Meta’s AI Helps You Predict Shape-And-Structure of Proteins

Meta's AI has the potential to change the way we analyze protein structures

Scientists have worked very hard to understand the structures and functions of proteins, and today the AI research team at Meta stated that they have developed a model that could predict the three-dimensional (3D) structure of proteins based on their amino acid sequences. Meta's AI is built on a language learning model rather than a shape-and-sequence matching algorithm, in contrast to earlier work in the field, such as that of DeepMind. In addition to publishing its preprint publication on this study, Meta will also open. The goal of Meta's AI model is to predict the structures of the hundreds of millions of protein sequences in metagenomics datasets using massive language models as inspiration. Researchers will learn more about these proteins' functions and the chemicals they interact with by gaining an understanding of the forms that these proteins take.

The first comprehensive characterization of metagenomics proteins has been produced by us. According to Alex Rives, a research scientist at Meta AI, the database, which has more than 600 million predictions of protein shapes, is being made available as an open science resource. The model and protein database was made available to the scientific community and business, with the statement "This encompasses some of the least understood proteins out there."

According to Rives, evolution cannot decide between two linear positions on its own since the structure would crumble if the incorrect piece were placed in this location. This implies that protein sequence patterns can reveal details about the folded structure because distinct places in the sequence will co-vary with one another. That will demonstrate something regarding the protein's basic biological characteristics

In contrast, DeepMind's novel strategy, which debuted for the first time in 2018, heavily on a technique called multiple sequence alignment. To locate proteins that are linked to the one for which it is generating a prediction, it essentially searches through enormous evolutionary databases of protein sequences. Instead of making the prediction from this collection of numerous related proteins and examining trends, Rives explains that his team's method makes the prediction straight from the amino acid sequence. These patterns were acquired differently by the language model. This indicates that since we don't have to process this set of sequences or look for related sequences, we can drastically simplify the structure prediction architecture. According to Rives, these elements make their model faster than competing technology.

How was this model trained to be able to perform this task? It required two steps. In the beginning, they had to pre-train the language model across a trope of proteins that span the evolutionary history, exhibit a variety of structural features, and originate from many protein families. By blanking out some of the amino acid sequences, they employed a modified version of the Masked Language Model and asked the algorithm to fill in the missing information. According to Rives, "The language training is unsupervised learning; it is only taught on sequences." The model learns patterns across these millions of protein sequences by doing this.

The language model was then frozen, and a folding module was trained on top of it. They apply supervised learning throughout the second training phase. A collection of protein databank structures that have been contributed by researchers from around the world make up the supervised learning dataset. Then, using DeepMind's technology called AlphaFold, predictions are added to that. This folding module effectively outputs the 3D atomic coordinates of the protein [from the amino acid sequences] after taking the language model input, says Rives. That results in these representations, which are then reflected by the folding head into the structure.

The function of a protein's active site at the molecular level, which is the knowledge that might be very useful for drug development and discovery, is one application Rives envisions for this model. Additionally, he believes that in the future, new proteins may even be created using AI.