Thread by @DrTomsLens
Stocks · · 1 min read · x.com ↗
Dr. Tomislav Marinovic @DrTomsLens 2026-04-21
A minor update to my $ABCL thesis on biology-first structural advantage.
Earlier today, I mentioned that in a world increasingly going AI-first, biology is exposing a key weakness of AI: all major AI models are trained on pretty much the same public internet data, but in antibody design, you need to get that data yourself.
Turns out, not only do open AI models for biology lack access to proprietary data, but the internet text data used to train models in biology may also be contaminated and mislead the model.
A lot of the recorded literature is actually incorrect, and many published results don’t replicate.
The AI drug discovery companies most likely to succeed may be the ones with a unique way to generate science tokens that don’t exist in the public domain.
2026-04-21
Alex on why AI drug discovery companies need to generate novel data to succeed:
"AI models based on the research that's available is a lot of garbage in and garbage out."
"A lot of the recorded literature is actually incorrect. There's been tons of studies that show if you go x.com/patrick_oshag/…
Thoughts on Healthcare Markets and Tech @thoughtson_tech 2026-04-22
The part that doesn't get discussed enough: proprietary data isn't uniform either.
There's a spectrum from "data we collected" to "data only we could collect," and the distance between those two positions is where actual moats form or fail. A company can own a massive internal
Dr. Tomislav Marinovic @DrTomsLens 2026-04-22
Good points, thanks for sharing!
Tyler Bossermn @tyler_bosserman 2026-04-22
This part of the interview was so fascinating. We all remember and know how much AI can hallucinate, especially in the early days. Biology is no different.
AbCellera’s ability to do binding and functional assays at scale is one of its superpowers. They can get a good sense of
Dr. Tomislav Marinovic @DrTomsLens 2026-04-22
Great points Tyler 🎯
Jarek @Jare92365776 2026-04-22
100% Right
Public domain stuff like here: https://rcsb.org
Funny is how those data are created.
Methods to obtain data - inperfect.
Data processing by AI models to 3D structures - inperfect.
"Accuracy / Resolution" is questionable.
Than you make molecule form this? LOL
Dr. Tomislav Marinovic @DrTomsLens 2026-04-22
Thanks for sharing this!
DCLXVI @shareshares_wme 2026-04-22
Very true