WorkPlan.md

Preliminary workplan

Please, add any modification/suggestion you wish.

Preliminary experiments to evaluate the predictability of the TCDB classes

Experiments using BLAST data and the proteins more than 1 annotation.

We should exlude TCDB classes having only 1 annotation, thus resulting in about 2436 classes (according to tcdb.1). The gzipped text file tcdb1.annotations.levels1234.MoreThan1.txt.gz in the directory Experiments/Data is a matrix with 12546 proteins (rows) and 2436 TC classes (columns): only classes with more than 1 annotation are included. In this way, in principle, using leave-one-out we could try to predict all the classes.

• At first we could use BLAST data just prepared by Su to understand whether we can predict something.

• As a second step we could try to add Interpro features.

• Metrics: AUC per class and precision at different level of recall (e.g. at recall levels varying from 0.1 to 1 at 0.1 steps). Evaluation: leave-one-out (loo). Results averaged across classes, across levels, and per class.

• Preliminary flat methods that could be considered:

  1. BLAST (best hit?) as (strong) baseline
  2. Nearest Neighbour
  3. Semi-supervised flat learning methods: GBA and RANKS (Giorgio)
  4. linear SVM and/or logistic regression or some other linear method.

Su, could you perform task i, ii and iv? (of course if you agree on this set-up :-)

Experiments using InterPro features

TO DO

Definition of other non-InterPro features useful for the prediction of transport proteins (e.g. TMS data)

TO DO

Design of Tree structured output methods for single-path predictions

TO DO

Design of hierarchical ensembles for single-path predictions

TO DO

Status API Training Shop Blog About Help