Less excitement than in the previous weeks for Techbio announcements. There were however two cool papers showing the “real world” performances of Foundation models for DNA and proteins, which were rather lackluster.

This Week in TechBio 2024/02/15

• 2024/02/14 biorXiv: A nice paper by the folks from Microsoft Research on how well Protein Language Models do at transfer learning. The paper found that while these models are good at building features based on structure, they fail at being good for functions based prediction. Here is the associated twitter thread.
• 2024/02/13 BusinessWire: BMS signes a deal with VantAI for up to \$647m to use their drug discovery AI platform.
• 2024/02/13 SEC: Exscientia’s board removes their CEO under allegation of inappropriate relationships with 2 employees.
• 2024/02/12 biorXiv: A review / benchmark on the ability of biological foundation models to work on tasks they were not specifically trained on, i.e. few-shots learning. The paper finds that these foundation models do not manage to transfer their knowledge to 3’UTR and are beaten by smaller model designed for that task. While it may seem “obvious” it should show that these models (enformer, DNABert, BigRNA) are most definitely not usable out of the box and would require some careful fine-tuning and optimization to be performant. Here is a twitter thread about it from Anshul Kundaje.
• 2024/02/12 Endpoints: Iambic Therapeutics publishes NeuralPLexer in Nature Machine Intelligence a protein and small molecules binding model. Interestingly, according to their whitepaper, the only data that they relied on was open data from PDB that they refined.
• 2024/02/09 Fierce Biotech: Unlearn, a digital twin company, raised a \$50m series C. The company designs “digital twins” to do in silico placebo arms for clinical trials, and their product has apparently been “accepted” for use in phase 2 and phase 3 by the FDA and EMA, however I assume it can only be used “on top” of an existing trial.
• 2024/02/08 Fierce Biotech: Elemind, a startup focused on using electrical stimulation as a therapeutric for neurological diseases, comes out of stealth. Their pitch is that they used AI to learn patterns of activity in the brain, and how to affect them in the “right” direction with external electrical stimulation. While their rational is that it allows them to avoid using pharmaceuticals that would have to affect the whole body, I have to admit that it sounds a bit creepy (it’s not rational of me, but it is “icky”).
• 2024/02/07 STAT: The US’s government plans on having “AI assurance laboratories”, where researchers and company could deploy their AI model for diagnostics, pathology, or general hospital use, and where hospitals could properly evaluate how good they run in a hospital environment (compared to a controlled environment). I find this initiative amazing because it will help speed up validation and development of these products, as well as help them get deployed in real life. There is however some equity concerns because only fancy reserach hospitals will get to deploy these, because they are the only ones knowing how to use and evaluate them, which could let the ones with less funding behind. While I understand these concerns, I find that this is counter productive and just lowers everyone down.

• 2024/02/07 STAT: STAT’s writers conducted many interviews with folks from industry and academia about how they see the role of AI for the pharmaceutical industry going forwards. While most companies believe that it will be extremely important for the future and invest in it, they seem not yet to know how or in which areas yet. This can be seen in the Novartis-Isomorphic contract where Novartis seems to have unloaded programs they got stuck on as a hedge, instead of important programs. Folks also want to use that tool for patient selection and recruitment, but it is currently unclear how and why that would work. There is interest about digital twins, but they rely on having a good representation of patients which I personnaly find unlikely, as if we could directly model the effect of a molecule on a patient, we would not bother with molecular biology. Overall the issue seems to be that while the technology is extremely promising, it is not yet a product that can be used, and everyone seems to explore how to turn it into one.

  In many diseases, there is simply not enough data to precisely map biological processes to make the development of therapies anything but a treacherous process with an uncertain outcome.

  Both biologists and machine learning researchers caution against extrapolating the success of generative AI in consumer tasks to the work of solving biological problems.

  “We’re at the point of (asking), ‘What can this do?’” said Fiona Marshall, president of biomedical research at Novartis. “Very little is proven, so we’re exploring the possibility, and doing that through test cases.”

• 2024/02/07 STAT: In the aftermath of the UnitedHealth scandal about refusing care based on an internal algorithm, with no margin of maneuver for its employees, Medicare wrote a memo warning insurers not to rely on “AI” before putting the patient needs first. While the idea is noble and probably correct, implementing that will be a big issue because AI is just a fancy word for statistics, and insurance companies run on actuarial science, another fancy word for statistics.
• 2024/02/07 STAT: Creation of Scion Life Sciences, a new VC firm that plans on focusing on very long term returns (exits in 20 years instead of 10), whose goal is to invest in a relatively small amount of companies that would become the new Recursion or Vertex.
• 2024/02/07 Fierce Biotech: Ambience Healthcare raises a \$70m series B. The company uses text generative AI tools to speed up, scribing, referals, billing, and after visit summaries. They are in production in a few hospitals that also participated in that funding round.
• 2024/02/07 Endpoints: Aizon, a company focused on designing AI tools for monitoring manufacturing at CDMOs and CMOs, raises a \$20m Series C. This new round will be used to increase their headcount and push their third product into production. It’s interesting to see some activity in that field that I would have tended to overlook (see C3’s deal with Genentech)
• 2024/02/06 Crick Institute: The crick Institute plans on opening on opening a 30k sqft and a 52k sqft lab spaces near King’s cross to host biotech and AI companies. This space will accommodate state of the art labs, as well as collaborations with the Crick. Such efforts are sorely needed to help Europe catch up to Boston and the Bay Area, and will be welcomed by Europeans who would rather not have to cross the Atlantic to find a lively bio community (me included).
• 2024/01/31 Fierce Biotech: BD and Techcyte collaborate for a cervical cancer detection software using pap smears and computer vision, just before Hologic’s one gets FDA clearance.

I probably missed quite a few announcements, don’t hesitate to DM me @gama_search if you see anything missing or needing corrections.