Designing proteins can open paths to new therapeutics and industrial processes, but the conventional approach involves proposing a three-dimensional shape and then searching for an amino acid sequence likely to fold into it. Microsoft’s EvoDiff takes a different route: it generates proteins from sequence data without needing a target structure.
Designing from sequence instead of structure
Proteins carry out important cellular functions, and understanding them can help researchers investigate disease mechanisms. Creating new proteins could also support new kinds of drugs, therapeutics and delivery methods.
In the traditional lab design process described by Microsoft, researchers first identify a structure that might perform a desired task. They then need to find a sequence of amino acids that can fold into that three-dimensional form. The source article describes this work as costly in both computational and human resources.
EvoDiff aims to remove the need to specify the target protein’s structure. Microsoft says its framework can generate high-fidelity, diverse proteins from protein sequences. Senior researcher Kevin Yang told TechCrunch that the project seeks to extend protein engineering beyond the structure-function approach toward sequence-first design.
How the AI framework generates proteins
At the center of EvoDiff is a 640-million-parameter model trained on proteins from different species and functional classes. Its training data came from the OpenFold data set for sequence alignments and UniRef50, a subset of UniProt, the database of protein sequence and functional information maintained by the UniProt consortium.
EvoDiff uses a diffusion model, an approach also used in image-generation systems such as Stable Diffusion and DALL-E 2. In broad terms, the model starts with a protein sequence that is mostly noise and gradually removes that noise, step by step, to produce a sequence.
That sequence-based approach gives the framework room to work with designs that are not defined by a final folded structure. Microsoft researcher Ava Amini said the team sees sequence generation as offering generality, scale and modularity, along with ways to guide designs toward functional goals.
Potential uses include filling design gaps
Microsoft describes EvoDiff as useful for generating new proteins, including enzymes that could be explored for therapeutics, drug delivery methods or industrial chemical reactions. These are potential applications; the source does not report that the generated proteins have already been shown to work in those settings.
The framework can also generate sequence around a specified portion of an existing protein design. For instance, if researchers provide a part that binds to another protein, EvoDiff can produce an amino acid sequence around that part according to a set of criteria. This could let researchers preserve a desired component while exploring surrounding sequence options.
Because EvoDiff operates in sequence space rather than designing around a fixed protein structure, it can also generate disordered proteins. These proteins do not settle into a final three-dimensional shape, yet the source notes they can play biological roles, including enhancing or decreasing the activity of other proteins.
Research remains at an early stage
The work behind EvoDiff has not been peer reviewed. Sarah Alamdari, a Microsoft data scientist who contributed to the project, said more scaling work would be needed before the framework could be used commercially.
Alamdari also said generation quality might improve if the model were scaled from 640 million parameters to billions of parameters. For more fine-grained control, the team would want to condition the framework on text, chemical information or other ways to describe a desired function.
The next step for the team is to test the proteins EvoDiff generated in a lab to find out whether they are viable. Those experiments will help determine whether the sequences translate into proteins that can function as intended. If they do, the team plans to begin work on a next generation of the framework.