TL;DR
This paper introduces the Compositional Perturbation Autoencoder (CPA), a novel deep learning model that learns interpretable representations of cellular responses to complex drug and genetic perturbations. CPA can accurately predict the effects of unseen combinations of treatments, enabling more efficient experimental design and hypothesis generation in high-throughput screening.
What Problem Does This Paper Address?
High-throughput screening (HTS) experiments using single-cell RNA sequencing (scRNA-seq) are revolutionizing our understanding of how cells respond to various perturbations, such as drugs or genetic modifications. However, the sheer number of possible combinations of these perturbations (e.g., multiple drugs, varying doses, different cell types) makes exhaustive experimental screening infeasible. This combinatorial explosion poses a significant challenge for identifying optimal therapeutic strategies or understanding complex biological interactions. Existing computational methods often struggle with interpretability, scalability, or the ability to predict responses to novel, unseen combinations of perturbations (Out-Of-Distribution or OOD data).
Methods
The core innovation is the Compositional Perturbation Autoencoder (CPA). CPA is a deep learning model that builds upon autoencoder architectures. Its key features include:
- Factorized Latent Space: CPA decomposes the cellular response into interpretable latent embeddings representing the cell's basal state, the specific perturbation(s) applied, and other experimental covariates (like cell type or species).
- Disentanglement: An adversarial training approach is used to ensure that the basal state embedding is disentangled from perturbation and covariate information, meaning it doesn't reveal what treatment was applied.
- Compositionality: The model learns to combine these disentangled embeddings additively to reconstruct the observed gene expression. This allows for the prediction of novel combinations by simply combining their respective learned embeddings.
- Interpretable Embeddings: The learned embeddings for drugs and cell types can be visualized and analyzed to understand drug similarity and biological relationships.
- Dose/Time Response Modeling: CPA incorporates flexible neural networks to model continuous covariates like drug dose or time, allowing for non-linear dose-response curves.
Datasets
The authors evaluated CPA on several publicly available datasets, showcasing its versatility across different experimental scenarios:
- Sci-Plex 2: Human lung adenocarcinoma cells treated with various doses of four drugs (dexamethasone, Nutlin-3a, BMS-345541, or vorinostat/SAHA). This dataset was used to test dose-response prediction.
- 96-plex scRNA-seq: HEK293T cells under combinations of four drugs. This dataset was used to evaluate predictions for drug combinations.
- Cross-species dataset (Hagai et al.): Phagocytes from mouse, rat, pig, and rabbit treated with LPS over time. This dataset was used to assess the model's ability to handle multiple covariates and time-series data.
- Perturb-seq dataset (Norman et al.): K562 cells with single and combinatorial genetic perturbations (gene overexpression). This dataset was used to study genetic interactions.
Key Insights and Ideas
- Predicting the Unseen: CPA demonstrates a remarkable ability to predict cellular responses to drug and genetic combinations that were not present in the training data, a crucial capability for guiding experimental design.
- Interpretable Drug Embeddings: The learned latent representations of drugs cluster based on their known mechanisms of action (e.g., epigenetic regulators, tyrosine kinase inhibitors), suggesting that CPA captures biologically relevant information.
- Uncertainty Quantification: The model provides a measure of uncertainty for its predictions, highlighting when predictions are less reliable, which is essential for practical applications.
- Modeling Complex Interactions: CPA can capture not only single-gene effects but also the complex synergistic and epistatic interactions between multiple genetic perturbations.
- Cross-Dataset Generalization: The learned drug embeddings show consistency across different datasets and modalities (e.g., bulk RNA-seq vs. scRNA-seq), indicating the discovery of universal relationships.
Key Results
- CPA achieved high R² scores (a measure of prediction accuracy) on various OOD prediction tasks, often outperforming simple baselines. For example, on the sci-Plex 2 dataset, CPA achieved R² scores of 0.92 for Nutlin and 0.94 for BMS on OOD splits, compared to linear interpolation baselines of 0.85 and 0.89, respectively.
- The model successfully predicted synergistic and epistatic genetic interactions, recapitulating known biological phenomena and proposing new hypotheses.
- CPA was able to generate in-silico predictions for 5,329 missing genetic combination perturbations, representing 97.6% of all possibilities, with diverse predicted interaction types.
- The learned latent space for drugs clustered based on pathway targets, with epigenetic, tyrosine kinase signaling, and cell-cycle regulation compounds forming distinct groups.
Limitations and Future Work
- Data Dependency: Like all machine learning models, CPA's performance is dependent on the training data. If test data is significantly different from the training data, predictions may be unreliable. The authors address this by implementing uncertainty quantification.
- Scalability to Very Few Combinations: While CPA excels with large, diverse datasets, its performance can degrade if trained on very few combinations, especially for predicting differentially expressed genes (DEGs).
- Beyond Gene Expression: The current model focuses solely on gene expression. Future work could extend CPA to incorporate other modalities like epigenomic data (ATAC-seq, CUT&Tag) or spatial transcriptomics for a more comprehensive understanding.
- Multimodal Integration: The authors suggest that CPA could be extended to include other modalities for a more holistic view of cellular responses.
- Mechanistic Understanding: While CPA provides interpretable embeddings, further work could focus on deeper mechanistic insights derived from the model's predictions.
Summary updated