Repeated-measures RNA-seq analysis

Project Overview

This project focused on reanalyzing a repeated-measures RNA-seq experiment after correcting sample misassignment. Because multiple tissue or treatment conditions were collected from the same rats, the analysis required models that account for within-individual similarity rather than treating all samples as independent. I compared the original differential expression strategy with two repeated-measures-aware approaches: within-individual normalization followed by limma/duplicateCorrelation, and a mixed model approach using DREAM with rat identity as a random effect. Follow-up gene set activity scoring aimed to identify biological signals that were statistically significant and consistently separable at the sample level.

Experiment Design

→ Look at the “mouse” COLUMN, this is a repeated measure design. For example:

• m1 → sampleID Cont_A, Cont_B and Cont_C
• m2 → sampleID Cont2_A, Cont2_B, Cont2_C
• m3 → sampleID Cont3_A, Cont3_B, Cont3_C

Original Formula Design

~ group1

→ Failed to consider either BATCH or MOUSE COLUMN

Goal and Challenges

• The difference between Control - Treatment in certain Cancer. We previously design this comparison:

  Treat_B vs Cont_B
  Treat_A vs Cont_A
  Treat_C vs Cont_C

  → However, in Treat_B there were 3 different individual, and in Cont_B there were another 3 different individual

Challenge 1 - Repeated Individuals in Different Samples

Difference between individuals overwrote the difference between groups.

→ The original treatment-only comparison did not account for individual effects across the six rats in Treat_B vs. Cont_B.

Challenge 2 - Sample Correlation

• Samples of cancer B and C were highly correlated, regardless of treatment.
• But the sample of “Cont + cancer C” was very different from cancer A…potentially a outlier?

Download original image: Sample correlations across cancer and treatment groups

Summary

The comparisons of interest were:

• The difference of Treatment in cancer B
• The difference of Treatment in cancer C

Main Challenges were…

1. The samples of cancer B and C were highly correlated, regardless of treatment.
2. Difference between individuals might overwrite the difference between groups.

Method1 - Within Individual Normalization

Since cancer “A” was actually NO CANCER control, we use “A” as base to normalize expression of the other 2 cancer types (B and C) in each individual. Take mouse 1 as example

• m1 → sampleID Cont_A, Cont_B and Cont_C

  

For example, we normalized for m1 and m4 using log

Next, we tested the effect of Treat vs. Cont in cancer B

The treatment-only design becomes easier to explain:

~ Treatment

Validation

From left (before within individual normalization) to right (after within individual normalization)

Download original image: Sample relationships after within-individual normalization

→ This method did a great job of separating Treatments of cancer B and C by mitigating the effects of individuals

duplicateCorrelation

• Furthermore we considered the individual effect as gene expression of same individual must be similar → Cont_B_norm must be similar to Cont_C_norm, since they were all came from mouse 1.

• Consider the interaction

  ~ Cancer*Treatment

DEG Result of Method 1

• Only 1 DEG (ENSRNOG00000006478) was found. Was it special?

The trend of this DEG was down-regulated by the Treatment in cancer B?

• From the frequency histogram, we confirmed that such down-regulation trend was not common.

GSEA of Method 1

With few DEGs, I chose GSEA to highlight GO term instead of ORA.

Among all NES>0 GO terms, we focused on this term (GO:0060077).

Sample-level

We calculated 2 kinds of score to quantify the activation of each sample in this term:

GO term - Samples

We calculated further 4 kinds of score to judge

• whether the significance we observed was real between treatment

For researcher, if not significant → this term might be hard to validate in the following experiment

Method 2: DREAM