LP-Micro: Analyzing Time-Varying Microbial Effects using Machine Learning

Introduction

This GitHub repository includes the R code for LP-Micro, a machine learning-based method to predict the complex traits (e.g., early childhood caries, obesity.) using longitudinal metatranscriptomics. Additionally, LP-Micro can be used to analyze the temporal association between disease onset and microbiome by identifying the most predictive microbial taxa and ideal prognostic time.

Contact and Citation

Yifan Dai: yfd@unc.edu, Di Wu: did@email.unc.edu

Dai Y, Qian Y, Qu Y, et al. LP-Micro Offers Interpretable Disease Outcome Prediction by Leveraging Microbial Biomarkers and Their Time-Varying Effects. bioRxiv. 2024

Installation

• Available for Linux and Mac.
• For Windows users, please install Rtools prior to install LP-Micro.

install.packages("devtools")
devtools::install_github("IV012/LPMicro")

Usage

LP-Micro includes the following components:

• Microbial Taxa Screening screen_taxa: LP-Micro combines group lasso and polynomial splines to screen disease-correlated taxa from longitudinal microbiome data.

• Visitwise Prediction visit_predict: LP-Micro supports popular machine learning methods for predicting the future disease onset using microbial profiles at a single time point, namely visitwise prediction in our paper. User can idenfity the most predictive time point of microbiome for disease outcomes.

• Cumulative Prediction cumulative_predict: LP-Micro also supports the prediction of disease outcomes using microbial up to a chosen visit, namely cumulative prediction. Cumulative prediction is usually more accurate than visitwise prediction with proper feature engineering using screen_taxa.

• Feature Interpretation taxa_interpret: LP-Micro provides statistical p-values to test whether the following features are useful for prediction, including (i) a microbial taxon, (ii) all information from one time point, and (iii) a microbial taxon at a chosen time point.

Feel free to check more details using help and apply LP-Micro to your data!

help(screen_taxa)
help(cumulative_predict)
help(visit_predict)
help(taxa_interpret)

Examples

In the following script, we introduce the general input and output of the taxa screening function.

# general input format
n <- 200 # 200 individuals
p <- 500 # 500  microbial taxa
q <- 5 # 5 repeated measurements
x <- matrix(rnorm(n*p*q), n, p*q) # simulated microbial matrix
y <- rnorm(n) # simulated response matrix
idx <- rep(1:p, q) # identifier of the microbial taxa across the columns of x

# microbial feature screening, lambda is the tuning parameter
screen_taxa(x, y, idx, lambda = c(1e-2, 1e-4)) # this returns a list of taxa of the same length as lambda

Next, we introduce the machine learning based feature importance. First, we need to specify the machine learning model and its parameter. Below is a template of the model specification. We take random forest as an example.

# models and parameters supported by LP-Micro
n_ensemble = 100 #100
n_perm = 100 #100

n_epoch = 1000 #1000
n_tree = 1000
node_size = 3
esCtrl = list(
  n.hidden = c(50, 40, 30, 20), 
  activate = "relu", 
  l1.reg = 10**-4, 
  early.stop.det = 1000, 
  n.batch = 50, 
  n.epoch = n_epoch, 
  learning.rate.adaptive = "adam", 
  plot = FALSE)

args_list = list(
  svm = list(method="svm", n.ensemble=n_ensemble),
  rf = list(
    method = "random_forest",
    n.ensemble=n_ensemble, 
    ntree=n_tree, 
    nodesize=node_size
  ),
  xgb = list(
    method="xgboost", 
    params=list(
      booster="gbtree", 
      objective="reg:squarederror", 
      eta=0.3, 
      gamma=0, 
      max_depth=5, 
      min_child_weight=1, 
      subsample=1, 
      colsample_bytree=1
    )
  ),
  lasso = list(method="lasso"),
  dnn = list(
    method = "ensemble_dnnet",
    n.ensemble = n_ensemble,
    esCtrl = esCtrl,
    verbose = 0
  )
)

mod_args <- args_list[[2]]

For cumulative/visit-wise prediction, we also need to specify the cumulative dataset. This records the predictive accuracy up-to or of each visit.

# taxa_list represents the microbial abundance measured up to a time point
taxa_list <- list()
for(t in 1:q){
  taxa_list[[t]] <- list(x=x[, 1:(p*t)], y=y)
}

cumulative_predict(taxa_list, mod_args, type="regression")

# taxa_list represents the microbial abundance measured at a time point
taxa_list <- list()
for(t in 1:q){
  taxa_list[[t]] <- list(x=x[, (p*(t-1)+1):(p*t)], y=y)
}

visit_predict(taxa_list, mod_args, type="regression")

Finally, we present the feature and visit importance from permutation importance test.

idx_tx <- rep(1:p, q) # identifier of the microbial taxa across the columns of x
idx_tp <- c() # identifier of the measured time point
for(t in 1:q){
  idx_tp <- c(idx_tp, rep(t, p))
}

taxa_interpret(x, y, idx_tx, idx_tp, mod_args) # This gives the taxa and visit importance