Problem With Limma On Two-Color Factorial Design: Coefficients Not Estimable
1
3
Entering edit mode
13.0 years ago
Stephen 2.8k

I'm running into a problem where my coefficients are not estimable when using Bioconductor/limma on a two-color factorial design for a microarray analysis.

I have microarray data I downloaded from Array Express using the ArrayExpress function in the ArrayExpress package. I managed to convert the NChannelSet object into an RGList object that I can use in limma using a tip I found in the archives (https://stat.ethz.ch/pipermail/bioconductor/2009-September/029705.html). After some bg subtraction and normalization, I have an MAList object with 32 arrays (red and green) by 34,944 probes.

The experiment consists of two cell types (WT or KO), by 4 treatments (control, CD70, CD80, CD70+CD80), at four different time points (2, 4, 8, 14 hours), with a dye-swap. I'm mostly interested in genes that are differentially expressed when WT cells are hit with CD70+CD80 versus CD80 alone.

I'm trying to create a targets file so I can use modelMatrix to create my design matrix. But I have a multifactorial design using a two color dye-swap design, and I'm not sure how to specify this in the targets file. The limma manual has information about factorial designs, but no examples for two-color experiments. I have the following factors:

1: Celltype: WT or KO 2: Treatment: control (x), CD70, CD80, or CD70+CD80 (CD7080). 3: Timepoints: 2, 4, 8, and 14 hours.

I tried collapsing all of factors and levels into a single string in my targets file:

> targets
       array          Cy3            Cy5
1  array3329    WT.CD70.2         WT.x.2
2  array2675  KO.CD7080.8    WT.CD7080.8
3  array2242  WT.CD7080.2    KO.CD7080.2
4  array3328       WT.x.2      WT.CD70.2
5  array3310       WT.x.8      WT.CD70.8
6  array2246  KO.CD7080.2    WT.CD7080.2
7  array3337  WT.CD7080.4      WT.CD80.4
8  array3323 WT.CD7080.14     WT.CD80.14
9  array2673    KO.CD70.8      WT.CD70.8
10 array1938 WT.CD7080.14   KO.CD7080.14
11 array2240    WT.CD70.2      KO.CD70.2
12 array3336    WT.CD80.4    WT.CD7080.4
13 array3322   WT.CD80.14 WT.CD7080.14.2
14 array2674  WT.CD7080.8    KO.CD7080.8
15 array3321   WT.CD70.14        WT.x.14
16 array2241    KO.CD70.2      WT.CD70.2
17 array2597  KO.CD7080.4    WT.CD7080.4
18 array3313  WT.CD7080.8      WT.CD80.8
19 array1939 KO.CD7080.14 WT.CD7080.14.2
20 array3335    WT.CD70.4         WT.x.4
21 array2672    WT.CD70.8      KO.CD70.8
22 array3320      WT.x.14     WT.CD70.14
23 array3334       WT.x.4      WT.CD70.4
24 array3311    WT.CD70.8         WT.x.8
25 array3331  WT.CD7080.2      WT.CD80.2
26 array1941   KO.CD70.14   WT.CD70.14.2
27 array2588    WT.CD70.4      KO.CD70.4
28 array2596  WT.CD7080.4    KO.CD7080.4
29 array3330    WT.CD80.2    WT.CD7080.2
30 array3312    WT.CD80.8    WT.CD7080.8
31 array1940 WT.CD70.14.2     KO.CD70.14
32 array2593    KO.CD70.4      WT.CD70.4

I then created a design matrix, where the reference group is WT cells, untreated, 2-hour timepoint.

design <- modelMatrix(targets, ref="WT.x.2")

This creates a 32x25 design matrix, but when I fit the model, I get a "Coefficients not estimable" error/warning:

> fit <- lmFit(d, design)
Coefficients not estimable: WT.CD70.14.2 WT.CD80.14 WT.CD80.2 WT.CD80.4 WT.CD80.8 WT.x.14 WT.x.4 WT.x.8 
Warning message:
Partial NA coefficients for 34944 probe(s)

Any help with how I can specify this multifactorial time-course design in a two-channel dye-swap experiment would be greatly appreciated. As I said, I'm most interested in the conditions on arrays 8-9, 13-14, 19, 26, and 30-31, where I'm looking at WT cells treated with CD70 and CD80 versus CD80 alone.

Thanks in advance for any help!

r bioconductor microarray • 5.4k views
ADD COMMENT
3
Entering edit mode
13.0 years ago
Ido Tamir 5.2k

It looks like the wt.x.2 reference you wanted to have is not connected to the subset you are really interested in which is then not estimable.

If I followed the graph correctly then its:

  • wt.x.2 / wt.cd70.x.2
  • wt.cd70.x.2 / ko.cd70.x.2

Then it seems to end.

Unless you go for "separate channel analysis for of two color data" chapter 9 in limma user guide, it will be difficult to answer the questions you are interested in. Maybe its easier to look only at the subset you are interested in, and forget about the rest.

ADD COMMENT

Login before adding your answer.

Traffic: 1400 users visited in the last hour
Help About
FAQ
Access RSS
API
Stats

Use of this site constitutes acceptance of our User Agreement and Privacy Policy.

Powered by the version 2.3.6