Con Proportional Hazards Style of Analysis

[DarioS]

Could LinDA have a survival time analysis (event indicator, follow-up days) mode?

[cafferychen777]

Hi @DarioS,

Thank you for this interesting feature request! I apologize for the delayed response.

Current LinDA Design

LinDA (Linear models for Differential Abundance) is specifically designed for cross-sectional and longitudinal microbiome studies using linear regression and mixed-effects models. It works by:

1. Applying centered log2-ratio transformation to handle compositional data
2. Fitting linear/mixed-effects models
3. Correcting bias due to compositional effects
4. Outputting log2 fold changes and associated statistics

Cox Proportional Hazards Analysis

Cox proportional hazards (survival analysis) requires a fundamentally different statistical framework:

• Outcome: Time-to-event data (survival time + event indicator)
• Model: Semi-parametric hazard function
• Effect measure: Hazard ratios (HR), not fold changes
• Assumptions: Proportional hazards, rather than linear relationships

Current Limitations

LinDA does not currently support survival analysis because:

1. The underlying statistical framework is incompatible (linear models vs. survival models)
2. The output metrics differ (log2FC vs. hazard ratios)
3. The data structure requirements are different

Recommended Alternatives

For microbiome data with survival outcomes, I recommend:

1. MaAsLin 2 with Cox Models

# MaAsLin 2 supports multiple model types including Cox
library(Maaslin2)
fit_data <- Maaslin2(
  input_data = abundance_table,
  input_metadata = metadata,
  output = "output_dir",
  analysis_method = "LM",  # Note: Cox not directly supported
  ...
)

2. Manual Cox Regression Approach

For each taxon, fit a Cox model with CLR-transformed abundances:

library(survival)
library(compositions)

# CLR transformation (similar to LinDA)
clr_transform <- function(x) {
  log_x <- log(x + 0.5)  # Add pseudo-count
  return(log_x - mean(log_x))
}

# Apply to your feature table
clr_data <- t(apply(feature_table, 1, clr_transform))

# Fit Cox model for each taxon
results <- list()
for (taxon in rownames(clr_data)) {
  surv_data <- data.frame(
    time = metadata$follow_up_days,
    event = metadata$event_indicator,
    abundance = clr_data[taxon, ]
  )
  
  cox_model <- coxph(Surv(time, event) ~ abundance, data = surv_data)
  results[[taxon]] <- summary(cox_model)
}

3. coxme Package for Mixed Effects

If you need random effects (like LinDA's mixed models):

library(coxme)
cox_mixed <- coxme(Surv(time, event) ~ abundance + (1|subject_id), 
                   data = surv_data)

Potential Future Development

This is an interesting feature that could benefit the microbiome community. However, implementing Cox models in LinDA would require:

• Substantial refactoring of the core algorithm
• Different bias correction approaches for compositional data in survival context
• Extensive validation with microbiome survival datasets

I'll keep this issue open as a feature request for potential future development. If there's sufficient community interest, this could be considered for a future major release.

In the meantime, the manual approach above should work well for your survival analysis needs while preserving LinDA's compositional data handling philosophy (CLR transformation).

Additional Resources

• Paper: "Compositional Data Analysis in Microbiome Studies" - discusses CLR transformation
• Tutorial: https://microbiome.github.io/tutorials/ - includes time-to-event analysis examples

Would you be willing to share your use case? Understanding specific applications would help prioritize this feature request.

Best regards,
Chen Yang

[DarioS]

Thank you. Alternatively, users could simply categorise their follow-up times, such as:
Good Prognosis: Alive, No Sign of Relapse for at least three years.
Poor Prognosis: Dead within one year of diagnosis.

[cafferychen777]

Hi @DarioS,

Thank you for that practical suggestion! You're absolutely right - categorizing follow-up times is a valid and often more interpretable alternative to Cox proportional hazards models.

Your proposed categorization approach:

• Good Prognosis: Alive, No Sign of Relapse for at least three years
• Poor Prognosis: Dead within one year of diagnosis

This is an excellent solution that allows you to use MicrobiomeStat's existing differential abundance framework. Here's how you can implement this:

Implementation Approach

# Create the categorized outcome variable
metadata$Prognosis <- ifelse(
  metadata$Status == "Alive" & metadata$RelapseTime >= 36, "Good",
  ifelse(metadata$Status == "Dead" & metadata$SurvivalTime <= 12, "Poor", NA)
)

# Remove intermediate cases (optional, depending on your study design)
metadata_clean <- metadata[!is.na(metadata$Prognosis), ]

# Use any MicrobiomeStat differential abundance function
# For cross-sectional analysis:
test_results <- generate_taxa_test_single(
  data.obj = your_data_obj,
  group.var = "Prognosis",
  adj.vars = c("Age", "Gender", "Treatment"), # adjust as needed
  feature.level = c("Genus"),
  feature.dat.type = "count"
)

# For visualization:
volcano_plot <- generate_taxa_volcano_single(
  data.obj = your_data_obj,
  group.var = "Prognosis",
  test.list = test_results,
  feature.sig.level = 0.05,
  feature.mt.method = "fdr"
)

Advantages of This Approach

1. Clear Clinical Interpretation: The categorization has direct clinical meaning
2. Compatible with Current Tools: Works seamlessly with all MicrobiomeStat functions
3. Robust to Violations: Avoids proportional hazards assumptions that might not hold
4. Focus on Extremes: By selecting clear good/poor outcomes, you increase effect size and statistical power

Additional Considerations

1. Sample Size: Make sure you have adequate samples in both categories after filtering out intermediate cases
2. Sensitivity Analysis: You might want to try different time cutoffs (e.g., 2 years vs 3 years for good prognosis)
3. Documentation: Clearly document your categorization criteria in your methods

This approach is well-established in prognostic biomarker studies and should work excellently with MicrobiomeStat's compositional data analysis framework.

Thanks again for the suggestion - this might be helpful for other users with similar survival analysis needs!

Best regards,
Chen