#Notes
Protein blocks with a lenght of five residues, we get torsions angles (phi, psi) wich is dihedral angles phi is angle between N-C psi C-C and omega is angle between amino acid (allowing rotation between aa, can go between 0 and 180)
In a protein block we do not consider first and last angles because they are linked to next or previous protein block. So we consider 8 angles Because of overlap and sliding window two protein blocks following are sharing 4 amino acids. So, in a protein sequence of 15 aa, we consider protein blocks
For example, in a protein of 10 aa:
- First protein block is psi1 to phi5
- Second pb is psi2 to phi6
- Third pb is psi3, phi3, psi4, phi5, psi5, phi6, psi6, phi7,psi7
PB-torsion file generation
Algorithm: An unsupervised algorithm classification. Using data it train to classify PB PBA: phi psi phi psi omega phi omega phi omega phi
- Each cells have a value. For each value that we want to classify, In our dataset there is no overlap because it s a dataset with only PBA, coming from different places.
1- Create a random matrix with 16 random values. Let s calling it v 2- Pick one random value from training dataset and compare it to each value of cell (=neuron). It compare by calculating RMSD. With put value in closest cell. This neuron will be the most educated, and neuron next to him will learn a bit less, decreasing with radius. Choosed neuron will be k w_next= w^k*(t+1)= w^k*(t) + [v-w^(k(t))]* eta e^(-v/2r²*(r^k-r^next)²) 3- Loop until results are constant
Summary:
- Obtaining datas from PISCES
- DSSP to obtain psi and phi angles
- Generate PB dataset composed by 16 protein blocks
- a, b, -> c alpha-N-cap
- d -> beta
- e, f -> c-cap-beta
- g, h -> coil
- m -> alpha
- n,o,p -> c-cap-alpha
For protein block a we have a training dataset.
Algorithm: random vector file (-180,180), Generate initial map. Using training dataset, Calculate RMSD and put it in right cell