What is the role of the GPT-4B Micro model in the RetroBioSciences collaboration, and how was it used to engineer improved OSKM factors?

In the OpenAI and RetroBio collaboration, GPT-4B Micro was a protein-focused foundation model created to support protein engineering, including sequence generation, targeted in-filling edits, and 3D structure generation. It was used to generate candidate improved OSKM factor sequences through an evolutionary-series prompting strategy, and filtered variants were experimentally tested by RetroBio in human fibroblasts.

GPT-4B Micro was not trained on protein sequences alone. It was mid-trained on multiple biological modalities, including protein sequences, tokenized three-dimensional structures, co-evolutionary information from multiple sequence alignments, protein-protein interaction data, and relevant scientific text. This allowed it to handle common protein design workflows: generating sequences, performing fill-in-the-middle edits that preserve critical regions while redesigning others, and producing structure tokens corresponding to plausible 3D conformations. In the OSKM effort, the model was used in a lab-in-the-loop workflow. A prompting strategy based on evolutionary series made the system steerable, so it learned patterns across related proteins and proposed new sequences conditioned on examples. The model generated thousands of candidate sequences, which were filtered to preserve essential domains and maintain diversity, then narrowed to a few hundred variants that a typical lab could feasibly test. RetroBio synthesized DNA for these variants, delivered them into human fibroblasts using lentiviral constructs, reprogrammed cells for roughly 10 to 14 days, and used cell-surface markers as readouts.