United StatesHealthcare and life sciences2025 – ongoingMetagenomi
Generating candidate CRISPR enzymes with a protein language model on AWS Inferentia
Metagenomi generated more than 1 million candidate enzymes for $2,613 by moving its protein model to AWS Inferentia
Problem
Metagenomi develops gene-editing therapies based on CRISPR enzymes. It fine-tunes a protein language model on its own enzyme data and needs to generate very large numbers of candidate sequences. Running that on Nvidia L40S GPUs was costly.
Approach
Metagenomi fine-tuned the ProGen2 model on its proprietary enzymes. With AWS, it traced the model for Inferentia2 and ran generation as batch jobs on Spot Instances. Generated sequences were then filtered with tools including ESMFold, mmseqs and hmmsearch.
Results
Takeaway
For long generation runs, the choice of inference hardware changed the price of screening candidates by a wide margin. Metagenomi still had to filter the output, and only some of the generated enzymes passed.
Categories
- Industry
- Healthcare and life sciences
- Business function
- IT and R&D
- Innovation type
- Process innovation
- AI technology
- Generative AI
- Adoption
- Off-the-shelf tool or platform
- Goal
- Cost savings
- Evidence
- Vendor case study
Sources
Source summaries
AWS authors wrote the post with Metagenomi. It describes running a fine-tuned ProGen2 model on Inferentia2 Spot Instances and reports up to 56% lower cost than L40S GPUs for longer sequences. A scale test produced well over 1 million enzymes for $2,613, and several hundred thousand passed the filters.
Full details are in the original sources.