Keynote 1: Understanding interactomes by...

Keynote 1: Understanding interactomes by data integration
Soren Brunak

Protein-protein interaction networks - topology, temporal, phenotypic aspects. Text mining, predict 'disease genes'.

Starting with the cell cycle (an intrinsically dynamic system).

Data-cleaning is a pain - quality, formats, ...

Data following linear growth (?not in exponential phaze yet?)

Different databases provide very different data (did he mean knowledge?) about interactions.

Don't trust other people's 'components list' - they are wrong, missguided - re-analize raw data e.g. they found that 10% of yeast proteins are periodic (cell-cycle) - found 600 proteins.

Most cell-cycle control is a mix of non-cyclic, constant-expression components and up/down components - just-in-time activation of nearly-assembled complexes. Periodically present proteins have more degredation signals, more phosphorylation, post-translational modification, ...

Very little conservation of periodic behavior between human, pombe, cerevisiae, arabadopsis. However, complexes have the same members in each organisms, but have per-linneage choice of periodic vs stable expression. The important thing is that the complex is periodically active, not which component gives rise to this. Complexes and their behavior is conserved, not the details of the complex members.

This is very important for drug testing - need to make sure that drugs are tested in model organisms with compattible regulation to humans e.g. chimp responds to that immunosupressent like human does, macack doesn't (where the guys swelled up and had massive anafalactic shocks) - be scared wherever there is phosphorylation and tight regulation - it may be one of these stable/variable complexes.

Similar symptoms from diseases with genetic components may be due to them arrising from impairment with different proteins, but both interacting with the same complex. Combine phenotype data with linkage data with interactions.

Project this knowledge into a vector space - apply standard kernel magick (in their case possibly just dot product) to ccreate a scoring function.

Lots of good, solid work. The bioinformatics really does seem to be tracking the biology in this case.