Analysis workflow2022

Cross-Study Microarray Integration & Batch Correction Workflow

#Microarray
#CrossStudyIntegration
#ComBat
#PCA

I brought gene-expression data from separate studies into one traceable matrix. Using and adapting R/Perl routines, I aligned gene identifiers, tracked sample origin and built a batch-correction model that retained the comparison group.

Identifier alignmentMatrix integrationModel designDiagnostic inspection

GitHub

One example run

Select a stage to inspect

Check the sample relationships

I inspected the integrated inputs through a selected sample-level view.

Input: Before and after ComBat matrices
Output: One selected diagnostic output

R::prcomp · full-matrix input

Saved inputs → aligned matrix → batch-aware adjustmentHover, focus or tap for sample details

Method notes & example context

Contribution and scope. I used and adapted saved R/Perl routines to align identifiers, combine expression matrices and map batches and comparison groups. This is a workflow example, not a published study, an independently validated biological conclusion or a claim to have invented ComBat. The workflow starts with supplied gene-level array data; it does not process DNA genomes or raw sequencing reads.

Example inputs and date. The saved example combines supplied, already normalized gene-level inputs from two rat hippocampal microarray studies, with two Sham and three TBI samples in each study. Presentation labels Control/Condition correspond to those groups. The integrated input has 19,273 shared gene rows and ten sample columns. Both saved inputs contain the same unique gene identifiers; this example does not remove any study-exclusive rows or reorder mismatched input rows. Study A/B and S01–S10 are local display aliases. Archived outputs dated May 2022 support the approximate project date.

Adjustment model. The recorded ComBat call uses study origin as the batch variable, includes group in its model matrix and uses parametric priors. Including the group is the model design, not proof that all biological information is retained or that every unwanted difference is removed. Diagnostics should be interpreted with the source scale and study design in mind. See the official sva method reference.

Selected diagnostic. PCA was recomputed for this portfolio using R on each complete archived matrix, before and after saved ComBat adjustment. Samples are observations; each gene is centered across samples and is not scaled to unit variance. Each stage has its own PCA basis, explained-variance percentages and padded axis limits. Marker colors and shapes identify study; open/filled markers identify group. No connecting paths or sample-motion animation imply a common coordinate system. Sample size and PCA coordinates are available in mark details.

Process excerpts and units. The eight displayed genes are selected at evenly spaced ranks after sorting all shared gene identifiers, not by significance, biological importance or improvement after adjustment. Heatmap color uses one fixed asinh display transform across the two saved matrices. Untransformed values are shown in details; full precision is retained in the embedded data. Expression is kept in supplied units. No validated upstream log2 scale is established for this archived branch, so neither these values nor the saved model coefficients are relabelled log2 fold changes.

Downstream boundary. The historical differential-expression branch also calls normalizeBetweenArrays after ComBat and fits a limma/eBayes comparison. It is not the input used for the selected ComBat diagnostic here. The final differential-expression thresholds and enrichment outputs are omitted to keep the focus on integration and avoid repeating the RNA-seq and parallel-microarray projects. This preview does not refit ComBat or limma and does not add the unrelated microglial analysis.

All projects