This was achieved by a ‘rolling input subset’ approach which involved
training the data containing inputs across a 3 kDa mass range, and then shifting
this along 1 kDa at a time, in order to create a new data block. For
example, the first block contained data within the 3–6 kDa mass range, the
next block ranged from 4 to 7 kDa, the next from 5 to 8 kDa and so on up to
30 kDa. Each model was then trained over 50 random training/test/validation
sub-models and relative importance values for each individual input were
recorded so that they could be ranked according to their influence upon correct
sample assignment. Relative importance values were calculated for each
input by:
(i) Multiplying the absolute weight values of the network from the input
node to the first hidden node with the absolute weight values of the
network between