SIAMCAT: Statistical Inference of Associations between Microbial Communities And host phenoTypes
Date last modified: 2020-04-04
Contents
1 About This Vignette
This vignette aims to be a short tutorial for the main functionalities of
SIAMCAT. Examples of additional workflows or more detailed tutorials can
be found in other vignettes (see the
BioConductor page).
SIAMCAT is part of the suite of computational microbiome analysis tools
hosted at EMBL by the groups of
Peer Bork and
Georg Zeller. Find
out more at EMBL-microbiome tools.
2 Introduction
Associations between microbiome and host phenotypes are ideally described by
quantitative models able to predict host status from microbiome composition.
SIAMCAT can do so for data from hundreds of thousands of microbial taxa, gene
families, or metabolic pathways over hundreds of samples. SIAMCAT produces
graphical output for convenient assessment of the quality of the input data and
statistical associations, for model diagnostics and inference revealing the
most predictive microbial biomarkers.
3 Quick Start
For this vignette, we use an example dataset included in the SIAMCAT package.
As example dataset we use the data from the publication of
Zeller et al, which demonstrated
the potential of microbial species in fecal samples to distinguish patients
with colorectal cancer (CRC) from healthy controls.
library("SIAMCAT")
data("feat_crc_zeller", package="SIAMCAT")
data("meta_crc_zeller", package="SIAMCAT")First, SIAMCAT needs a feature matrix (can be either a matrix, a
data.frame, or a phyloseq-otu_table), which contains values of different
features (in rows) for different samples (in columns). For example, the
feature matrix included here contains relative abundances for bacterial
species calculated with the mOTU profiler for 141 samples:
feat.crc.zeller[1:3, 1:3]## CCIS27304052ST-3-0 CCIS15794887ST-4-0
## UNMAPPED 0.589839 0.7142157
## Methanoculleus marisnigri [h:1] 0.000000 0.0000000
## Methanococcoides burtonii [h:10] 0.000000 0.0000000
## CCIS74726977ST-3-0
## UNMAPPED 0.7818674
## Methanoculleus marisnigri [h:1] 0.0000000
## Methanococcoides burtonii [h:10] 0.0000000dim(feat.crc.zeller)## [1] 1754 141Please note that
SIAMCATis supposed to work with relative abundances. Other types of data (e.g. counts) will also work, but not all functions of the package will result in meaningful outputs.
Secondly, we also have metadata about the samples in another data.frame:
head(meta.crc.zeller)## Age BMI Gender AJCC_stage FOBT Group
## CCIS27304052ST-3-0 52 20 F -1 Negative CTR
## CCIS15794887ST-4-0 37 18 F -1 Negative CTR
## CCIS74726977ST-3-0 66 24 M -1 Negative CTR
## CCIS16561622ST-4-0 54 26 M -1 Negative CTR
## CCIS79210440ST-3-0 65 30 M -1 Positive CTR
## CCIS82507866ST-3-0 57 24 M -1 Negative CTRIn order to tell SIAMCAT, which samples are cancer cases and which are
healthy controls, we can construct a label object from the Group column in
the metadata.
label.crc.zeller <- create.label(meta=meta.crc.zeller,
label='Group', case='CRC')## Label used as case:
## CRC
## Label used as control:
## CTR## + finished create.label.from.metadata in 0.001 sNow we have all the ingredients to create a SIAMCAT object. Please have a
look at the vignette about input formats for more
information about supported formats and other ways to create a SIAMCAT object.
sc.obj <- siamcat(feat=feat.crc.zeller,
label=label.crc.zeller,
meta=meta.crc.zeller)## + starting validate.data## +++ checking overlap between labels and features## + Keeping labels of 141 sample(s).## +++ checking sample number per class## +++ checking overlap between samples and metadata## + finished validate.data in 0.032 sA few information about the SIAMCAT object can be accessed with the show
function from phyloseq (SIAMCAT builds on the phyloseq data structure):
show(sc.obj)## siamcat-class object
## label() Label object: 88 CTR and 53 CRC samples
##
## contains phyloseq-class experiment-level object @phyloseq:
## phyloseq@otu_table() OTU Table: [ 1754 taxa and 141 samples ]
## phyloseq@sam_data() Sample Data: [ 141 samples by 6 sample variables ]Since we have quite a lot of microbial species in the dataset at the moment, we
can perform unsupervised feature selection using the function filter.features.
sc.obj <- filter.features(sc.obj,
filter.method = 'abundance',
cutoff = 0.001)## Features successfully filtered4 Association Testing
Associations between microbial species and the label can be tested
with the check.associations function. The function computes for each species
the significance using a non-parametric Wilcoxon test and different effect
sizes for the association (e.g. AUC or fold change).
sc.obj <- check.associations(sc.obj, log.n0 = 1e-06, alpha = 0.05)
association.plot(sc.obj, sort.by = 'fc',
panels = c('fc', 'prevalence', 'auroc'))The function produces a pdf file as output, since the plot is optimized for a
landscape DIN-A4 layout, but can also used to plot on an active graphic device,
e.g. in RStudio. The resulting plot then looks like that:
5 Confounder Testing
As many biological and technical factors beyond the primary phenotype of
interest can influence microbiome composition, simple association studies may
suffer confounding by other variables, which can lead to spurious results.
The check.confounders function provides the option to test the associated
metadata variables for potential confounding influence. No information is stored
in the SIAMCAT object, but the different analyses are visualized and saved to
a combined pdf file for qualitative interpretation.
check.confounders(sc.obj, fn.plot = 'confounder_plots.pdf',
meta.in = NULL, feature.type = 'filtered')The conditional entropy check primarily serves to remove nonsensical variables from subsequent checks. Conditional entropy quantifies the unique information contained in one variable (row) respective to another (column). Identical variables and derived variables which share the exact same information will have a value of zero. In this example, the label was derived from the Group variable which was determined from AJCC stage, so both are excluded.
Conditional Entropy Plot
To better quantify potential confounding effects of metadata variables on
individual microbial features, check.confounder plots the variance explained
by the label in comparison with the variance explained by the metadata variable
for each individual feature. Variables with many features in the upper left
corner might be confounding the label associations.
Variance Explained Plot
6 Model Building
One strength of SIAMCAT is the versatile but easy-to-use interface for the
construction of machine learning models on the basis of microbial species.
SIAMCAT contains functions for data normalization, splitting the data into
cross-validation folds, training the model, and making predictions based on
cross-validation instances and the trained models.
6.1 Data Normalization
Data normalization is performed with the normalize.features function. Here,
we use the log.unit method, but several other methods and customization
options are available (please check the documentation).
sc.obj <- normalize.features(sc.obj, norm.method = "log.unit",
norm.param = list(log.n0 = 1e-06, n.p = 2,norm.margin = 1))## Features normalized successfully.6.2 Prepare Cross-Validation
Preparation of the cross-validation fold is a crucial step in machine learning.
SIAMCAT greatly simplifies the set-up of cross-validation schemes, including
stratification of samples or keeping samples inseperable based on metadata.
For this small example, we choose a twice-repeated 5-fold cross-validation
scheme. The data-split will be saved in the data_split slot of the SIAMCAT
object.
sc.obj <- create.data.split(sc.obj, num.folds = 5, num.resample = 2)## Features splitted for cross-validation successfully.6.3 Model Training
The actual model training is performed using the function train.model.
Again, multiple options for customization are available, ranging from the
machine learning method to the measure for model selection or customizable
parameter set for hyperparameter tuning.
sc.obj <- train.model(sc.obj, method = "lasso")The models are saved in the model_list slot of the SIAMCAT object. The
model building is performed using the mlr R package. All models can easily be
accessed.
# get information about the model type
model_type(sc.obj)## [1] "lasso"# access the models
models <- models(sc.obj)
models[[1]]$model## <LearnerClassifCVGlmnet:classif.cv_glmnet>
## * Model: cv.glmnet
## * Parameters: alpha=1, s=0.0441
## * Packages: mlr3, mlr3learners, glmnet
## * Predict Types: response, [prob]
## * Feature Types: logical, integer, numeric
## * Properties: multiclass, selected_features, twoclass, weights6.4 Make Predictions
Using the data-split and the models trained in previous step, we can use the
function make.predictions in order to apply the models on the test instances
in the data-split. The predictions will be saved in the pred_matrix slot of
the SIAMCAT object.
sc.obj <- make.predictions(sc.obj)
pred_matrix <- pred_matrix(sc.obj)head(pred_matrix)## CV_rep1 CV_rep2
## CCIS27304052ST-3-0 0.10109512 0.16441672
## CCIS15794887ST-4-0 0.13744065 0.31452241
## CCIS74726977ST-3-0 0.32604131 0.60417075
## CCIS16561622ST-4-0 0.46672609 0.11842978
## CCIS79210440ST-3-0 0.21227138 0.08580803
## CCIS82507866ST-3-0 0.03522074 0.134500327 Model Evaluation and Interpretation
In the final part, we want to find out how well the model performed and which
microbial species had been selected in the model. In order to do so, we first
calculate how well the predictions fit the real data using the function
evaluate.predictions. This function calculates the Area Under the Receiver
Operating Characteristic (ROC) Curve (AU-ROC) and the Precision Recall (PR)
Curve for each resampled cross-validation run.
sc.obj <- evaluate.predictions(sc.obj)## Evaluated predictions successfully.7.1 Evaluation Plot
To plot the results of the evaluation, we can use the function
model.evaluation.plot, which produces a pdf-file showing the ROC and PR
Curves for the different resamples runs as well as the mean ROC and PR Curve.
model.evaluation.plot(sc.obj)
7.2 Interpretation Plot
The final plot produced by SIAMCAT is the model interpretation plot, created
by the model.interpretation.plot function. The plot shows for the top
selected features the
model weights (and how robust they are) as a barplot,
a heatmap with the z-scores or fold changes for the top selected features, and
a boxplot showing the proportions of weight per model which is captured by the top selected features.
Additionally, the distribution of metadata is shown in a heatmap below.
The function again produces a pdf-file optimized for a landscape DIN-A4 plotting region.
model.interpretation.plot(sc.obj, fn.plot = 'interpretation.pdf',
consens.thres = 0.5, limits = c(-3, 3), heatmap.type = 'zscore')The resulting plot looks like this:
8 Session Info
sessionInfo()## R version 4.2.1 (2022-06-23)
## Platform: x86_64-pc-linux-gnu (64-bit)
## Running under: Ubuntu 20.04.5 LTS
##
## Matrix products: default
## BLAS: /home/biocbuild/bbs-3.15-bioc/R/lib/libRblas.so
## LAPACK: /home/biocbuild/bbs-3.15-bioc/R/lib/libRlapack.so
##
## locale:
## [1] LC_CTYPE=en_US.UTF-8 LC_NUMERIC=C
## [3] LC_TIME=en_GB LC_COLLATE=C
## [5] LC_MONETARY=en_US.UTF-8 LC_MESSAGES=en_US.UTF-8
## [7] LC_PAPER=en_US.UTF-8 LC_NAME=C
## [9] LC_ADDRESS=C LC_TELEPHONE=C
## [11] LC_MEASUREMENT=en_US.UTF-8 LC_IDENTIFICATION=C
##
## attached base packages:
## [1] stats graphics grDevices utils datasets methods base
##
## other attached packages:
## [1] ggpubr_0.4.0 SIAMCAT_2.0.1 phyloseq_1.40.0 mlr3_0.14.0
## [5] forcats_0.5.2 stringr_1.4.1 dplyr_1.0.10 purrr_0.3.4
## [9] readr_2.1.3 tidyr_1.2.1 tibble_3.1.8 ggplot2_3.3.6
## [13] tidyverse_1.3.2 BiocStyle_2.24.0
##
## loaded via a namespace (and not attached):
## [1] readxl_1.4.1 uuid_1.1-0 backports_1.4.1
## [4] corrplot_0.92 plyr_1.8.7 igraph_1.3.5
## [7] splines_4.2.1 listenv_0.8.0 GenomeInfoDb_1.32.4
## [10] gridBase_0.4-7 digest_0.6.29 foreach_1.5.2
## [13] htmltools_0.5.3 magick_2.7.3 lmerTest_3.1-3
## [16] fansi_1.0.3 magrittr_2.0.3 checkmate_2.1.0
## [19] googlesheets4_1.0.1 cluster_2.1.4 tzdb_0.3.0
## [22] globals_0.16.1 Biostrings_2.64.1 modelr_0.1.9
## [25] mlr3tuning_0.14.0 matrixStats_0.62.0 vroom_1.6.0
## [28] prettyunits_1.1.1 colorspace_2.0-3 rvest_1.0.3
## [31] haven_2.5.1 xfun_0.33 crayon_1.5.2
## [34] RCurl_1.98-1.8 jsonlite_1.8.2 lme4_1.1-30
## [37] survival_3.4-0 iterators_1.0.14 ape_5.6-2
## [40] glue_1.6.2 mlr3measures_0.5.0 gtable_0.3.1
## [43] gargle_1.2.1 zlibbioc_1.42.0 XVector_0.36.0
## [46] car_3.1-0 Rhdf5lib_1.18.2 shape_1.4.6
## [49] BiocGenerics_0.42.0 abind_1.4-5 scales_1.2.1
## [52] infotheo_1.2.0.1 DBI_1.1.3 rstatix_0.7.0
## [55] Rcpp_1.0.9 progress_1.2.2 palmerpenguins_0.1.1
## [58] bit_4.0.4 stats4_4.2.1 glmnet_4.1-4
## [61] httr_1.4.4 RColorBrewer_1.1-3 ellipsis_0.3.2
## [64] farver_2.1.1 pkgconfig_2.0.3 sass_0.4.2
## [67] dbplyr_2.2.1 utf8_1.2.2 labeling_0.4.2
## [70] tidyselect_1.1.2 rlang_1.0.6 reshape2_1.4.4
## [73] PRROC_1.3.1 munsell_0.5.0 cellranger_1.1.0
## [76] tools_4.2.1 cachem_1.0.6 cli_3.4.1
## [79] generics_0.1.3 ade4_1.7-19 broom_1.0.1
## [82] evaluate_0.16 biomformat_1.24.0 fastmap_1.1.0
## [85] yaml_2.3.5 bit64_4.0.5 knitr_1.40
## [88] fs_1.5.2 lgr_0.4.4 beanplot_1.3.1
## [91] bbotk_0.5.4 future_1.28.0 nlme_3.1-159
## [94] paradox_0.10.0 xml2_1.3.3 compiler_4.2.1
## [97] curl_4.3.2 ggsignif_0.6.3 reprex_2.0.2
## [100] bslib_0.4.0 stringi_1.7.8 highr_0.9
## [103] lattice_0.20-45 Matrix_1.5-1 nloptr_2.0.3
## [106] vegan_2.6-2 permute_0.9-7 multtest_2.52.0
## [109] vctrs_0.4.2 pillar_1.8.1 lifecycle_1.0.2
## [112] rhdf5filters_1.8.0 BiocManager_1.30.18 jquerylib_0.1.4
## [115] LiblineaR_2.10-12 data.table_1.14.2 bitops_1.0-7
## [118] R6_2.5.1 bookdown_0.29 gridExtra_2.3
## [121] mlr3misc_0.11.0 IRanges_2.30.1 parallelly_1.32.1
## [124] codetools_0.2-18 boot_1.3-28 MASS_7.3-58.1
## [127] assertthat_0.2.1 rhdf5_2.40.0 mlr3learners_0.5.4
## [130] withr_2.5.0 S4Vectors_0.34.0 GenomeInfoDbData_1.2.8
## [133] mgcv_1.8-40 parallel_4.2.1 hms_1.1.2
## [136] grid_4.2.1 minqa_1.2.4 rmarkdown_2.16
## [139] carData_3.0-5 googledrive_2.0.0 pROC_1.18.0
## [142] numDeriv_2016.8-1.1 Biobase_2.56.0 lubridate_1.8.0