Multi-omics experimental design in animal production

Practical 15: Genomics, systems biology and multi-omics

Learning objectives

Define the biological question and experimental unit; plan matched assays and balanced processing; diagnose a rank-deficient design.

Biological question and data

How should an individually allocated diet study collect molecular profiles and later performance measurements?

The hypothetical design allocates diet to individual animals. Pen-allocated diets would require a different unit and replication plan. The practical specifies metadata and tests design identifiability; it does not establish sample size or a collection protocol for a real assay.

No biological download is required. Numerical inputs are hypothetical, not observed animal data. The runnable examples use base R. Count and abundance units are specified where used.

1. Build a matched sample sheet

Keep animal, diet and batch separate from the assay-specific specimen identifiers.

Show R code 
plan <- data.frame(animal=paste0("a",1:12),
 diet=factor(rep(c("control","supplement"),6)),
 batch=factor(rep(c("A","A","B","B"),3)),
 tissue="liver",collection_day=14)
plan$rna_id <- paste0(plan$animal,"_RNA_d14")
plan$protein_id <- paste0(plan$animal,"_protein_d14")
stopifnot(!anyDuplicated(plan$animal),!anyDuplicated(plan$rna_id))
plan
   animal       diet batch tissue collection_day      rna_id      protein_id
1      a1    control     A  liver             14  a1_RNA_d14  a1_protein_d14
2      a2 supplement     A  liver             14  a2_RNA_d14  a2_protein_d14
3      a3    control     B  liver             14  a3_RNA_d14  a3_protein_d14
4      a4 supplement     B  liver             14  a4_RNA_d14  a4_protein_d14
5      a5    control     A  liver             14  a5_RNA_d14  a5_protein_d14
6      a6 supplement     A  liver             14  a6_RNA_d14  a6_protein_d14
7      a7    control     B  liver             14  a7_RNA_d14  a7_protein_d14
8      a8 supplement     B  liver             14  a8_RNA_d14  a8_protein_d14
9      a9    control     A  liver             14  a9_RNA_d14  a9_protein_d14
10    a10 supplement     A  liver             14 a10_RNA_d14 a10_protein_d14
11    a11    control     B  liver             14 a11_RNA_d14 a11_protein_d14
12    a12 supplement     B  liver             14 a12_RNA_d14 a12_protein_d14
Show R code 
xtabs(~diet+batch,plan)
            batch
diet         A B
  control    3 3
  supplement 3 3

Each animal has two linked assays from a specified tissue and day. The sample sheet still needs documented collection/preservation and assay-specific handling.

2. Check estimability

Contrast a crossed design with perfect treatment–batch confounding.

Show R code 
X <- model.matrix(~diet+batch,plan)
bad <- plan;bad$batch <- bad$diet
Xbad <- model.matrix(~diet+batch,bad)
rbind(crossed=c(columns=ncol(X),rank=qr(X)$rank),
 confounded=c(columns=ncol(Xbad),rank=qr(Xbad)$rank))
           columns rank
crossed          3    3
confounded       3    2
Show R code 
stopifnot(qr(X)$rank==3,qr(Xbad)$rank==2)

Full rank permits separate estimation of these fixed-effect columns. It does not certify power or remove unmeasured confounding.

3. Plan decision timing and missingness

Specify the predictor window and outcome window before inspecting results.

Show R code 
decisions <- data.frame(item=c("genotype","day14 assays","future performance"),
 available_day=c(0,14,28),role=c("baseline covariate","candidate day14 predictor","outcome"))
decisions
                item available_day                      role
1           genotype             0        baseline covariate
2       day14 assays            14 candidate day14 predictor
3 future performance            28                   outcome
Show R code 
stopifnot(all(decisions$available_day[decisions$role=="candidate day14 predictor"]<=14))

The future outcome can evaluate a day14 prediction; it cannot be included as input to that prediction. Predefine how absent assays and failed QC affect the analysis population.

4. Compare allocation with observation count

A pen-allocated experiment has fewer independent treatment units than animal rows. Construct a hypothetical design with four pens and record both levels.

Show R code 
pen_plan <- data.frame(pen=rep(paste0('pen',1:4),each=5),
 diet=rep(c('C','C','T','T'),each=5))
pen_plan$animal <- paste0('animal',seq_len(nrow(pen_plan)))
c(animal_observations=nrow(pen_plan),allocation_units=length(unique(pen_plan$pen)))
animal_observations    allocation_units 
                 20                   4
Show R code 
unique(pen_plan[c('pen','diet')])
    pen diet
1  pen1    C
6  pen2    C
11 pen3    T
16 pen4    T

5. Plan a design audit

Check allocation unit, group counts within processing batches, specimen timing and missing-assay contingencies before collecting data. A full-rank fixed design is one useful check but does not assess pen dependence or generalization across farms.

Predefine a primary contrast and relevant exclusion rules. Biological allocation, blinded handling and randomized processing order protect different aspects of the study. A variance and effect-size sensitivity analysis can support planning, but does not replace a clear unit and sampling hierarchy.

Interpretation challenge

Would sequencing each animal twice double the diet replication in the four-pen design?

No. Duplicate libraries add technical evidence for the same specimens; there are still four independently allocated pens. The analysis must respect that hierarchy.

Interpretation and limits

Randomization and blocking, tissue/time, sex, farm, batch, handling and exclusion rules need a justified study-specific plan. Technical aliquots are not independent animals. Use NC3Rs design guidance for the design principles; the invented sheet is not a laboratory sampling instruction.

Exercises

  1. If diet is assigned to pens, what changes?
  2. Why can adding animals within the same confounded batches fail to solve the problem?
  3. What must be decided before selecting an integration method?
  1. Pen-level allocation and dependence become central; sufficient independent pens are needed.
  2. The treatment and batch columns remain dependent.
  3. Question, samples, matching, time, measurement scales, missingness, covariates and validation target.

Continue learning

B01 Foundations, B02 Design and B18 Synthesis.

Continue with multi-omics quality assessment. Return to the practical index or the unified map.