EFGPP: Exploratory framework for genotype-phenotype prediction

Technical Summary: EFGPP – An Exploratory Framework for Genotype-Phenotype Prediction

Problem Statement
The prediction of complex phenotypes from human genetic data is increasingly constrained not by data availability, but by the challenge of prioritizing and integrating heterogeneous sources of signal. The field possesses a wide array of genetic representations (e.g., genotype matrices, principal components), Polygenic Risk Score (PRS) tools, and source Genome-Wide Association Studies (GWAS), yet lacks a systematic, reproducible framework to determine which combinations are most informative for a given phenotype under realistic data limitations. This is particularly critical for complex traits with moderate polygenic signals, where the marginal contribution of each data modality is small, and selection decisions disproportionately impact out-of-sample performance. Furthermore, the potential value of integrating cross-trait data (e.g., depression-derived inputs for migraine prediction) remains uncertain, as indiscriminate feature accumulation can lead to redundancy and overfitting without guaranteed performance gains.

Methodology: The EFGPP Framework
The authors present EFGPP (Exploratory framework for genotype-phenotype prediction), a systematic workflow designed to generate, rank, prune, and integrate heterogeneous genetic data representations. The framework operates through six main stages:

1. Data Source Assembly: Heterogeneous inputs are assembled from target cohorts and external discovery resources, including genotype data, covariates (clinical and metabolomic), principal components (PCA), functional annotations, and GWAS summary statistics.
2. Dataset Generation: Multiple individual datasets are constructed, including:
   • Covariates (e.g., NMR metabolomic biomarkers, comorbid conditions).
   • PCA representations to account for population stratification.
   • Genotype-derived matrices (weighted/unweighted, annotated/non-annotated) derived from GWAS summary statistics using P-value thresholding.
   • PRS representations generated using four distinct tools: PLINK, PRSice-2, ANNOPRED, and LDAK-GWAS.
3. Ranking and Pruning: Candidate datasets are benchmarked across 12 machine learning algorithms (including logistic regression, ensemble methods, and neural networks) using a fold-wise evaluation design. Redundant representations are removed using similarity-based pruning (Kolmogorov–Smirnov statistic).
4. Category Selection: The strongest retained datasets are grouped into representative families (covariates, PCA, genotype, PRS) and ranked using a composite score that incorporates validation AUC, generalization gaps, and stability metrics.
5. Multimodal Search: Shortlisted datasets are systematically combined (2–5 datasets) to evaluate if integrated representations improve performance beyond single modalities. Combinations are filtered based on validation performance.
6. Final Evaluation: The best individual and multimodal models are compared using cross-validated performance metrics on held-out test subsets to quantify gains from structured integration.

Experimental Setup
The framework was evaluated using UK Biobank data from 733 individuals (53 migraine cases, 680 controls). Two configurations were tested:

• Configuration 1 (Migraine-focused): Utilized migraine-derived GWAS inputs only.
• Configuration 2 (Cross-trait): Expanded the search space to include depression-derived GWAS and PRS inputs to test the utility of related-trait information.

The study employed a strict fold-wise design where all data-dependent processing (dataset construction, model selection, pruning) occurred within training/validation folds, reserving held-out test sets exclusively for final evaluation to prevent information leakage.

Key Results

• Single-Modality Performance: No single genetic representation matched the performance of a covariates-only baseline (Test AUC = 0.639). The best individual genetic representation was a weighted genotype dataset derived from migraine GWAS (Test AUC = 0.644). Among PRS tools, PLINK performed best for migraine-derived scores, but depression-derived PRS (specifically LDAK-GWAS) outperformed all migraine-derived PRS configurations (Test AUC = 0.564 vs. 0.521), suggesting shared genetic architecture.
• Multimodal Integration: Structured integration significantly improved performance.
  • In Configuration 1, the best multimodal model (combining covariates, PCA, and a weighted non-annotated migraine genotype matrix) achieved a Test AUC of 0.688, outperforming the best single modality (0.644) and the covariate baseline (0.639). Notably, the top-performing models in this configuration did not include PRS features, favoring direct genotype matrices.
  • In Configuration 2, the best model (combining unweighted annotated migraine genotype and depression-derived LDAK-GWAS PRS) achieved a Test AUC of 0.663. While competitive, this was lower than the migraine-focused configuration, indicating that cross-trait data must be empirically prioritized rather than assumed to improve performance.
• Benchmarking Context: The best EFGPP multimodal result (0.688) exceeded the best single-tool full-model result (0.629) from a companion benchmark of 46 PRS tools, demonstrating that structured multimodal integration offers measurable gains beyond individual PRS tools.

Significance and Claims
The authors position EFGPP not as a clinically deployable prediction model, but as a reproducible proof-of-concept framework for systematic data prioritization. The study claims that for complex trait prediction, the primary challenge is often data prioritization rather than data scarcity. EFGPP provides a structured decision-making process to reduce large representation spaces into a smaller set of evidence-based modeling choices.

The paper emphasizes that:

1. Integration is key: Structured multimodal integration can improve prediction beyond any single data modality, even in modestly sized cohorts.
2. Prioritization matters: Disciplinary feature prioritization (selecting the right combinations) is more critical than raw feature volume; adding more datasets does not guarantee better generalization.
3. Cross-trait utility: Related-trait inputs (e.g., depression for migraine) can contain useful predictive signal, but their inclusion requires empirical prioritization within a structured framework rather than indiscriminate addition.

The authors conclude that EFGPP serves as a practical foundation for future studies to move from opportunistic feature assembly toward systematic, evidence-based construction of genotype-phenotype prediction models, while acknowledging limitations regarding sample size, lack of external validation, and the need for further calibration and benchmarking against larger cohorts.