BlankBio bets RNA design needs small contrastive models, not billion-parameter scale

An applied AI lab training RNA foundation models — from an open mRNA design toolkit toward models that predict disease progression and patient response for pharma.

In biology, sequence is mostly noise, so scaling laws underperform: joint-embedding small models beat billion-parameter ones on usable RNA function per GPU-hour.

Philip Fradkin and Ian Shi studied self-supervised objectives for RNA during their PhDs; Jonny Hsu led product and strategy at Valence Discovery through its 2023 acquisition by Recursion. The trio open-sourced Orthrus and mRNABench and joined YC's Summer 2025 batch to productize RNA design.

Orthrus, a 10M-parameter model trained on 4 GPUs in 14 hours, outperforms Evo2 — a 40B-parameter model trained on 1000 GPUs for a month — on mRNA half-life prediction, and on the team's 59-task mRNABench, Orthrus-MLM beats Evo2 while being roughly 700x smaller; the founders report they could not get Evo2-40b to perform well even working with its authors. Over the following year the company's positioning broadened from an mRNA design toolkit to 'RNA intelligence for precision medicine': models that capture tumor-specific biology to help pharma raise clinical-trial success rates.

Only about 2% of the human genome encodes protein; next-token prediction spends capacity modeling intergenic static, so standard scaling under-delivers on nucleotides.

Contrastive joint embeddings group sequences by function — what biologists actually select for — instead of autocompleting nucleotides.

4 GPUs for 14 hours versus 1000 GPUs for a month keeps training inside a startup's budget and iteration loop.

Open-sourcing Orthrus and mRNABench before selling anything made the benchmark the company's proof and its distribution at once.

Where data is sparse and signal hides in noise, the bitter lesson bends: matching the pre-training objective to the structure of the problem can beat another zero in parameter count.

As of September 2026, Blank Bio is listed as Active in YC's directory, describing itself as an applied AI research lab working to raise clinical-trial success rates. The public artifacts remain the open benchmark and model — mRNABench on GitHub, Orthrus embeddings with API access — and no external funding round had been announced publicly by that date. The field is heavily funded: the founders' launch post sized rivals including Strand ($153M), Orna ($221M), and Sail Biomedicines ($440M), arguing the money has outrun the tooling.