MEDUSA input and output

MEDUSA takes as input an amino acid sequence in .fasta format. Using HHblits tool, it extracts evolutionary information and combines it with physico-chemical properties of each residue to predict a flexibility class of each position using a convolutional neural network.

MEDUSA performs four different predictions in terms of the normalized B-factor value (Bnorm). It creates two 2-class predictions following the nomenclature introduced by Schlessinger & Rost in 2006 [1]:

and multi-class predictions:

For all the predictions MEDUSA also returns the raw output of the network last layer allowing the user to estimate the reliability of the class attribution for each residue (confidence of prediction).

MEDUSA main algorithm

The general MEDUSA workflow is given in the figure below:

MEDUSA workflow

MEDUSA algorithm proceeds in eight main steps:

  1. Extract evolutionary information: MEDUSA finds homologs of the query sequence by HHblits search.
  2. MEDUSA filters the resulting Multiple sequence alignment (MSA) file using hhfilter
  3. The final MSA is translated into a probability profile using position specific score matrix: each position of the sequence is thus encoded by 21 numerical values corresponding to 20 amino acid types and gap.
  4. MEDUSA translates each amino acid to 58 numerical values, which encode its physico-chemical properties (using AA INDEX scheme).
  5. MEDUSA creates one hot encoding of each amino acid and adds a flag for the sequence terminus.
  6. Using a sliding window of 15 amino acids, MEDUSA creates input vectors for each sequence position for all the considered features.
  7. Different features are merged to create an input vector for the prediction of dimensions 15x100.
  8. Finally, the neural network performs binary and multi-class predictions and provides the general summary as well as flexibility prediction and confidence value for each amino acid.