Skip to contents

Deep Learning with OmicSelector

Source: vignettes/deep-learning.Rmd
deep-learning.Rmd

Overview

This vignette demonstrates a full deep learning workflow with OmicSelector, including autoencoder feature compression, mlr3torch training, checkpoint export/import, and fine‑tuning.

Requirements: - torch - mlr3torch

install.packages("torch")
torch::install_torch()
install.packages("mlr3torch")

Load Package and Data 

library(OmicSelector)

data("original_TCGA_data", package = "OmicSelector")

Prepare a Small TCGA Subset 

feature_cols <- grep("^hsa\\.", names(original_TCGA_data), value = TRUE)
feature_cols <- head(feature_cols, 300)

subset_df <- original_TCGA_data[, c("patient", "sample_type", feature_cols), drop = FALSE]
subset_df <- as.data.frame(subset_df)
subset_df$sample_type <- factor(subset_df$sample_type)

set.seed(1)
idx_tumor <- which(subset_df$sample_type == "PrimaryTumor")
idx_normal <- which(subset_df$sample_type == "SolidTissueNormal")
subset_df <- subset_df[c(sample(idx_tumor, 60), sample(idx_normal, 60)), ]
subset_df <- subset_df[sample(nrow(subset_df)), ]

Create Pipeline 

pipeline <- OmicPipeline$new(
  data = subset_df,
  target = "sample_type",
  positive = "PrimaryTumor",
  patient_id = "patient"
)

Build Deep Learning GraphLearner

This pipeline adds a torch autoencoder before feature selection and trains an MLP via mlr3torch.

learner <- pipeline$create_graph_learner(
  filter = "anova",
  model = "mlp",
  n_features = 50,
  autoencoder = list(
    latent_dim = 32,
    hidden_layers = c(128, 64),
    epochs = 50,
    batch_size = 64,
    early_stopping = TRUE,
    patience = 10
  )
)

Run Nested CV 

result <- pipeline$benchmark(
  learners = learner,
  outer_folds = 3,
  inner_folds = 2,
  seed = 42,
  parallel = TRUE,
  threads = 1
)

print(result)

Fit Final Model 

fit <- pipeline$fit(learner, seed = 42)
fit$selected_features

Export / Import Checkpoint 

export_omicfit_checkpoint(fit, "mlp_checkpoint.pt")

# Load into a trained learner (for inference)
fit <- import_omicfit_checkpoint(fit, "mlp_checkpoint.pt")

Fine‑tune on a New Task 

task <- pipeline$get_task()

finetuned <- finetune_mlr3torch_checkpoint(
  learner = learner,
  task = task,
  path = "mlp_checkpoint.pt",
  epochs = 30
)

# Or stay at the OmicFit level
finetuned_fit <- finetune_omicfit_checkpoint(
  fit = fit,
  task = task,
  path = "mlp_checkpoint.pt",
  epochs = 30
)

Notes

  • Autoencoders can also be trained separately via autoencoder_fit() and passed into the pipeline using pretrained and freeze_encoder.
  • For transfer learning across datasets, export checkpoints and re‑train with finetune_mlr3torch_checkpoint().