adding new samples

[I-am-Shir]

I understood that in order to perform downstream analysis on multiple samples I must run them together so the annotation will include all of them, is that correct?
Could you clarify what the recommended workflow is when additional samples need to be added later?
In my case I’m working with many large datasets, so re-running the full pipeline on all samples each time a new one is added would be computationally expensive. Is there a recommended strategy for incrementally adding samples while keeping the isoform set consistent across runs?

Sincerely,
Shir

[mschilli87]

@I-am-Shir: This is a very good question.

I think there are generally three solutions:

1. Use the same annotation for all samples based on all samples.
2. Use a different filtered annotation for each sample.
3. Use a fixed filtered annotation for all samples.

AFAIK, only the 1st option is supported as of now.
However, @Franzx7 is working on a new version for nf-core that might already support different modes. If not, this might be a great time to discuss if it could/should.

As you pointed out, if you already ran SCALPEL on 50 samples and now you want to add 5 new ones, the 1st option would require re-analyzing all of them.
However, while the 2nd option would elimilate this problem, an additional filtering step would be required later when combining the results from several samples. The only strategy that comes to my mind is to use the intersection of all per-sample annotations. This could lead to the 'loss' of certain isoforms/genes compared to the 1st option, require more disk space (to accommodate all the annotations), but make re-runs with additional samples faster and would make per-sample results independent (i.e. the quantification results for any of the 50 samples analysed before won't suddenly change just because a newly added sample resulted in the inclusion of some new isoforms).
Option 3 would be a compromise between the first two: You could run option 1 for your first 50 samples, and later decide that the expected differences of the 5 new samples do not justify rebuilding the annotation, but instead provide the annotation created by your first run as 'fixed' annotation for the 2nd run. This is technically challenging as SCALPEL would somehow need to 'know'/figure out that this fixed annotation is the same that those first 50 samples had been analysed with already to be able to actually skip the re-run. The advantage would be that you could also define your annotation in a different way (e.g. based on bulk (long read) sequencing). Note that in this case, if you'd have specified your 'custom' annotation for the first run already, SCALPEL must still realize this to skip re-running those 50 samples on the second run. Because you'd probably expect it to re-run all samples when you do a re-run with a different 'fixed' annotation than before.

Thinking of it while I write this wall of text (apologies), maybe the best option would be

4. explicitly allow re-running without re-filtering the annotation.

This option would not provide the full flexibility of option 3, but simply make the annotation depend on only those samples, for which quantification results already exist. That seems straight-forward to implement and would result in automatically determining both, the annotation and the previous quantification results as up-to-date, and implicitly using the existing annotation from the first run for the quantification of the additional samples.

However, this would still introduce a bias towards the 'first' samples. Option 1 will always remain the only one that would make sure the results truly capture all information provided by the (final) full dataset.
Also, reproducing your results later on would require running SCALPEL it in two batches (your first, and your second plus all potential later ones) as well.

@Franzx7: Please share your thoughts on this as well. Maybe you already have a plan for this use case? Or is this something you never thought about during the design until now?

@I-am-Shir: Please let me know what you think and how common it is for you to get data piece-by-piece.

[I-am-Shir]

@mschilli87 Thanks so much for the thorough breakdown!

My situation definitely leans toward Options 3 or 4. My use case is often iterative: a user runs SCALPEL on an initial set of samples, and based on those results, decides to pull in additional relevant samples to strengthen the analysis.
Being able to use a 'fixed' annotation from the first run would be ideal. It not only saves time but, as you mentioned, ensures the quantification for the original samples stays consistent and doesn't 'shift' just because new data was added.
But it could definitely be helpful for those who are using the tool alongside the creation of their data.

It’s definitely a common scenario in my workflow to have the dataset grow as the hypothesis evolves.