Attention based neural networks for protein structure prediction

  • Author(s) / Creator(s)
  • Protein structure is one of its most important characteristics. It is through its structure that a protein is able to interact with each other and different molecules. Therefore, the ability to properly predict how the 3D structure of a protein is given configures a key advancement in the process of discovering new forms of fighting infectious diseases or even tackling environmental problems. Machine learning has been proving a great ally on the path to achieve an effective estimation of the protein structure down to presenting atomic precision on its angles of fold. New methods have been developed and revolutionized the field. The topic of the research is the use of a simplified Attention based neural network for predicting and visualizing protein angles from amino acid sequences with precision. Attention networks work by applying different weights on different parts of the input in order to retrieve which is the information that the model should focus its attention on. This approach can be applied to properly identify the critical interactions between different parts of the protein sequence that contribute to its folding. These important parts of the sequence can include amino acids that form critical interactions, such as hydrogen bonds, without necessarily having a score function for hydrogen bonding, wich makes it more efficient on a limited dataset while being able to cope with the variety and complexity of the structural data, as assessed by benchmarking on CASP13 and CASP14 datasets. Overall, attention based networks have proved to be a promising lead of research on protein folding, with a great potential to enhance our understanding of the structures and interactions of proteins. The intrinsic challenge of experimental structure determination has prevented an expansion in our structural knowledge before, but these new methods, when allied to a large and well-curated database of structures and sequences can provide grounds for a fast evolving body of knowledge on bioinformatics and other biophysical problems.

  • Date created
    2023
  • Subjects / Keywords
  • Type of Item
    Conference/Workshop Poster
  • DOI
    https://doi.org/10.7939/r3-zeak-sv57
  • License
    Attribution-NonCommercial 4.0 International