Amino acid chain fragments predict protein function.

Just a few months ago, DeepMind launched alphaFold, a system called AlphaGo, which predicts and generates protein 3D structures.
, researchers from MIT recently developed a new research model that directly predicts how amino acid chain fragments determine protein function.
discovery could help researchers design and test new proteins for drug development and biological research.
we all know that protein is a large and complex substance necessary to sustain our lives.
protein exactly what functions it can perform depends on its unique three-dimensional structure.
understanding the structure of proteins is therefore a very important part in predicting their response to certain drugs.
However, despite decades of research and the aid of multiple imaging techniques, we have learned only a small fraction of the numerous protein structures, and many unknown structures have yet to be revealed.
in response to this, researchers from MIT developed a way to "learn" the easily calculated characterizations at each amino acid position in the protein sequence.
, the researchers then entered these characterizations into a machine learning model, which allowed them to directly predict the function of individual amino acid fragments without any data on protein structures.
first, the researchers used about 22,000 proteins from the Protein Structure Classification Database (SCOP) to classify them according to the similarity of structure and amino acid sequences, and trained machine learning models.
for each pair of proteins, the researchers calculated a structural similarity score based on their SCOP category.
, the researchers then input random protein structure pairs and their amino acid sequences into machine learning models, which are represented by encoders, called embedding. each
contains similarity information for a pair of amino acid sequences.
the model aligns the two embedded, and then calculates a similarity score to predict the similarity of the three-dimensional structure of the protein it represents.
, the computer then compares this score to the real SCOP similarity score and sends a feedback signal to the encoder.
if the model's predictive score is far from the true score, some adjustment will be made. At the same time,
the model predicts each embedded "contact map", the distance of each amino acid to the other amino acids in the protein, compares the predicted contact graph with a known contact graph from SCOP, and then sends a feedback signal to the encoder.
this step helps the model better identify the exact position of amino acids in the protein structure, thus providing a better understanding of the function of each amino acid.
for an amino acid chain, the model can generate an embedded one for each amino acid position in a three-dimensional structure.
, the machine learning model can then use these sequences to embed and accurately predict the function of each amino acid based on its predicted 3D structural contact map.
in an application example, the researchers used the model to predict which proteins could pass through the cell membrane, and the predictions were more accurate than existing advanced models.
next, the researchers plan to apply the model to more predictive tasks, such as figuring out which sequence fragments can bind to small molecules, which is critical to drug development.
researchers say the study could eventually be applied to human health and pharmacogenomics because it could help detect harmful mutations that damage protein structures.
References: 1. Bepler, et al., (2019). Learning protein sequence embeddings using information from structure. ICLR 2019, arXiv: 1902.08661.2 MIT CSAIL's AI predicts a protein's function sfrom of amino acids. Retrieved March 26, 2019, from learns how individual amino acids syds sein function. Retrieved March 26, 2019, from.Source:Supplied.