Network Tools
ML Predictor · train your own classifier
Give the predictor a set of genes you already believe in and a set you believe are
unrelated. It trains five classifiers on the network neighbourhood
of those genes together with expression and protein measurements, then scores every
gene in the genome with the trained models. Use it when you have a pathway, a GO term
or a co-expression module and want a ranked shortlist of the genes that behave like
its members.
Trainable genes
17,877
present in network, expression and protein
Genes per list
10minimum, 30 recommended
Typical runtime
15–35minutes on an idle server
Before you start
Both lists need at least 10 genes, and
30 or more is recommended — the models learn to separate
the two sets, so a handful of genes gives them almost nothing to work with.
A gene is only usable if it appears in the network, the expression matrix
and the protein matrix. The page drops the rest and tells you which
ones it dropped before the job starts.
A run takes about 15–35 minutes on an idle server. It
trains 20 rounds, and this compute node is shared — when other users are
running large jobs a round can take twenty minutes instead of forty seconds.
The job page counts the completed rounds, so you can always tell it is still
working rather than stuck. It gives you a download link, so you can close the
tab and come back — the link stays valid for 14 days.
Inside the zip: genome-wide predictions
(predresult.csv / .xlsx), the fitted models,
model performance (train_auc_f1) and feature importances.