OpenCRISPR, by Profluent
Just like ChatGPT predicts one word after the other, the company Profluent has developed an LLM model that predicts amino acid sequences… like new Cas enzymes of all types, including base editors, that they’re open-sourcing, which may just free innovation in the field from the long-debated patent on CRISPR.
In their paper, they share how they produced 4 million sequences with comparable or improved activity and specificity relative to Cas9 that are at least 400 mutations away in sequence from any previous enzymes. Half of these were generated from the model, half guided from a natural protein towards a particular CRISPR family using <50 N or C-terminal residues as prompts.
The model was pre-trained on non-CRISPR proteins across the evolutionary tree (unsupervised pre-training) then by using more than 1 million CRISPR-Cas operons of different types found in microbial genomes and metagenomes (fine-tuning).
Some other technical details I found interesting: in total, only 48.2% of generated proteins were functionally complete (had core Cas9 domains), and the generated proteins were predicted by AlphaFold2 (pLDDT > 80 for 81% of structures) which feels a lot like making two AIs hold a biological conversation.
As a company, Profluent is a spin-out of a Salesforce-funded project to develop an LLM for amino acids called ProGen. Their Chief of Business, Hilary Eaton, has shared their plans to research institutions and drug developers to safely expedite the development of new CRISPR genetic therapies. For now, their CRISPR enzymes (not the model!) will be available for use free of cost under a term sheet that restricts use in germline cells and non-reproducing plants.
Though I understand this at an extremely superficial level, I’m still wondering, on the bio IP side, what the limits or guidelines are and how generative AI will redefine them. Is there a percentage or number of amino acids or bases that mark the limit between infringing a patent and not?
The Halo, by Prophetic
Keep reading with a 7-day free trial
Subscribe to Biopunk to keep reading this post and get 7 days of free access to the full post archives.