Challenges in data generation and harmonization

[jaybee84]

1. add seminal work references
2. add description

[allaway]

Techniques to manage disparities in data generation are required to power robust analyses in rare diseases: Rarity of patients leads to heterogeneity in sample collection, causing disparities in the data. We will discuss how rigorous normalization and methodologies capturing sample-wise gene-set level information can help appropriate integration of disparate data points to power machine learning approaches11–13.

[allaway]

There's a lot to possibly talk about here, so let's break this down by data type:

Gene expression: assessing heterogeneity

• MDS
• PCA
• tSNE
• UMAP

Gene expression: correcting heterogeneity

• batch effect correction methods (regression-based modeling of batch effects - ComBat, limma)
• gene expression heterogeneity analysis methods (sva)
• among many others

Variant data: assessing heterogeneity

• dimensionality reduction methods on binarized variant data

Variant data: correcting heterogeneity

• Filtering to enrich for computationally predicted high-impact mutations (SIFT, Polyphen, etc)
• Filtering out common variants in the population (ExAC, dbSNP VAF < 0.01)
• Reprocessing raw data with a common pipeline
• Comparing to targeted sequencing panels that generally have a better detection rate
• Ensembling multiple variant calling tools. Anecdotally, we observed that a tumor variant detected on a targeted sequencing panel was only detectable in the whole exome sequencing data from the same sample using DeepVariant- this manuscript is accepted at Sci Data so need to revise the doi before submitting
• Methods like SHEAR for structural variant detection

[jaybee84]

adding this paper for consideration in the high-impact mutation prediction point (along with SIFT and Polyphen) ... seems like a good resource for diseases with multigenic possibilities!

(conflict of interest disclaimer: I may have a soft spot for ensemble RFs :P )

[allaway]

An important thing to acknowledge: batches/processing cores/institutions are often confounded by the biological variables, like tumor type, disease state, etc.