@dominik_scherm Looks cool. I would audit the agent strategies on the VHH-IL-6 task very carefully. IME there aren't any methods with good AUROC for this task, but there are tons of subtle data leak exploits that get nominally high scores. Generating a genuinely clean holdout is not easy.
New CROs may benefit from studying the airline industry. Limited differentiation, brutal price competition, high actual and reputational costs for mistakes. Not much matters other than being on the cost-reliability Pareto frontier.
That said, if new CROs operated by people from the tech world result in better price transparency and more ergonomic ways to order experiments that's a win. Some incumbents are hyper optimized on the backend but have user experience problems.
@TensorTwerker An agent controlling headless blender w/ molecular nodes via the API is so much fun. That way the code to render the vis can be version controlled and is independent of the agent/session. Modal is useful for faster rendering if you're making animations.
@theanantpai@SynBio1@lucas_nivon The assets are shopped around and the data are offered with them. If the assets aren't sellable the data often goes nowhere. In the last 6 mo I've heard of several people working on aggregating data from failed biotechs into packages for AI labs. Consignment model seems popular.
@labordabio GlycoSHIELD is great for this. https://t.co/H2QVIneaR5 can work in a pinch too. Even if only some glycan conformations are occluded it will affect affinity due to entropic penalty
Agree discovery isn't the main bottleneck, but program failure due to not engaging the target doesn't usually show up in ph II/III stats. If you're initiating ph II without strong evidence of in vitro target engagement idk what you're even doing. I imagine those 3% failures are something like limited target-site exposure, not literally failing to engage the target in vitro.
How many discovery programs fail to progress due to not finding a sufficient quality lead? Unknowable, but >3%. Maybe 10-20%?