Tatta Bio

What an incredible experience here at the Defense Advanced Research Projects Agency (DARPA) AI BTO Pitch Day! As the event continues, I love engaging with top innovators in biotech and AI. Michael Koeris showed up today dressed in his pitch-day duds with an enthusiasm level that matched his suit! Mike will also be at #SynBioBeta2025 this May in San Jose. I was also able to catch up with George Church, and yesterday, I chatted with Alexander Titus to explore how he sees the AI + Bio landscape and provide updates about the National Security Commission on Emerging Biotechnology (NSCEB). The energy and ideas flowing here are a testament to how rapidly biotech and AI are evolving. Congratulations to all who will receive funding!

on Monday we chat about the OMG dataset: Open MetaGenomic Corpus for mixed-modality genomic language modeling https://lnkd.in/eer7BW9W With the author Andre Cornman from Tatta Bio Join us on zoom at 11am EDT / 5pm CEST: https://lnkd.in/eXnrUCsc

Today we're open-sourcing DGEB, the first large-scale functional evaluation benchmark for biological language models ?? DGEB fills a key gap in BioML by providing diverse functional benchmarks for assessing pLM and gLM representations. We address two critical problems in existing functional evals: 1) Bias in evaluation datasets and 2) Lack of evaluations that assess functional relationships *between* genomic elements. Biology is incredibly diverse, yet biological research is biased towards a few model organisms. Foundation models of biology must learn from data across the tree of life. Existing evaluations using only model organisms perpetuate this bias, hindering diverse discovery and design.DGEB features 18 expert-curated datasets that span all three domains of life! ?????? For underrepresented taxa (e.g. #Archaea), we find that model performance is poor and do not improve with scaling, highlighting a key limitation of existing models. Inferring biological function requires understanding the evolutionary and functional relationships *between* biomolecules. DGEB features multi-sequence tasks, allowing us to ask questions such as how well do learned embedding distances recapitulate phylogenetic distances? Can we use pLMs to retrieve bacterial homologs given archaeal proteins? Do pLMs learn co-regulatory interactions or distinguish paralogs from orthologs? Can multi-element functions (e.g. biosynthetic gene clusters) be captured in representations? DGEB is inspired by NLP benchmarks, in particular MTEB from Hugging Face. Our field can only progress with transparent, collaborative and diversified evaluation datasets. We encourage the contribution of expert-curated datasets through our github repo: https://lnkd.in/gHJ_WWsT This was a true multi-disciplinary effort by our team at Tata Bio: Jacob West-Roberts Joshua Kravitz Nishant J. and Andre Cornman. This work is made possible with generous support by Schmidt Futures! Read our manuscript at: https://www.tatta.bio/dgeb