This makes protein- interactions WAY more interpretable by reasoning. this is real interpretability… - like a foundationally more important one than even jude stiel’s
https://x.com/i/status/2035013002244866547
https://x.com/Radii2323/status/2035012134979961132
See new posts
Conversation
·
today we launched bioreason-pro, try using it: http://app.bioreason.net
·
we believe that reasoning can unlock a lot of potential and hypotheses in discovery and science. however, most of the biological data is not in natural language. to tackle this, we have combined biological foundation models and LLMs to reason across biological modalities and give
Show more
·
bioreason-pro can hypothesize deep molecular functions which later has been validated in the lab by careful experiments.
the model has also shown that it can reason into very depth on the mutations and structure of the models
·
·
it also achieved SOTA performance on Gene Ontology term prediction and provided more in-depth annotations than what scientists currently use.
·
we trained the model in two steps: SFT and RL. the SFT was based on synthetic reasoning traces by GPT-5 which were grounded by the ground truth of our biological data. RL made the model more realistic and robust to hallucinations which lead to increase in accuracy. We used a
Show more
·
Thanks for the amazing bioreason team,
and
for advising the project,
and
for supporting us, and ofc the amazing effort form
and
for this project
·
Special thanks to
for designing the beautiful website! And huge thanks to
,
, and
team for providing us the inference engine to deploy the model