Ensemble TWAS Weights via Stacked Regression

Minimal extension to pecotmr for learning optimal ensemble combination weights across multiple TWAS prediction methods and datasets.

Why

pecotmr offers 10+ TWAS weight methods. Each is best for some genetic architectures but worse for others. Currently users face two bad options:

1. Pick one method — arbitrary, sub-optimal for most genes
2. Run several and test all — pay a multiple-testing penalty that hurts power

Ensemble learning is a statistically principled answer: produce one weight vector per gene by learning the optimal convex combination of methods, and run a single downstream test.

What this does

Given TWAS weight predictions from K methods (e.g., SuSiE, LASSO, Elastic Net, mr.ash, BayesR, ...), learn non-negative combination coefficients that sum to 1, maximizing cross-validated prediction R². This follows the SR-TWAS stacked regression approach.

Key features:

• Uses 11 prediction methods by default
• Learns optimal method combination via constrained QP (quadprog)
• Preserves per-method R² for comparison
• Supports multi-dataset ensemble (e.g., CUMC1 + MIT → one combined weight)
• Minimum changes to pecotmr (~100 lines new, 1 line modified)

Repository structure

.claude/skills/twas-ensemble/SKILL.md    # Claude Code skill (/twas-ensemble)
R/ensemble_weights.R                      # The ensemble_weights() function
tests/testthat/test_ensemble_weights.R    # Unit tests (30 tests, pecotmr style)
scripts/test_ensemble_twas.R              # Simulation test using simxQTL
patches/pecotmr_ensemble.patch            # Patch for pecotmr integration

Status

• Patch applies cleanly to StatFunGen/pecotmr@main
• R syntax check: passes
• Unit tests: 30/30 pass (1 skipped without glmnet)
• Sanity check (3 scenarios): coefficients non-negative and sum to 1, best method gets most weight, multi-dataset combination verified, weight combination numerically correct

Quick start

1. Install dependencies

remotes::install_github("StatFunGen/pecotmr")
remotes::install_github("StatFunGen/simxQTL")
install.packages(c("glmnet", "ncvreg", "L0Learn", "quadprog", "qgg"))

2. Single-dataset ensemble

library(pecotmr)
source("R/ensemble_weights.R")  # or apply the patch to pecotmr

# After running twas_weights_pipeline with CV:
res <- twas_weights_pipeline(X, y, cv_folds = 5, weight_methods = methods)

# Learn ensemble weights
ens <- ensemble_weights(
  cv_results = res$twas_cv_result,
  Y = y,
  twas_weight_list = res$twas_weights
)
ens$method_coef           # combination weights (sum to 1)
ens$method_performance    # per-method R²
ens$ensemble_performance  # ensemble R² and correlation
ens$ensemble_twas_weights # final combined weight vector

3. Multi-dataset ensemble

# Run pipeline on each dataset
res1 <- twas_weights_pipeline(X1, y1, cv_folds = 5, weight_methods = methods)
res2 <- twas_weights_pipeline(X2, y2, cv_folds = 5, weight_methods = methods)

# Learn joint weights across datasets
ens <- ensemble_weights(
  cv_results = list(res1$twas_cv_result, res2$twas_cv_result),
  Y = list(y1, y2),
  twas_weight_list = list(res1$twas_weights, res2$twas_weights)
)

4. Claude Code skill

From this repo directory, use the /twas-ensemble skill:

/twas-ensemble 500 200

11 Default Methods

#	Method	pecotmr function	Type
1	SuSiE	susie_weights	Bayesian fine-mapping
2	LASSO	lasso_weights	L1 penalty
3	Elastic Net	enet_weights	L1 + L2 penalty
4	mr.ash	mrash_weights	Adaptive shrinkage
5	BayesR	bayes_r_weights	4-component mixture
6	BayesL	bayes_l_weights	Laplace prior
7	BayesA	bayes_a_weights	t-distribution prior
8	BayesC	bayes_c_weights	Spike-and-slab
9	SCAD	scad_weights	Non-convex penalty
10	MCP	mcp_weights	Minimax concave
11	L0Learn	l0learn_weights	L0 regularization

References

• SR-TWAS: Dai et al. (2024). Nature Communications. doi:10.1038/s41467-024-50983-w
• pecotmr: github.com/StatFunGen/pecotmr
• simxQTL: github.com/StatFunGen/simxQTL
• SuSiE: stephenslab.github.io/susieR