Add RTX prediction notebook

[jaclyn-taroni]

In this PR, I am adding a notebook that performs a supervised ML analysis using the RTX data. I attempt to predict response labels (nonresponder and two types of responders: tolerant and nontolerant) from the baseline samples expression data using multinomial LASSO. Unfortunately, there are only 36 baseline samples with response labels. With this sample size, I use leave-one-out cross-validation (LOOCV) and do not have a hold-out set. I am comparing three data entities if you will:

• Expression data -- gene-level measurements that have been filtered to only include genes that are in the recount2 model
• recount2 LVs -- this RTX data in the recount2 PLIER model latent space
• RTX PLIER LVs -- dataset-specific model

My hypothesis was that the multiPLIER LVs (recount2) would outperform the expression data in this prediction task. The results in this notebook refute that hypothesis. A few notes & observations, although I have no intention of further exploring this avenue given the limitations:

• The range of the recount2 LVs is smaller / is "flatter" than the other entities (can be seen at the bottom of the notebook). It's possible that scaling the features would provide a better comparison between the entities. In general, there's probably room for improvement in training/tuning, etc.
• My original hypothesis wrt multiPLIER LVs is that this approach would generalize better than models trained on gene-level expression. I can't test that with this dataset + sample size. So I note this here as a future direction.

Notebook html: 25-predict_response.nb.zip

[huqiwen0313]

Looks good to me. Some minor comments.

[huqiwen0313]

Is it worth to tune lasso parameter to achieve the best performance using caret::train ?

[jaclyn-taroni]

Totally agree that this would be my next move if I was going to work on this a bit more!

[huqiwen0313]

RTX LVs have a prediction accuracy equal to 1... seems like overfitting ?

[jaclyn-taroni]

Yes, absolutely

[huqiwen0313]

I agree, the LVs values may affect prediction, do you think scale them to the same level will make the accuracy looks more similar ?

[jaclyn-taroni]

It's possible, but I think there is not much to be done with this sample size (especially given the overfitting as you point out above!)

[jaclyn-taroni]

Thanks for the comments @huqiwen0313. I agree with them and am glad they will be recorded here! I am going to merge this the way it is because I will not investigate this particular avenue further until I have a larger or more appropriate dataset.