A Preliminary Study on the Prediction of Human Protein Functions - Foundations on Natural and Artificial Computation - page 336

From this ontology we selected all the GO terms which annotate from 20
up to 3'000 human proteins. Note that we discarded all the proteins that were
associated to the GO terms by electronic annotation, because we do not want
to rely on GO annotations that have been predicted on the basis of features
that are part of our input vectors (for example InterPro domains). Finally, we
obtained 1'930 GO terms representing output targets for the supervised models
to be trained. Since a protein can have more than a function, an input vector
can be associated to more than a GO term.

Table 1. Input features

Protein Feature Internet Location

InterPro domains www.ebi.ac.uk/interpro/

PROSITE patterns/profiles expasy.org/prosite/

Coiled coil domains www.ch.embnet.org/software/COILS form.html

Transmembrane domains In-house program following [5]

Transmembrane domains saier-144-21.ucsd/memsat.html

Transmembrane domains www.cbs.dtu.dk/services/TMHMM/

GPI lipid anchors mendel.imp.ac.at/gpi/gpi server.html

N-Glycosylation sites www.cbs.dtu.dk/services/NetNGlyc/

O-glycosylation sites www.cbs.dtu.dk/services/NetOGlyc/

N-terminal myristoylation mendel.imp.ac.at/myristate/SUPLpredictor.htm

N-terminal myristoylation expasy.org/tools/myristoylator/

Presence of poly-aminoacids In-house program

Sequence repeats andrade/papers/rep/search.html

Signal peptide cleavage sites www.cbs.dtu.dk/services/SignalP/
Tyrosine sulfation sites expasy.org/tools/sulfinator/
Subcellular location and cleavage site www.cbs.dtu.dk/services/TargetP/
Disordered regions anchor.enzim.hu/
Interaction with other human proteins string-db.org/

www.embl.de/

∼

4 Experiments

In the first series of experiments, we took into account 1930 distinct classifica-
tion problems, for which positive examples represent proteins annotated by a
particular GO term, while the negative examples are selected outside the branch
of its associated GO term. This classification strategy corresponds to the “ one-
versus-all ” approach. We performed both cross-validation trials and tests on the
independent testing set.