mokume Optimization Discussion

[yueqixuan]

1. During feature normalization (features2peptides), it may be necessary to first apply a series of filters to peptides (features), and then compute peptide frequencies and normalization factors (e.g., global median), to ensure that the results are not affected by contaminants, decoys, or other artifacts. (corresponding code)

Suggestions from Tony:

1. Keep only intensity > 1. Reasoning: You'll notice when seeing spectra visually that intensities 0 to 1 typically appear to be background noise
2. Log2 transform intensity data (log transform intensity before any normalization / feature removal / summarization). Reasoning: Intensities are typically right-skew, which can lower the power of differential abundance analysis (also logs are easier to handle in statistics)
3. [Merge fractions] MSstats: if a feature is measured across multiple fractions for a sample, MSstats takes the maximum intensity among them.  Assumption here is that a peptide ion should elute dominantly in one fraction (and signal in other fractions is likely noise)
4. Batch effect correction may not be appropriate, especially when the definition of a batch is unclear here.  Also since each run is associated with a single biological condition, treating each run as a batch would remove the biological effect of interest.

[ypriverol]

Issue tackle in e6d0367

[yueqixuan]

1. Add Run-level normalization
The current global normalization is performed at the sample level (sample_accession) (possibly involving get_median_map(), analyse_sdrf(), feature.iter_samples(), etc.). Should we add normalization at the Run level instead, processing by reference_file_name?

2. A small issue: unify output file formats.
In features2peptides: writer_parquet_task or write_csv_task
In peptides2protein: to_csv or to_parquet
That is, when running features2peptides, there is no need to provide a file extension, whereas when running peptides2protein, a file extension is required.