Handling technical replicates in Stabl

[RMebane]

Hello,

We are hoping to use Stabl on a large spatial transcriptomics dataset, and we were hoping to get some advice on including features from multiple samples from the same patient. Should we aggregate features from these samples ourselves, or does Stabl include methods for including features from multiple replicates.

Thank you!

[xavdurand]

Hello @RMebane ,

Firstly, to provide a clear answer, Stabl does not include any internal mechanism to specifically handle multiple samples from the same patient.

1. If the multiple samples represent different timepoints:

   • Concatenate the data: Combine the data from multiple samples of a single patient, resulting in one sample per patient. Afterward, you can perform a standard analysis on this aggregated dataset.
   • Multivariate time analysis: Use the relationship between timepoints to create new features. Examples include:

   1. Calculating differences in features between timepoints.
   2. Extracting trends (e.g., slope or tendency) for specific features over time.
   3. Including engineered features as new variables (omics) or concatenating them with the patient’s original feature set.

2. If no assumptions are made about the relationship between samples: Treat each sample as independent for analysis. This approach assumes that the goal is to identify a global signature across all samples that is predictive of the outcome. For example, for timepoint data, this means finding a signature that is consistent across all timepoints.

Each approach has its merits and should be selected based on the specifics of your study design and the biological or clinical context of your data.

In the latter case, it is crucial to carefully handle the samples during the model training process, particularly when performing cross-validation or creating a training-validation split. Specifically, you must ensure that all samples from a single patient are either included entirely in the training set or entirely in the test set. This prevents data leakage and ensures that the model is evaluated on truly independent samples.

I hope it helps you in your study analysis design,
Xavier Durand