TL;DR: We’ve built the foundations of a pharma in the age of superintelligence: data partnerships, biology models and autonomous research agents. We have SOTA results on finding overlooked drugs and predicting clinical trial success rates.

Hey everyone! We (Sanjukta Bhattacharya, Christian Gensbigler, and I) started Atlas Discovery to get treatments to patients who need them most.

The problem: Today's pharma works for only ~150 diseases. Below roughly $150M in expected revenue a program never gets funded, because the cost of developing a drug is fixed by the knowledge work it requires, not the size of the market. 10,000+ diseases remain without an approved treatment even where the biology is tractable. Meanwhile, there are 100,000+ existing drugs, and barely 1% of the drug-disease combinations have ever been tried. The bottleneck is no longer supply of compounds; it's the judgment needed to search that space and run the trials, and the removal of human coordination costs.

Our thesis is that AI will drop the cost of drug development enough to push the viable threshold from ~$150M toward ~$15M, opening a ~$500B market incumbents have structurally ignored. A company built for that world needs three things: proprietary data, models that predict clinical trial outcomes, and agent systems that replace knowledge work. We use agents to repurpose existing drugs for the long tail of diseases, then run the clinical trials faster and cheaper.

Progress: In three months we've built the first version of all three layers. Data: an 8 figure LOI with an AI lab and 6 figures in contracted revenue. Models: state-of-the-art clinical trial prediction results published at three conferences,. Agents: three rare disease foundations are now testing drugs surfaced by our agents.

On drug repurposing using agents, we see nearly 5x improvement over current frontier LLMs. Read more about this benchmark here.

We've also shown that models trained on in vitro cell data (Latin for "in glass") and mouse data can translate to gains in predicting patient response to drugs in human clinical trials.

We’ve also shown SOTA results in being able to predict clinical trial success when trained on over 100,000 clinical trials data and evaluated on a held out set of 10,000 trials. Read more here

We’ve presented our research in 3 top conferences (ICLR, CSHL, ICML) and shown that new AI architectures like discrete diffusion can learn these gene expression patterns and predict response to treatments with over 10x improvement in accuracy while training on 50x less data.

Reach us at founders@atlasdiscovery.bio

Blog: https://atlasdiscovery.bio/blog