Allow Multi loops in interaction region

[PatrickRWright]

[martin-raden]

@PatrickRWright think I need a depiction for that to get the idea...

[PatrickRWright]

Something like figure 14 in my thesis.

[martin-raden]

planned (while the intramolecular structure is incorporated using the ensemble energy of the enclosed part) but most likely not available before v2.1

have to discuss:

• maximal distance between interactions that include intramolecular base pairs (critical for performance)
• whether or not each interaction segment has to show an individual seed region (critical for sanity and performance)
• how exactly do we define an optimal interaction in that case and how to compute efficiently

[martin-raden]

Experimentally verified example in accessfold paper and according supplementary material

[martin-raden]

position-wise ED values (used in AccessFold) directly enable such predictions.

advantages:

• easy to incorporate without runtime losses

disadvantages

• ED values are too high since positions are treated independently
• in AccessFold countered with a general scaling factors to all such ED values
• seems to be not very well performing (Umu+Gardner,2016, in press) preprint @PatrickRWright do you know this paper already? check the preprint!

[martin-raden]

bistaRNA (Poolsap et al., 2011) enables multi-site predictions but with overall low performance according to (Umu+Gardner, 2016)

I checked the manuscript: they apply an MEA algorithm (no NN-energy-model with stacking etc.) using intermolecular bp-probabilities from an RNAduplex-hack. Furthermore, they consider from the (smaller) query RNA only accessible regions (above threshold) for interaction prediction and screen via sliding window for interactions within the target. Finally, given interaction sites, they identify the best interaction by applying MEA on the target RNA in the presence of one or more interaction sites, to incorporate accessibility of the target.

[martin-raden]

heuristic to reduced time consumption (Rolf):

• store for each sequence position i the best k Qm entries with left boundary i
• during interaction prediction: for each position i check only the best k ML-inclusions
• results in O(1) time ML-case if k fixed and small

open questions:

• what k sufficient? maybe enable length-dependency via k=log(n)
• do we need max Qm spanning to avoid too large inclusions (that enable higher energies)?

[martin-raden]

check the IRIS pub for further examples

Dmitri D. Pervouchine. IRIS: intermolecular RNA interaction search. Genome Inform, 15(2):92{101, 2004.