How to detect alternative splicing when there is no replicate

[metheuse]

I'm working on a pilot RNAseq project which compares paired-end reads between two samples, and there is only a single experiment in each sample.
I've been searching for tools to handle alternative splicing and found things like DEXseq and SplicingCompass. But both of them seem to require replicates. I'm wondering how I can use them (or any other tools) on my data.
I understand that without replicates, the results would lack statistic meaning, but since this is just a pilot project, the thing is to find an appropriate pipeline to analyze the data.
By the way, could I randomly separate the reads into two sets and treat them as pseudo replicates? I know this is definitely against the definition of biological replicates, but is it a way that would worth a try?
Thanks for any suggestions.

[syfo]

the thing is to find an appropriate pipeline to analyze the data.

That's the point: an appropriate pipeline requires replicates.

could I randomly separate the reads into two sets and treat them as pseudo replicates? I know this is definitely against the definition of biological replicates, but is it a way that would worth a try?

This is what I would do to set up a pilot pipeline.

[metheuse]

Thanks! By the way, do you know how to randomly separate the reads into two groups? What I think of is to shuffle the reads and then pick the first half and the last half. Should I separate them equally?

[Simon Anders]

The wole point of a pilot experiment is to figure out whether it is worthwhile to do the real experiment, i.e., whether you have a chance to detect interesting biology once you scale things up.

To address this question, you have to know effect size (how big are the typical differences between control and treatment) and noise strength (how big are meaningless differences that also appear between two control or two treatment samples). Without performing at least one condition in at least duplicates, there is no way of knowing how much noise you are facing and hence how strong effects have to be in order to be meaningful.

So, I simply don't get it: What is the point of unreplicated pilot experiments, and why do people waste their time on them?

If your aim is solely to tets your pipeline, simply take some published data. If you split your unreplicates data in two halves, you will only test the unrealistic situation of nearly every gene with more than a few dozen read counts being called significant.

[metheuse]

Thanks for your reply, Simon.
I perfectly understand your point.
This is not my own project. I'm just in a position to help some biologists to deal with bioinformatics stuff. The unreplicate data is what they have right now and they want me to explore the tools that can process the data. I told them there is no way to determine significance without replicates and they absolutely understand it. They will sequence more replicates as the next step.
Simply, I just want to construct a pipeline that can take the data and output results. But I understand I can't well evaluate the pipeline since the result would lack statistic meaning.

[syfo]

At least you can set up a bioinformatics pipeline in terms of format, data processing etc. I would just start by splitting the fastq files in two. Shuffling, random subsampling etc sounds better but as Simon pointed out the biological variability cannot really be simulated anyway. I understand your approach though, it makes sense to ensure your workflow can actually process the data the way it comes from your lab. Published data from another source might differ in many ways (format, naming system, etc).