Skip to content

PredictDB

Summary

PredictDB is a database hosting genetic prediction models of tissue-specific gene expression and splicing, developed by the Im Lab. It is primarily designed to facilitate transcriptome-wide association studies (TWAS) using the MetaXcan (and S-PrediXcan) framework. PredictDB models map genetic variants (SNPs) to molecular traits to impute gene expression in cohorts where only genotype data is available.

Core Models and Data

PredictDB models are trained on datasets with paired genotype and transcriptome data: - GTEx Bulk Gene Expression: Comprises tissue-specific transcriptome models across 49 human tissues trained using GTEx V8 (17,382 RNA-seq samples from 948 donors across 54 tissues). - Training Methodologies: Built using statistical models such as Elastic Net and MASHR (Multivariate Adaptive Shrinkage) to capture genetic effects on expression and splicing.

Model Family Comparison (MASHR vs. Elastic Net)

A systematic comparison of four transcriptome-prediction strategies — Elastic Net (EN-M), cross-tissue CTIMP (CTIMP-M), fine-mapping-weighted Elastic Net (DAPGW-M, using dap-g posterior inclusion probabilities to down-weight the penalty on likely-causal variants), and MASHR (MASHR-M, applying multivariate adaptive shrinkage to dap-g-selected variants' effect sizes) — found that raw cross-validated prediction accuracy is not the right criterion for choosing a model family, since the goal of TWAS is reliable causal gene discovery, not expression-prediction accuracy per se.

  • Fine-mapping-informed families produced far more usable models (518,537 for DAPGW-M, 84% more than EN-M; 686,241 for MASHR-M, 143% more than EN-M).
  • On independent GEUVADIS lymphoblastoid-cell-line data, MASHR-M, DAPGW-M, and CTIMP-M all modestly outperformed EN-M in prediction correlation (mean Spearman ρ gains of 0.028, 0.027, and 0.018 respectively over EN-M).
  • Colocalization rate: among significant gene-trait associations, MASHR-M showed the highest colocalization rate with GWAS signal (11.7% via enloc, vs. 8.18–9.01% for the other three families) while also producing more prioritized candidates overall (36.6 genes/trait on average for MASHR-M vs. 25.5–36.7 for the others) — a result that held (with a smaller margin) when re-validated with an independent colocalization method (coloc) to rule out dap-g circularity.
  • Silver-standard gene recovery: against OMIM-curated (1,592 associations) and rare-variant-based (101 associations) gold-standard gene-trait pairs, MASHR-M had the best ROC/precision-recall performance for OMIM (DAPGW-M a close second); the smaller rare-variant standard could not distinguish among the four families.

These results led the authors to adopt MASHR as the default PredictDB model family for expression and splicing prediction across GTEx tissues, distributed publicly at predictdb.org.

Cross-Ancestry Model Comparison: Elastic Net vs. Non-Linear ML

Okoro, Schubert, Guo, et al. (2021) asked whether non-linear/non-parametric machine learning algorithms — random forest (RF), support vector regression (SVR), K-nearest-neighbors (KNN) — could out-predict the elastic net (EN) model used by PredictDB/PrediXcan, and whether cross-population portability differs by algorithm. Models were trained on blood monocyte genotype/transcriptome data from the Multi-Ethnic Study of Atherosclerosis (MESA; African American, Hispanic American, and European American participants) and tested out-of-cohort on the Modeling the Epidemiology Transition Study (METS; African-ancestry participants from Ghana and Chicago).

  • EN generally wins, but not uniformly: across all training populations, EN had the highest average cross-validated R² (ALL cohort: EN=0.0733, SVR=0.0476, RF=0.0409, KNN=0.0103), but RF matched or exceeded EN for a subset of genes — particularly when the training and test ancestries were more disparate — suggesting RF's non-linear representation may be more robust, if less accurate on average, when transferring across populations.
  • Ancestry match dominates algorithm choice: regardless of which of the four algorithms was used, prediction accuracy in the held-out METS (African-ancestry) cohort was consistently highest for models trained on ancestry-matched or ancestry-similar MESA populations (AFA, or the combined AFA+HIS "AFHI" set) than for models trained on the more distant CAU population — reinforcing that training-population diversity, not model family, is the primary lever for cross-ancestry transcriptome prediction portability. This is directly relevant to why MASHR/MetaXcan-family TWAS results are less reliable in non-European ancestries and motivates multi-ancestry model development (see Multi-Ancestry Proteome-Wide Association Studies for an analogous protein-level effort).
  • RF recovers TWAS hits EN misses: applying EN- and RF-trained expression models to a TWAS of HDL cholesterol in held-out MESA genotypes, both EN and RF (and SVR) independently and concordantly recovered the well-established CETP-HDL association (EN p=4.1×10⁻¹¹, RF p=2.1×10⁻¹³, SVR p=3.9×10⁻⁸; KNN was not significant, p=0.016) — but RF also built a usable expression model for 16 genes with no corresponding EN model, including ST8SIA4, which showed a suggestive (sub-Bonferroni) HDL association (p=3.2×10⁻³) invisible to an EN-only TWAS.

The practical implication for PredictDB users is that a purely elastic-net-based (or MASHR) model catalog can systematically miss real gene-trait associations for genes whose cis-genetic architecture is better captured by a non-linear model, particularly outside the European ancestries dominating GTEx.

Citations

[1] Barbeira, A.N., Melia, O.J., Liang, Y., Bonazzola, R., Wang, G., Wheeler, H.E., Aguet, F., GTEx Consortium, Ardlie, K.G., Wen, X., & Im, H.K. Fine-mapping and QTL tissue-sharing information improve the reliability of causal gene identification. Genetic Epidemiology (NIH author manuscript; DOI not present in the retrieved copy). Source paper: nihms-1640792.pdf - Barbeira, A. N., et al. (2018). Exploring the phenotypic consequences of tissue specific gene expression variation inferred from GWAS summary statistics. Nature Communications, 9, 1825. - Gusev, A., et al. (2016). Integrative approaches for large-scale transcriptome-wide association studies. Nature Genetics, 48(3), 245–252. [2] Okoro, P.C., Schubert, R., Guo, X., Johnson, W.C., Rotter, J.I., Hoeschele, I., Liu, Y., Im, H.K., Luke, A., Dugas, L.R., & Wheeler, H.E. (2021). Transcriptome prediction performance across machine learning models and diverse ancestries. Human Genetics and Genomics Advances, 2(2), 100019.