Does modelling how drugs change gene expression help predict response?
A variational autoencoder that learns both how drugs shift gene expression and whether cells respond predicted drug sensitivity better than standard classifiers for over half of the drugs tested.
Source
Dr.VAE: improving drug response prediction via modeling of drug perturbation effects
Study at a glance
- Design
- Computational / modelling — Dr.VAE fitted separately for each of 26 drugs on CTRPv2 sensitivity and CMap-L1000 perturbation data; 100 train/validation/test splits; compared with ridge logistic regression, random forest, RBF-SVM, SSVAE, PCA and PertVAE embeddings
- N
- 26 drugs evaluated; cell-line counts vary by drug (up to 855 cell lines with viability data, up to 60 with perturbation data), so no single N
- Population
- Cancer cell lines from the CTRPv2 drug-sensitivity screen and the LINCS CMap-L1000 perturbation screen
- Outcome
- Test AUROC and AUPR for classifying cell lines as drug responders or non-responders
Structured fields used in claim comparison tables when every cited study has a complete layer.
What they did
The authors built Dr.VAE, a deep generative model that compresses gene expression into a low-dimensional latent space, learns how a drug moves a cell line within that space, and predicts whether the cell line responds. It was trained on drug-sensitivity labels plus before-and-after-treatment expression profiles, separately for 26 FDA-approved drugs, over 100 data splits. They compared it with ridge logistic regression, random forests and SVMs, and ran ablations: a VAE without perturbation data (SSVAE), Dr.VAE with its perturbation function replaced by an identity, and classifiers fitted on unsupervised PCA or VAE embeddings.
What they found
Dr.VAE beat all three standard classifiers on at least 14 of the 26 drugs, tied on nine, and lost on only three. It significantly beat the perturbation-free SSVAE on 9 drugs and lost on one. Swapping the learned perturbation function for an identity lowered performance on 16 drugs, but most of the gain seemed to come from a better latent representation. Standard classifiers also did better on an unsupervised VAE embedding than on PCA.
The limits
What it doesn't show
Everything is in cancer cell lines, so it says nothing directly about predicting response in patients. Only gene expression was used, and perturbations were taken only at 6 hours at one dose, which the authors note may miss slower feedback effects (as with fluvastatin). For some drugs perturbation data existed for very few cell lines, and improvements over baselines were often a few percentage points of AUROC.
Key terms
- Variational autoencoder (VAE)
- A generative neural network that encodes data into a probabilistic low-dimensional latent space and learns to decode it back.
- Semi-supervised learning
- Training that combines labelled examples with unlabelled ones, here perturbation profiles that lack response labels.
- Drug perturbation signature
- The change in a cell line's gene expression after short-term exposure to a drug.
- Ablation
- Removing or replacing one part of a model to see how much that part contributes.
- AUROC
- Area under the ROC curve: the chance a random responder is ranked above a random non-responder.
Flashcards
0 of 10 answers reviewed
Research intelligence for this paper
See its role on concept claims, tensions it is part of, placement history, and related discoveries.
Quiz yourself
On how many of the 26 drugs did Dr.VAE beat all three standard classifiers?
Common questions
Why not just use a standard neural network classifier?
There are far more genes than labelled cell lines, so discriminative networks overfit; jointly modelling the data generatively acts as a regulariser.
Did the perturbation modelling itself matter?
Partly: replacing it with an identity hurt performance on 16 drugs, but the authors conclude most of the benefit came from a better learned representation.
Why did Dr.VAE fail for fluvastatin?
Statin sensitivity depends on feedback that peaks after more than 8 hours, which 6-hour perturbation data cannot capture.
More on Generative models