Representation learning
Do deep learned features help predict disease from gene data?
Open access · cc by · source: Europe PMC
Across dozens of prediction tasks from gene expression data, fancy learned representations did not beat a simple regularised linear model on properly normalised data.
Study at a glance
- Design
- Computational / modelling — Systematic benchmark of normalisations, gene sets, learned representations and classifiers across many RNA-seq prediction tasks using 5-fold nested cross-validation.
- N
- About 37k samples trained the unsupervised representations; supervised tasks (24 classification, 26 survival) each had typically a few hundred samples.
- Population
- Human RNA-seq samples from the recount2 database (GTEx, TCGA and SRA).
- Outcome
- Out-of-sample AUC (binary), accuracy (multiclass) and concordance index (survival).
Structured fields used in claim comparison tables when every cited study has a complete layer.
Key findings
Models using many genes beat those using the single best gene, and log-ratio normalisation beat raw relative abundances. Representations from PCA, autoencoders and VAEs lowered out-of-sample performance, and the semi-supervised version helped slightly only on tasks similar to its training data. Random forest won 13 tasks and logistic regression nine, but logistic regression was more consistent, so an l2-regularised linear model on centred log-ratio data is recommended as the baseline.
Methodology
Using public RNA-seq data, the authors set up 24 classification tasks and 26 survival tasks and compared thousands of models with 5-fold nested cross-validation. They varied the normalisation, the gene set, whether features came from PCA, a denoising autoencoder, a variational autoencoder or none, and the classifier (k-nearest neighbours, random forest, regularised logistic regression). They also tested a semi-supervised autoencoder trained jointly with several prediction tasks.
Limitations
The result covers bulk RNA-seq phenotype prediction with the specific architectures and 512-dimensional embeddings tested; other designs or much larger labelled datasets might change it. Supervised tasks had only a few hundred samples each, which may favour simple models. The authors suggest simple methods may already be near the limit of what bulk expression can predict, but they do not test that directly.
How this study connects
Role on claims
Each row is a claim on a concept or method page where this paper supports, challenges, or qualifies the statement. Roles are hand-checked — not a model guess.
Learned representations are not always better than raw features with a simple model.
For phenotype prediction from bulk transcriptomics, PCA, autoencoder and VAE representations lowered out-of-sample performance compared with regularised linear models or random forests on log-ratio normalised data.
Evidence for the claim as stated.
Pretrained embeddings helped for CT imaging, protein and metagenome tasks but hurt for bulk transcriptomic phenotype prediction with a few hundred labelled samples.
Same question, contrary or null result.
Open questions
Tensions this paper is part of
From concept pages' “where studies disagree.” Disagreement means the same question; scope means different assays, populations, or outcomes.
Pretrained embeddings helped for CT imaging, protein and metagenome tasks but hurt for bulk transcriptomic phenotype prediction with a few hundred labelled samples.
Discoveries this paper informs or conflicts with
- On a few hundred samples, cross-validation flatters models and deep features lose to linear baselines
This paper informs this development.
Related papers in this topic
Same topic cluster — not a recommendation engine.
- Can a small protein language model match much bigger ones?
- Can a transformer predict which drugs bind which proteins?
- Can one language model read protein sequence and structure?
- Can random negative pairs teach a better genome-sorting model?
- Can a self-taught CT model find cancer markers with little data?