Presentation on theme: "Accounting for variation in designing greenhouse experiments Chris Brien 1, Bettina Berger 2, Huwaida Rabie 1, Mark Tester 2 1 Phenomics & Bioinformatics."— Presentation transcript:
Accounting for variation in designing greenhouse experiments Chris Brien 1, Bettina Berger 2, Huwaida Rabie 1, Mark Tester 2 1 Phenomics & Bioinformatics Research Centre, University of South Australia; 2 Australian Centre for Plant Functional Genomics, Adelaide.
Outline 1.The issues. 2.The PA experiment. 3.Results of the experiment. 4.Uniformity trials to compare designs. 5.Current designs. 6.Conclusions. 2
1.The issues The Plant accelerator ® Latest technology in high throughput plant imaging Plants are first grown in a Greenhouse then moved to the imaging room (Smarthouse) Automatic, non-destructive, repeated measurements of the physical attributes (phenotype) of plants in Smarthouse. 3
Issues in designing PA experiments At least two phases: Greenhouse and Smarthouse phases. Should one worry about design at all? o Perhaps better to rearrange location of plants during the experiment to average out microclimate effects. Even if design Smarthouse phase, do we need to worry about design in the Greenhouse phase? If do use designs, what design to use in a phase? What N-S or E-W trends should be accounted for? Is there spatial correlation? Does movement in PA have a thigmomorphogenic or other effect of movement? Other possible effects of movement are soil compaction and or root damage due to soil movement. 4
2.The PA experiment Ran a two-phase wheat experiment in PA. Brien & Bailey (2006) and Brien et al (2011) discuss such experiments; They generally involve multiple allocations and/or randomizations. In this case, a Greenhouse and a Smarthouse phase. All plants are Gladius. Specifically designed soil substrate to circumvent soil movement effects. 5
Greenhouse phase 6 East Western door NorthSouth Air con 288 pots 2Sides 2Blocks 3Rows in S 24Columns in B The 2 Sides by 2 Blocks correspond to 4 Locations in the Greenhouse. No allocations
Smarthouse phase: allocation of pots to carts 7 North West South Zone 1 Zone 2 Zone 3 Zone 4 288 carts 4Zones 3Lanes in Z 24Positions Smarthouse 288 pots 2Sides 2Blocks 3Rows in S 24Columns in B Greenhouse East North South Air con Zone 1 Zone 2 Zone 3 Zone 4 Solid lines indicate randomization while dashed lines indicate systematic assignment.
Smarthouse tactics Four tactics, each of 3 rows of 24 carts, were applied in the Smarthouse: 1. Bench: Plants placed on benches at the end of the conveyer system and not moved – no relocation; 2. Same lane: always return to the same position after watering or imaging – standard practice; 3. Half lane: After watering or imaging, move pots forward half a lane, which will result in pots changing sides from East to West and vice-a-versa with each move – restricted relocation; 4. Next lane: After watering or imaging, move the whole lane forward to the next lane in the Smarthouse – restricted relocation. 8
Allocation of Smarthouse tactics Pots have been allocated to carts. Four tactics are systematically allocated to zones. 9 288 carts 4Zones 3Lanes in Z 24Positions Smarthouse 288 pots 2Sides 2Blocks 3Rows in S 24Columns in B Greenhouse 4 treatments 4Tactics
Smarthouse phase 10 North West South Air con Imaging Bench Same Half Next North Zone 4 – Next lane West Air con Zone 3 – Half lane East
3.Results: data obtained Fresh weight at the end of the trial Total area (pixels) on Mon, Wed & Fri from day 21 to day 51. Height (cm) on day 51, from which derived a Density index ( = Total area / height). 11
Profile plots of the longitudinal data Next lane has slower growth is more variable. 12
Predicted growth curves using splines Next lane has slowest growth Half-lane has fastest growth 13
Total area measurements for Days 21 and 51 Focus on these: Day 21 represents the effect of the Greenhouse; Day 51 represents the combined effect of the Greenhouse and Smarthouse. Mixed models: Tactics + Tactics Lanes + td(Positions) + td(Positions) Tactics | idh(Tactics) Lanes ar1(Positions) o td means investigate trend over this factor; o idh mean investigate unequal variances between levels; o ar1 mean investigate autocorrelation between levels. 14
Results of mixed model analyses Similar models for Day 21 and Day 51. Differences in means and variances between the Tactics. However, no differences between bench and same lane for any responses (including density index). No evidence of spatial correlation. No differences between the three Lanes within each Tactic. Trends over Columns in the greenhouse and Positions in the Smarthouse that differ between Tactics. 15
Column/Position trends in Total area 16 Day 21Day 51 Area increases eastwards in the Greenhouse, mainly in south (light?). Increasing slope for all on Day 51, except for half-lane.
Position trends for Day 51 adjusted for Day 21 For same lane (and probably bench) there is a trend in the Smarthouse that increases from West to East (air in W). The Position trend in next lane parallels Column trend in Day 21 total areas greenhouse or Smarthouse? For half lane, no Smarthouse Position trend little Column trend in north-east and no Smarthouse contribution. 17 Day 51
Lane trend 18 Three split plots Exp. 1 – 22 lines and 3 conditions in 8 blocks; conditions in 3 vertical subplots. Exp. 2 – 153 lines and 2 conditions in 3 blocks; lines partially replicated; conditions in 2 horizontal subplots. Exp. 3 – 214 lines and 2 conditions in 8 blocks; lines partially replicated; conditions in 2 vertical subplots. Jo Tillbrooks 2011 experiments – fill Smarthouse
Estimated lane trend Plants in Lanes towards north grow less no. lanes with lower area depends on time of the year. Seems about 4 lanes are homogeneous. It would appear that the lower total area for next- lane tactic is due to shading in the northern zone. 19 Ribbons are CIs
Relative efficiencies from taking lane & position into account Experiment 123 No trend100.0 Lane trend139.9236.07146.8 Position trend102.895.998.0 Lane + Position trend148.8230.1147.7 20 The efficiencies are relative to the no-tend analysis. Gains vary. Gains in efficiency of 40% or more can be expected from allowing for lane trends. A 10% gain in efficiency can be expected from allowing for position trends in Experiment 1 only, and provided that lane trends allowed for.
Relocation during the PA experiment In half-lane tactic: Plants spend half time in eastern half and western half; Plants not equal in exposure to trend: when carts 13–24 moved to positions 1–12, relative east west positions maintained; Result is unable to detect trend, but greater individual plant variability (30% less precision). In next-lane tactic: Plants spend equal amount of time in shaded lanes; 5 or less days difference in entry of 1 st and exit of 3 rd lanes; No difference between lanes of next-lane tactic supports uniform exposure of plants to lane trend. 21
4.Uniformity trials to compare designs Each tactic, 3 Lanes 24 Positions, is essentially a uniformity trial (all Gladius, all treated equally). Perfect for comparing different designs to deal with position trends: Superimpose treatments (lines) on a zone using different designs; Analyse the total area according to the design; Compute the relative efficiencies (%) of designs: Repeat for a random sample of possible randomizations of the designs. 22
Relative efficiency (relative to a CRD) For a Proposed Design or Analysis (PDA): 23 where each It is a modified A-optimality criterion, the F taking into account any differences in denominator df. It compares the average sizes of the confidence intervals for pairwise differences between predictions for treatments. PDA with RE PDA > 100 is more efficient than a CRD: it has smaller s.e.d.s and so better able to detect treatment (line) differences.
Equally-replicated lines Consider the following designs & analyses with 36 (24) lines: 1) A CRD, without and with adjustment for Position trend; 2) An RCBD with two 3 12 (three 3 8 & 1 24) blocks, without and with adjustment for Position trend; 3) (Nearly) Trend-free designs for CRD & RCBD (DiGGer); 4) Resolved IBDs with blocks 3 1, 1 4 & 3 6 (3 1, 1 4 & 3 4) (CycDesigN); 5) Resolved row-col designs with two 3 12 (three 3 8) rectangles. 24 3 2 1 123456789101112131415161718192021222324 3 2 1 123456789101112131415161718192021222324 3 2 1 123456789101112131415161718192021222324 3 2 1 123456789101112131415161718192021222324
Equally- replicated lines Look for designs which give > 10% increase for all tactics. For 36 lines: small blocks, CRD + Adj, or TFD; but, TFCBD3 12EqLin best for same & next. For 24 lines: small blocks, CRD + Adj, RCBD 3 8 ( RRCD 3 8); TF or NTF no advantage. 25
Partially-replicated lines, with 2 conditions (an initial investigation) A split-plot design for 72 carts with: 1) 6 (or 8) duplicated lines, 20 (or 16) unreplicated lines and 2 control lines replicated twice; 2) Lines applied to 36 main plots, of 2 consecutive carts in the same lane, using an augmented block design; 3) 2 conditions randomized to the 2 subplots (carts) of a main plot. 26 3 2 1 123456789101112131415161718192021222324 3 2 1 123456789101112131415161718192021222324 Again, looked at designs with varying block sizes. 3 2 1 123456789101112131415161718192021222324 3 2 1 123456789101112131415161718192021222324
Partially- replicated lines, with two conditions Look for designs which give > 10% increase for all tactics. Line comparisons: best is main plots (2 carts) of 3 3 (= 3 Lanes 6 Positions) for t6 & t8, and 3 2 (= 3 Lanes 4 Positions) for t6. Conditions comparisons: little affected (as assigned to carts), but same designs best. 27 t6 t8
5.Current designs Smarthouse experiments run with 24 lanes x 22 positions 528 carts. Maximum of 23 carts per row because of weight limitations of conveyor system. Smarthouse divided into: (6 zones of 4 lanes) x (2 halves of 10 & 12 positions). Block designs for 12 blocks. Nearly trend-free designs: Lines are balanced across positions so that they are unaffected by a linear trend across positions in the Smarthouse. Often partially replicated designs: Parents/controls several replicates, 20% replicated twice, rest unreplicated. Generated using DiGGer software, an add-in to the statistical programming language R. (free) 28
Fleet x Commander mapping pop n 29 Germination and initial growth of 528 single-plant pots on 10 tables in southern space in Smarthouse. SmarthouseTable 1Table 3Table 5Table 7Table 9 Position123451234512345123451234 2Lane1Lane2Lane3Lane4Lane5Lane6Lane7Lane8Lane9Lane10Lane11Lane12Lane13Lane14Lane15Lane16Lane17Lane18Lane19Lane20Lane21Lane22Lane23Lane24 3Lane1Lane2Lane3Lane4Lane5Lane6Lane7Lane8Lane9Lane10Lane11Lane12Lane13Lane14Lane15Lane16Lane17Lane18Lane19Lane20Lane21Lane22Lane23Lane24 4Lane1Lane2Lane3Lane4Lane5Lane6Lane7Lane8Lane9Lane10Lane11Lane12Lane13Lane14Lane15Lane16Lane17Lane18Lane19Lane20Lane21Lane22Lane23Lane24 5Lane1Lane2Lane3Lane4Lane5Lane6Lane7Lane8Lane9Lane10Lane11Lane12Lane13Lane14Lane15Lane16Lane17Lane18Lane19Lane20Lane21Lane22Lane23Lane24 6Lane1Lane2Lane3Lane4Lane5Lane6Lane7Lane8Lane9Lane10Lane11Lane12Lane13Lane14Lane15Lane16Lane17Lane18Lane19Lane20Lane21Lane22Lane23Lane24 7Lane1Lane2Lane3Lane4Lane5Lane6Lane7Lane8Lane9Lane10Lane11Lane12Lane13Lane14Lane15Lane16Lane17Lane18Lane19Lane20Lane21Lane22Lane23Lane24 8Lane1Lane2Lane3Lane4Lane5Lane6Lane7Lane8Lane9Lane10Lane11Lane12Lane13Lane14Lane15Lane16Lane17Lane18Lane19Lane20Lane21Lane22Lane23Lane24 9Lane1Lane2Lane3Lane4Lane5Lane6Lane7Lane8Lane9Lane10Lane11Lane12Lane13Lane14Lane15Lane16Lane17Lane18Lane19Lane20Lane21Lane22Lane23Lane24 10Lane1Lane2Lane3Lane4Lane5Lane6Lane7Lane8Lane9Lane10Lane11Lane12Lane13Lane14Lane15Lane16Lane17Lane18Lane19Lane20Lane21Lane22Lane23Lane24 11Lane1Lane2Lane3Lane4Lane5Lane6Lane7Lane8Lane9Lane10Lane11Lane12Lane13Lane14Lane15Lane16Lane17Lane18Lane19Lane20Lane21Lane22Lane23Lane24 12Lane1Lane2Lane3Lane4Lane5Lane6Lane7Lane8Lane9Lane10Lane11Lane12Lane13Lane14Lane15Lane16Lane17Lane18Lane19Lane20Lane21Lane22Lane23Lane24 13Lane1Lane2Lane3Lane4Lane5Lane6Lane7Lane8Lane9Lane10Lane11Lane12Lane13Lane14Lane15Lane16Lane17Lane18Lane19Lane20Lane21Lane22Lane23Lane24 Table 2Table 4Table 6Table 8 Table 10 14Lane1Lane2Lane3Lane4Lane5Lane6Lane7Lane8Lane9Lane10Lane11Lane12Lane13Lane14Lane15Lane16Lane17Lane18Lane19Lane20Lane21Lane22Lane23Lane24 15Lane1Lane2Lane3Lane4Lane5Lane6Lane7Lane8Lane9Lane10Lane11Lane12Lane13Lane14Lane15Lane16Lane17Lane18Lane19Lane20Lane21Lane22Lane23Lane24 16Lane1Lane2Lane3Lane4Lane5Lane6Lane7Lane8Lane9Lane10Lane11Lane12Lane13Lane14Lane15Lane16Lane17Lane18Lane19Lane20Lane21Lane22Lane23Lane24 17Lane1Lane2Lane3Lane4Lane5Lane6Lane7Lane8Lane9Lane10Lane11Lane12Lane13Lane14Lane15Lane16Lane17Lane18Lane19Lane20Lane21Lane22Lane23Lane24 18Lane1Lane2Lane3Lane4Lane5Lane6Lane7Lane8Lane9Lane10Lane11Lane12Lane13Lane14Lane15Lane16Lane17Lane18Lane19Lane20Lane21Lane22Lane23Lane24 19Lane1Lane2Lane3Lane4Lane5Lane6Lane7Lane8Lane9Lane10Lane11Lane12Lane13Lane14Lane15Lane16Lane17Lane18Lane19Lane20Lane21Lane22Lane23Lane24 20Lane1Lane2Lane3Lane4Lane5Lane6Lane7Lane8Lane9Lane10Lane11Lane12Lane13Lane14Lane15Lane16Lane17Lane18Lane19Lane20Lane21Lane22Lane23Lane24 21Lane1Lane2Lane3Lane4Lane5Lane6Lane7Lane8Lane9Lane10Lane11Lane12Lane13Lane14Lane15Lane16Lane17Lane18Lane19Lane20Lane21Lane22Lane23Lane24 22Lane1Lane2Lane3Lane4Lane5Lane6Lane7Lane8Lane9Lane10Lane11Lane12Lane13Lane14Lane15Lane16Lane17Lane18Lane19Lane20Lane21Lane22Lane23Lane24 23Lane1Lane2Lane3Lane4Lane5Lane6Lane7Lane8Lane9Lane10Lane11Lane12Lane13Lane14Lane15Lane16Lane17Lane18Lane19Lane20Lane21Lane22Lane23Lane24 Colours represent 6 Zones of 4 Lanes on the conveyor system. Within a lane, 22 pots arranged in same order as will be placed on conveyor system.
Fleet x Commander mapping pop n 6 Zones each of 2 blocks of 4 lanes x 10 & 12 Positions. 2 parents replicated 6 times (blue), 36 lines replicated twice (grey), 180 lines unreplicated (green). 16.7% replicated. 2 consecutive carts have 2 conditions (no & added salt) randomized to them. Asymmetrical in 26–1 to distance from air con. 30 A partially-replicated, nearly-trend- free, block design with split plots.
Barley GWAS experiment 168 lines from a barley diversity panel from James Hutton Institute, each replicated thrice. 3 Australian varieties as controls, each replicated 8 times. 2 watering conditions to study drought tolerance. A total of 1024 pots requiring 2 Smarthouses. In one Smarthouse, controls have 4 replicates, 84 lines replicated twice and 84 occur only once. A partially-replicated, nearly-trend-free, block design with split plots used in each Smarthouse. Initial growth on tables with lanes kept in blocks. 31
6.Specific conclusions Not much Greenhouse column trend, except in south-east. There are substantial lane and, to a lesser extent, position trends in the Smarthouse. Designs in the Smarthouse should be block or trend-free designs, not row-and-column designs, nor spatial designs. The blocks in such design should be no larger than 4 Lanes by 12 Positions and smaller would be better. These conclusions need to be re-evaluated in other situations. 33
General conclusions No evidence of a thigmomorphogenic or other movement effect in the Smarthouse. (Bench & Same Lane tactics do not differ.) Rearranging carts only minimizes plant variability where exposure of the plants to microclimates is equalized. Designed experiments and statistical analysis can more easily and reliably achieve same as rearranging carts. Have aligned Greenhouse and Smarthouse features, e.g. blocks and trends, so both dealt with simultaneously. 34