Identification of important regressor groups, subgroups and individuals via regularization methods: application to gut microbiome data

Garcia, Tanya P.; Müller, Samuel; Carroll, Raymond J.; Walzem, Rosemary L.

doi:10.1093/bioinformatics/btt608

Cited by 35 publications

(35 citation statements)

References 25 publications

(35 reference statements)

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Our primary objective is to extend the feature selection process to incorporate a second level of grouping information. To achieve this, we follow a similar approach to the sparse group‐subgroup lasso (SGSL) method proposed by Garcia …”

Section: Methodsmentioning

confidence: 99%

“…Thus, the first component in the orthogonal setting is

β^{s P L S} = β^{l a s s o} \frac{{(β^{l a s s o})}^{⊺} β^{L S}}{{(β^{l a s s o})}^{⊺} β^{l a s s o}} .

Taking this 1 step further, we can see that in this setting, the PLS regression weight is the regression of the lasso weight on the least squares estimate. These calculations, in the orthogonal setting, carry over for the sparse group‐subgroup PLS, so we find

β^{s g s P L S} = β^{s g s L} \frac{{(β^{s g s L})}^{⊺} β^{L S}}{{(β^{s g s L})}^{⊺} β^{s g s L}},

where β s g s L is the regression coefficient in the sparse group‐subgroup lasso . The nonzero estimates from lasso‐based methods tend to be biased towards 0, so multiplying these coefficients by the scalar

\frac{{false(β^{s g s L} false)}^{⊺} β^{L S}}{{false(β^{s g s L} false)}^{⊺} β^{s g s L}}

could improve the prediction error of the model.…”

Section: Parameter Estimation For the Sgspls Methodsmentioning

confidence: 99%

“…Despite their success, the current regularised PLS methods are limited to analysis where the covariates can be clustered in at most 1 grouping layer. In many practical problems, datasets with multiple grouping layers arise naturally . By using multiple grouping layers of information, one can gain a better understanding and obtain more meaningful biological interpretations from the analysis …”

Section: Introductionmentioning

confidence: 99%

“…Importantly, we note that the method may be applied more generally in multivariate statistical problems where the covariates of the dataset have a natural group and subgroup structure. An alternative example would be using the group‐subgroup structure present in the multiple taxonomy levels of Gut microbiota data . Another natural grouping example of Omics data is the single nucleotide polymorphisms (SNP) data with multiple group levels corresponding to gene pathway group level and individual gene subgroup level.…”

Section: Introductionmentioning

confidence: 99%

“…Drawing inspiration from the recently proposed sparse group‐subgroup lasso (sgsLasso) we propose a new sparse group‐subgroup PLS (sgsPLS) method. In this paper, we will show that our new method has many applications: Applying both canonical and regression PLS methods to gain insight in a supervised or unsupervised setting. Applying penalised PLS regression for a single component with orthogonal predictors gives an analogue to well‐known penalised methods such as the lasso. Applying sgsPLS to data with group and subgroup structure can provide new biological insights by selecting variables at multiple grouping levels. …”

Section: Introductionmentioning

confidence: 99%

See 4 more Smart Citations

Sparse partial least squares with group and subgroup structure

Sutton

Thiébaut

Liquet

2018

Statistics in Medicine

View full text Add to dashboard Cite

Integrative analysis of high dimensional omics datasets has been studied by many authors in recent years. By incorporating prior known relationships among the variables, these analyses have been successful in elucidating the relationships between different sets of omics data. In this article, our goal is to identify important relationships between genomic expression and cytokine data from a human immunodeficiency virus vaccine trial. We proposed a flexible partial least squares technique, which incorporates group and subgroup structure in the modelling process. Our new method accounts for both grouping of genetic markers (eg, gene sets) and temporal effects. The method generalises existing sparse modelling techniques in the partial least squares methodology and establishes theoretical connections to variable selection methods for supervised and unsupervised problems. Simulation studies are performed to investigate the performance of our methods over alternative sparse approaches. Our R package sgspls is available at https://github.com/matt-sutton/sgspls.

show abstract

Section: Methodsmentioning

confidence: 99%

“…Thus, the first component in the orthogonal setting is

β^{s P L S} = β^{l a s s o} \frac{{(β^{l a s s o})}^{⊺} β^{L S}}{{(β^{l a s s o})}^{⊺} β^{l a s s o}} .

β^{s g s P L S} = β^{s g s L} \frac{{(β^{s g s L})}^{⊺} β^{L S}}{{(β^{s g s L})}^{⊺} β^{s g s L}},

\frac{{false(β^{s g s L} false)}^{⊺} β^{L S}}{{false(β^{s g s L} false)}^{⊺} β^{s g s L}}

could improve the prediction error of the model.…”

Section: Parameter Estimation For the Sgspls Methodsmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

See 3 more Smart Citations

Sparse partial least squares with group and subgroup structure

Sutton

Thiébaut

Liquet

2018

Statistics in Medicine

View full text Add to dashboard Cite

show abstract

DeepBiome: A Phylogenetic Tree Informed Deep Neural Network for Microbiome Data Analysis

Zhai,

Choi,

Yang

et al. 2024

Stat Biosci

View full text Add to dashboard Cite

Background: Evidence linking microbiome and human health is rapidly growing. Microbiome profile can be a novel predictive biomarker for many diseases. However, bacteria count tables are typically sparse and bacteria are classified at a hierarchy of taxonomic levels, ranging from species to phylum. Existing analysis tools focus on identifying microbiome associations either at the community level or at a specific pre-defined taxonomic level. They fail to incorporate the evolutionary relationship between bacteria and cannot learn from the data to aggregate microbiome contribution, thus leading to less accurate and less interpretable results in prediction, classification or selection.Results: We present DeepBiome, a phylogney-informed neural network architecture for predicting phenotypes from microbiome counts and uncovering the microbiome-phenotype association network. It takes microbiome abundance as the input and let the phylogenetic taxonomy guide the neural network architecture. Commonly used neural network architectures are targeted towards image and text analysis and typically require huge amount of training data, which is scarce in biomedical applications. By leveraging the phylogenetic information, DeepBiome relieves the heavy burden of tuning for the optimal deep learning architecture, avoids overfitting, and more importantly enables visualizing the path from microbiome counts to disease. It is applicable to both regression and classification problems. The simulation study and real-life data analysis demonstrate that DeepBiome is highly accurate and efficient and and provides a deep understanding of complex microbiome-phenotype associations even using small to moderate training sample sizes. Conclusions:In practice, it is unknown at which taxonomic level that microbiome clusters tag the association. Therefore, the central advantage of the presented method over other analytical methods is that it offers an ecological and evolutionary understanding of host-microbe interactions which is important for microbiome-based medicine. DeepBiome is implemented using Python packages Keras and the Tensorflow. It is an open-source tool available at (https://github.com/Young-won/DeepBiome).

show abstract