Prediction of Protein Ligand Binding Affinity Using Deep Networks

Gholamzadeh Lanjavi, Atena | 2021

1037 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 54197 (19)
  4. University: Sharif University of Technology
  5. Department: Computer Engineering
  6. Advisor(s): Kalhor, Hamid Reza; Motahhari, Abolfazl
  7. Abstract:
  8. Protein-ligand binding affinity is extremely important for finding new candidates in drug discovery and computational biochemistry. One of the physical characteristics for protein ligand interactions has been dissociation constant (KD) which can be obtain experimentally. However, there have been tremendous efforts to predict KD using modeling and computational approaches for protein-ligand interactions. In this project, we have exploited Convolutional Neural Network (CNN) model based on KDeep design, PDBBind version 2016 refined set training data, and examining it with KDeep core set test data. In order to modify KDeep,instead of 24 rotations (0, 90, 180 and 270 degrees in selection of two pages from three pages), only two random rotations were applied. Our results surprisingly gave rise to a much better Pearson’s coefficient of 0.96 (0.14 better than KDeep model) and Root Mean Squared Error of 1.25 (0.02 less error than KDeep model). One of the reasons for enhancing the results obtained in our model, might be due to much lesser rotations but in a random manner, allowing neural network to learn faster with less data.Additionally to better analyze protein-ligand interactions, we have performed systematic analysis. Protein-ligand interactions based on their affinity values (KD) were divided in to two efficient and non-efficient groups; subsequently systematic analysis were performed on types of amino acids residues involved, types of noncovalent binding, and role of water in the protein-ligand interactions (PDBBind version 2019 data). Our results demonstrated that amino acid tyrosine was the most abundant residue from the proteins involved in the protein ligand interactions. The role of water molecules turned out to be significant as to be found in %92 of interactions. Moreover, the amino acid residue most often involved in Hydrogen bonding was identified as histidine
  9. Keywords:
  10. Convolutional Neural Network ; Proteins ; Ligands ; Pearson’s Coefficient ; Protein-Ligand Binding Affinity ; Root Mean Squared Error

 Digital Object List


...see more