Synthetic proteins are being built with the help of AI models
…Tweaking proteins to do human bidding is nothing new. Enzymes and antibodies, for example, have long endured such outrages. But that is not what Dr Ennist is up to. Rather than modifying existing proteins, his versions are being designed from scratch, using artificial-intelligence (AI) models, to be optimised for the task at hand. To start with, they would be inserted into a suitable organism, such as a plant or bacterium, to do their thing there. But eventually they could, he hopes, operate independently and thus form the basis of a new type of solar cell—one that turns out petrol rather than electricity.
…Others are taking a slightly different tack. Profluent, in Emeryville, California and EvolutionaryScale, in New York, are building protein-design AI models that resemble not image-generating software, but large language models (LLMs) of the sort that power the world’s chatbots. These firms’ models treat the amino-acid sequences in protein chains like the words in a piece of text—analysing relationships found in zillions of exemplars to design novel useful structures.
According to Ali Madani, Profluent’s chief executive, the firm is particularly focused on creating new CRISPR-Cas gene-editing tools. Here, its USP is a curated database of around 5m CRISPR-Cas protein complexes on which its AI model has been trained in order to design new versions.
