stages of clinical studies. I have some of my own ideas* how to encourage the release of such data, but what do you and everyone else here think would work?
*Additional financial incentives would probably help, but what forms they'd take needs to be carefully thought through and modeled.
You are right to say that negative results are important not only in AI drug discovery (training models on both positive and negative results) but also for scientific progress in general. Incentives are clearly misaligned, as recent research shows that publishing negative results generally leads to much fewer citations. One workaround is to standardize public data repositories or other easy ways to report on negative results without committing to a full-scale publication. Other ideas include that grant programs could require all data be deposited and reported alongside positive results, and be treated as equal outcomes for the purpose of effort billing. This would lead to direct financial incentives to honestly report on everything, not only positive results. It would be harder in a commercial setting, though... and here, may be one way to go about it is some forms of industry partnerships where companies would agree to fund efforst leading to even negative results, may be with certain clauses requiring certain ration of positive/negative (to avoid situation when nobody cares for positive results at all, because being paid for whatever comes out anyway).
I want to use AI to crunch literature and form hypotheses as you said. Specifically, I want it to form hypotheses of different potentially effective ways of treating or curing adrenal/endocrine diseases. How can I approach this best? What tools can I use? What would you recommend?
Thank you for your comment and question. I think to illustrate how you can use AI for this sort of tasks, the best way to go about answering this question is do exactly that: use AI. So, I literally asked your question to Gemini, and here is the step-by-step approach:
***Recommended Approach
Prioritize RAG over Fine-Tuning: Fine-tuning teaches a model style and general domain patterns, but RAG lets the AI read exact source documents dynamically, preventing hallucinations and saving costs.
Curate High-Quality Corpora: Gather specific papers, clinical trials, and reviews from databases like PubMed instead of raw unvetted texts.
Use Agentic Workflows for Hypothesis Generation: Set up multi-step prompts where the AI first extracts biological mechanisms, then contrasts pathways, and finally proposes non-obvious therapeutic hypotheses.
Recommended Tools
Elicit: AI research assistant that finds relevant papers, extracts key data points, and synthesizes findings across multiple studies.
Consensus: AI search engine that reads scientific literature and tells you what the consensus is on specific medical treatments.
Sematic Scholar / APIs: For building custom programmatic pipelines or connecting local vector databases if you need a heavy developer approach.
***
In practise, what I would do is simply create a project in Claude/Claude Cowork and update a well-curated corpus of relevant data about adrenal/endocrine diseases to create a retrieval layer, and then simply use that project for brainstorming.
So, my point is, AI's best capability is exactly that -- brainstorming and learning, and you can use it to go deeper on each of the items above; for example, ask it to explain RAG and guide you on how to set it up in terms of an actual technical project. I use AI all the time to help me manage my business (accounting, taxes, legal aspects, etc) and it is very helpful. It is NOT substituting final decision-making; I still need an actual accountant/lawyer to vet whatever AI suggests, but it is undoubtedly a game changer.
I have done stuff (similar) to this. But, it keeps hallucinating and being very much anchored to existing domains I discussed with the AI in question (whether it's claude, gemini, or an open-weight model through Hermes or smth).
I think, this will only work when I determine the theses; 'these fields are interesting' and then give it in a loop to AI which can do a semantic search and answer the simple question: Did you find anything on this topic?
I think it is impossible to complete eliminate hallucinations xhen relying on LLM-based systems. isn't? We are only talking about achiving some practically viable level. Here I reviewed a case with agentic AI that might be relevant: https://www.moleculesempires.com/p/who-wins-when-agents-come
The line about handoffs being where the years disappear matches what I see inside regulated labs, and I think it carries more weight than the section it sits in gives it.
I'd draw one distinction. You describe each function holding its own data, its own vocabulary, its own people. The vocabulary gap is real, but it's visible, so people build glossaries and mapping tables and route around it. The gap that costs years is lineage. The receiving function can't tie a result back to the sample, the instrument run, and the hypothesis that produced it. That failure stays invisible until an auditor or a diligence team asks you to reconstruct the chain, and by then you're rebuilding it from lab notebooks and email threads.
This matters for the second half of your argument. The connective layer you describe, one system carrying a molecule from design through submission and tracing a safety signal back to the design that caused it, only works if that trace physically exists in the data. A model reasoning over records that lost their provenance at capture gives you fast answers you can't defend to a regulator. The incumbents' proprietary data has the same problem in a different form. Decades of trial data is worth what its traceability lets you claim, and a lot of it was never captured with that in mind.
The unglamorous fix comes first: deterministic normalization at the instrument, checksums on every file, an append-only audit trail, and an explicit link from hypothesis to outcome. I build that layer for life science teams heading into FDA submissions and technical diligence, which is why your framing caught me.
The question I keep landing on is whether a newcomer renting CROs and CDMOs actually ends up with better lineage than an incumbent, or worse. Every rented function is another handoff across an organizational boundary, and the data comes back in whatever shape the vendor chose. Do you see the connective layer solving that, or does outsourcing recreate the same fragmentation with contracts in between?
This is such a great piece Andrii, thank you!
I've been trying to learn more about this space, and so far this has been one of the most illuminating articles I've read.
I really appreciate your positive feedback, Majose!
Agree with the direction, and the hardest part of that shift is rarely the org chart - it is the operating model underneath it. If you move from a fully integrated company to a network of partners, CDMOs and platform vendors, you no longer need armies of doers but a thin layer of people who can specify, contract and hold others accountable: MSAT, tech transfer, vendor QA, CMC leads, programme managers with real decision rights. Those are exactly the profiles that are scarcest in Europe right now, so the future pharma company is bottlenecked on orchestration talent rather than on capital. We mapped how life sciences companies redesign their operating model for scale here: https://www.sire-search.com/how-life-sciences-companies-redesign-operating-models-for-scale/
Interesting perspective. I usually hear an opinion that Europe is a cradle of research and talent, but is lacking capital for scale-up and growth phases of biotech ventures. Anyway, the practical question is how to make Europe more competitive when it comes to global biotech/pharma positioning... Thanks for the links with some insights, also feel free to share insights directly here in the chat.
A key data source that (IMHO) would really help are details of negative results, especially from the late preclinical and all stages of clinical studies. I have some of my own ideas* how to encourage the release of such data, but what do you and everyone else here think would work?
*Additional financial incentives would probably help, but what forms they'd take needs to be carefully thought through and modeled.
You are right to say that negative results are important not only in AI drug discovery (training models on both positive and negative results) but also for scientific progress in general. Incentives are clearly misaligned, as recent research shows that publishing negative results generally leads to much fewer citations. One workaround is to standardize public data repositories or other easy ways to report on negative results without committing to a full-scale publication. Other ideas include that grant programs could require all data be deposited and reported alongside positive results, and be treated as equal outcomes for the purpose of effort billing. This would lead to direct financial incentives to honestly report on everything, not only positive results. It would be harder in a commercial setting, though... and here, may be one way to go about it is some forms of industry partnerships where companies would agree to fund efforst leading to even negative results, may be with certain clauses requiring certain ration of positive/negative (to avoid situation when nobody cares for positive results at all, because being paid for whatever comes out anyway).
Insightful piece, thank you!
I want to use AI to crunch literature and form hypotheses as you said. Specifically, I want it to form hypotheses of different potentially effective ways of treating or curing adrenal/endocrine diseases. How can I approach this best? What tools can I use? What would you recommend?
Many thanks 🧡💙
Thank you for your comment and question. I think to illustrate how you can use AI for this sort of tasks, the best way to go about answering this question is do exactly that: use AI. So, I literally asked your question to Gemini, and here is the step-by-step approach:
***Recommended Approach
Prioritize RAG over Fine-Tuning: Fine-tuning teaches a model style and general domain patterns, but RAG lets the AI read exact source documents dynamically, preventing hallucinations and saving costs.
Curate High-Quality Corpora: Gather specific papers, clinical trials, and reviews from databases like PubMed instead of raw unvetted texts.
Use Agentic Workflows for Hypothesis Generation: Set up multi-step prompts where the AI first extracts biological mechanisms, then contrasts pathways, and finally proposes non-obvious therapeutic hypotheses.
Recommended Tools
Elicit: AI research assistant that finds relevant papers, extracts key data points, and synthesizes findings across multiple studies.
Consensus: AI search engine that reads scientific literature and tells you what the consensus is on specific medical treatments.
Sematic Scholar / APIs: For building custom programmatic pipelines or connecting local vector databases if you need a heavy developer approach.
***
In practise, what I would do is simply create a project in Claude/Claude Cowork and update a well-curated corpus of relevant data about adrenal/endocrine diseases to create a retrieval layer, and then simply use that project for brainstorming.
So, my point is, AI's best capability is exactly that -- brainstorming and learning, and you can use it to go deeper on each of the items above; for example, ask it to explain RAG and guide you on how to set it up in terms of an actual technical project. I use AI all the time to help me manage my business (accounting, taxes, legal aspects, etc) and it is very helpful. It is NOT substituting final decision-making; I still need an actual accountant/lawyer to vet whatever AI suggests, but it is undoubtedly a game changer.
Thank you for your reply.
I have done stuff (similar) to this. But, it keeps hallucinating and being very much anchored to existing domains I discussed with the AI in question (whether it's claude, gemini, or an open-weight model through Hermes or smth).
I think, this will only work when I determine the theses; 'these fields are interesting' and then give it in a loop to AI which can do a semantic search and answer the simple question: Did you find anything on this topic?
Thoughts?
I think it is impossible to complete eliminate hallucinations xhen relying on LLM-based systems. isn't? We are only talking about achiving some practically viable level. Here I reviewed a case with agentic AI that might be relevant: https://www.moleculesempires.com/p/who-wins-when-agents-come
I appreciate you and your work. Keep it up
Thanks a lot!
The line about handoffs being where the years disappear matches what I see inside regulated labs, and I think it carries more weight than the section it sits in gives it.
I'd draw one distinction. You describe each function holding its own data, its own vocabulary, its own people. The vocabulary gap is real, but it's visible, so people build glossaries and mapping tables and route around it. The gap that costs years is lineage. The receiving function can't tie a result back to the sample, the instrument run, and the hypothesis that produced it. That failure stays invisible until an auditor or a diligence team asks you to reconstruct the chain, and by then you're rebuilding it from lab notebooks and email threads.
This matters for the second half of your argument. The connective layer you describe, one system carrying a molecule from design through submission and tracing a safety signal back to the design that caused it, only works if that trace physically exists in the data. A model reasoning over records that lost their provenance at capture gives you fast answers you can't defend to a regulator. The incumbents' proprietary data has the same problem in a different form. Decades of trial data is worth what its traceability lets you claim, and a lot of it was never captured with that in mind.
The unglamorous fix comes first: deterministic normalization at the instrument, checksums on every file, an append-only audit trail, and an explicit link from hypothesis to outcome. I build that layer for life science teams heading into FDA submissions and technical diligence, which is why your framing caught me.
The question I keep landing on is whether a newcomer renting CROs and CDMOs actually ends up with better lineage than an incumbent, or worse. Every rented function is another handoff across an organizational boundary, and the data comes back in whatever shape the vendor chose. Do you see the connective layer solving that, or does outsourcing recreate the same fragmentation with contracts in between?