Workflow from Scientific Research

Advances in AI change the process of de novo protein design. User-defined goals (left) and inputs (middle) are used to generate proteins with new structures and functions (right). Categories 14 depict increasingly straightforward prompts leading to increasingly complex design outputs. Boxes indicate design goals with experimentally validated examples. (1) AI-based methods to design new protein structures can be unconstrained (generating diverse protein folds; helices shown in red and strands in yellow) or constrained to diversify a particular fold. (2) Most current methods to design function specify a motif with defined residue positions and orientations in a functional site. In a second step, a protein is generated de novo that surrounds and stabilizes the precise functional site geometry. This process is called motif scaffolding. (3) Advances in AI-based methods are in development that only define the target, and the design method generates a predicted binder. (4) Starting from a target function (for example, converting substrate S to product P), an AI method could generate a protein with the requirements for that function. Currently, protein language models trained on specific protein families or large experimental datasets can generate new sequences with functions similar to those in the training set.

DOI