Problems of predicted values with real sequences

[wjiangal]

Hi there, I try to implement the shaker model with in-cell training set on some real sequences. I find out there are possible problems and really wonder why them happened:

1. The results among sequences are nearly the same, the variance of shape values is nearly zero.
2. When the input test sequence is long(>500), the shape value will drop to zero(10^-43). It doesn't make sense...

[smautner]

Hi Lucian,

1. similar values are expected. This is not a problem as downstream analysis (i.e. viennaRNA) usually scales the data to its needs.
2. I have not explored the relationship between sequence length and predicted value.

To solve 1 you could scale the values or perform feature selection (I assume, for example, that 'radius=0' features are not very good). To explore 2 one could use LOOCV to obtain values for the data in this repository to explore the impact of sequence length on the average predicted values.

Please update this thread with your findings :D