A randomised factorial trial is usually designed to detect the smallest clinically meaningful difference for all interventions, absent interaction. When a negative interaction is apparent without a sufficient sample size to test for it, the question arises as to how to best report the results, given the constraints of the statistical analysis plan (SAP). We present the rationale and decisions made in this regard when reporting the results from the Topical Analgesia Post-Haemorrhoidectomy randomised factorial trial, in which a significant difference between a lidocaine-only and a lidocaine-diltiazem formulation injected unexpected variability in the primary outcome, a pain visual analogue scale. While the SAP dictated the reporting of the main effects only, appropriate translation to clinical practice required separate reporting of the four arms. We describe the trial’s design, longitudinal data collection process, main analysis, and explain how we dealt with this issue.

Fall armyworm (FAW) was first detected in 2020 in Queensland. It quickly spread to several locations in Australia, causing up to 80% losses in sweet corn and maize1. This paper describes the design and analysis of a plot trial conducted in north Queensland which assessed the susceptibility and preference of commercially available sweetcorn varieties to FAW infestation.

Genstat was used to generate the (latinized) doubly resolvable row-column design of four replicates of eight varieties. Variables measured included FAW oviposition rate (egg masses), number of larvae and the intensity of foliage damage. Analyses were conducted in Genstat using hierarchical generalised linear models and with the use of the VAROWCOLDESIGN procedure. This procedure allowed the analysis of the row-and-column design by REML, with the convenience of automatic selection of the best random and spatial covariance model. Significant differences were found between varieties in the number of total alive FAW larvae.

Plant breeding multi-environment trials (METs) are an important component of identifying superior varieties as they allow an assessment of variety by environment interaction (VEI). A key consequence of the presence of VEI is that it is both statistically invalid and biologically meaningless to base variety selections on measures of overall performance across all environments in the MET. In this poster we address the issue for genomic selection within the framework of a single-step linear mixed model analysis in which a factor analytic model is used for the variety effects in each environment. Our approach for variety selection involves the formation of groups of environments, called interaction classes (iClasses), within which there is minimal crossover VEI. It is then valid to obtain predictions of variety performance across the environments in each iClass. We demonstrate the methods using a commercial Wheat MET dataset.

Interactions between stressors are defined as synergistic if the combined effect is greater than the sum of each effect independently, or antagonistic if less. This implies that the null hypothesis (of no interaction) is that the stressors combine additively (since it is the sum). But would a null hypothesis of combining multiplicatively be more sensible? For instance, if both stressors reduce growth by 50%, should one expect that the combined effect would be 0% or 25% growth if there was no interaction? The motivation for research comes from investigations into viroid interactions on the growth of citron (C. medica L.) plants. The issues surrounding fitting statistical models to studies examining interactions between stressors, such as pathogen interactions, will also be discussed.

Rhizoctonia root rot is a fungal soilborne disease that affects cereal and broadleaf crops. It is caused by a long-lived necrotrophic fungus which grows through the top layer of the soil decaying and pruning the plant roots it infects.

In field trials involving Rhizoctonia root rot, disease severity has traditionally been assessed using bespoke rating schemes on plants removed from plots with ratings that reflect visual markers of infection, such as lesion colour and the presence of pruned roots, and a grading of how badly a plant is affected. The resulting qualitative ordinal data is then often inappropriately treated as metric data.

We present alternative quantitative traits for assessing disease severity which were used alongside the conventional approach in experiments conducted to assess Rhizoctonia root rot incidence and severity. These quantitative traits can be analysed with standard methods, are easier to interpret, and offer an approach which can be consistently applied across trials and through time.

Joint matrix factorization is a popular method for extracting lower dimensional representations of multi-omics data. It disentangles underlying biological signals, facilitating efficient sample clustering, disease subtyping, or biomarker identification. However, with a limited number of samples, the effectiveness of matrix factorization is reduced. Addressing this limitation, we introduce MOTL (Multi-Omics Transfer Learning), a novel framework for multi-omics matrix factorization with transfer learning based on MOFA (Arguelet et al. 2018). MOTL infers latent factors for a small multi-omics dataset, with respect to those inferred from a large heterogeneous learning dataset. We designed two evaluation protocols based on simulated and real multi-omics data. We observed that MOTL improves the factorization of multi-omics datasets, comprised of a limited number of samples, when compared to factorization without transfer learning. We showcase the usefulness of MOTL on a small sample glioblastoma dataset, revealing an enhanced delineation of cancer status and subtype thanks to transfer learning.

The Gaussian graphical model (GGM) has become a popular tool for biological network analysis to study conditional dependence between biological variables and predict gene functions, phenotypes and patterns of molecular regulation. We consider the problem of learning a GMM for paired data where observations come from two dependent groups sharing the same variables. Comparing the distribution of a set of variables between two experimental conditions, or groups, is common to many real-world applications such as cancer genomic data or brain networks from fMRI data. To address the problem of high dimensionality and heterogeneity, we apply a mixture of GGMs for paired data and develop a penalized EM algorithm for these models where model estimation and selection can be performed simultaneously. The efficiency of our approach is then demonstrated on constructing gene networks using both real and simulated transcriptome data sets.

As part of the Australian Government’s efforts to reduce greenhouse gas emissions, farmers can gain Carbon Credits by sequestering soil carbon through improved management practices[1]. However, measuring changes in soil organic carbon (SOC) is difficult, as changes in SOC are typically small, whilst variability in the field (both spatially and temporally) is often substantial and measurement errors can be considerable. Additionally, the sample size is generally small; the Government requires a minimum of 9 samples, whereas plots can be as large as 200 ha. This work therefore aims to identify improved sampling protocols to reliably detect changes in SOC in a practically feasible way. As a first step, a large sampling campaign was undertaken, in which over 1000 cores were extracted from 9 farms in eastern Australia. This poster highlights preliminary results on variability in SOC estimates, and showcases the experimental design for the second round of sampling.

We introduce a mutual influence model for two-mode network that can effectively capture both inward and outward influences within the network. To estimate the unknown parameters, we employ quasi-maximum likelihood estimation and derive the asymptotic properties of our estimators. Furthermore, we propose multiple hypothesis tests to evaluate the inward and outward influence of the network. To assess the model performance, we conduct numerical studies and real data analysis.

We propose a Bayesian Inference approach to infer the unknown parameters of a Gaussian plume model to predict the behaviour of a marine cloud brightening (MCB) Plume on the Great Barrier Reef. Normalised data from 12 experiments collected from an MCB plume has been used alongside prior knowledge of atmospheric stability to understand the overall time averaged behaviour of the plume. This research employs two different priors for the inferred parameters. The study focuses mostly on the diffusion parameters of the Gaussian plume model and how they relate to the Pasquill-Gifford Stability classes. It compares the effectiveness of a single and multi-mode prior distribution. While each method has its advantages and disadvantages, we have concluded that the multi-modal Prior stands out as the method which has the best balance between interpretability and performance.

In New Zealand, 7% of deaths are related to respiratory diseases, with Pacific people at higher risk. Our work investigates the causal effects of early-life risks and resilience factors on early-adulthood lung function amongst Pacific Islands Families Study (PIFS) cohort members (n=1,398). 466 from the cohort participated in the respiratory study. Primary outcome was forced expiratory volume in 1 second (FEV1) z-score at age 18 years. FEV1 and healthy lung function (HLF), defined as the z-score being larger than -1.64, were secondary outcomes. A previous study had evaluated the effects of early-life nutrition factors on the respiratory health of Pacific youth. The results suggested a positive impact of consuming more fruit and vegetables during childhood on respiratory health later in life. The follow-up study will continue to explore the effects of factors from relevant domains based on the PIFS cohort, where a new integrated model will be applied. A simulation will be conducted to determine this model.

We investigate the relationships between local environmental variables and the geochemical composition of the Earth in a region spanning over 26,000 km^2 in the lower South Island of New Zealand. Part of the Southland–South Otago geochemical baseline survey—a pilot study pre-empting roll-out across the country—the data comprise the measurements of 59 chemical trace elements, each at two depth prescriptions, at several hundred spatial sites. We demonstrate construction of a hierarchical spatial factor model that captures inter-depth dependency; handles imputation of left-censored readings in a statistically principled manner; and exploits sparse approximations to Gaussian processes to deliver inference.