Predicting Cellular Responses to Novel Drug Perturbations at a Single-Cell Resolution

1 Introduction

• Single-cell HTSs reveal subtle gene-expression changes and cellular heterogeneity, but costly library preparation limits screens to no more than 200 drugs in nuclear hashing.The paper frames this cost and scale constraint as a barrier to screening thousands of molecules.Source: §1 · p. 1
• chemCPA combines compound-structure representations with bulk RNA HTS transfer to improve single-cell generalization for unseen compounds.The authors describe the model as flexible and interpretable like CPA while supporting lower-resolution, higher-throughput bulk assays.Source: §1 · p. 2
• chemCPA outperforms existing methods on unobserved drug-covariate combinations and is evaluated on generalization tasks unavailable to prior methods.The stated contributions include transfer learning with identical and differing source and target gene sets.Source: §1 · p. 2

2 Related Work

• Deep learning can model nonlinear, cell-type-specific responses and genome-wide measurements, and CPA models perturbations in scRNA-seq data.CPA does not generalize to unseen compounds, limiting virtual screening of drugs not measured by scRNA-seq [Lotfollahi et al. (2021)].Source: §2 · p. 2
• Bulk RNA methods predict gene-expression profiles for de novo chemicals, but translating them to single-cell datasets with fewer compounds and often different gene sets remains unclear.Source: §2 · p. 2

3 Chemical Compositional Perturbation Autoencoder

• The architecture combines gene-expression encoding and decoding, attribute embedders, and adversarial classifiers to generate counterfactual expression states from chosen attribute combinations.Source: §3 · p. 3, Figure 1 · p. 3

3.1 Gene encoder and decoder

• The encoder maps measured expression to a latent basal state, and the decoder reconstructs gene-expression distributions after attribute updates.The encoder is an MLP; the decoder supports negative-binomial or Gaussian distributions, with the authors reporting better convergence using a Gaussian likelihood.Source: §3.1 · p. 3

3.2 Attribute embedding and additive latent space

• Drug and cell-line embeddings are added to the basal state, with dosage scaling the drug embedding; this linear structure supports component-wise analysis and dose interpolation or extrapolation.The perturbation network maps molecular representations and dosage into a latent drug effect, allowing prediction for molecules absent from training.Source: §3.2 · pp. 3–4

3.3 Adversarial classifiers for invariant basal states

• Adversarial classifiers predict drug and cell-line attributes from the basal state, while reversed classification loss trains the encoder to remove that attribute information.The authors describe this disentanglement as an approximation that makes the problem tractable and supports attributing effects to specific sources.Source: §3.3 · p. 4

4 Datasets and transfer learning

• The main evaluation uses sci-Plex3, with L1000 used for bulk pretraining.The datasets overlap in cell lines and many compounds, while L1000 spans a more diverse molecular space.Source: §4 · p. 5
• L1000 contains about 1.3 million bulk RNA observations across 978 genes and almost 20,000 drugs, while sci-Plex3 measures 649,340 cells across 7,561 genes and 188 drugs.sci-Plex3 covers three human cancer cell lines and four dosages; all cell lines and about 150 compounds overlap with L1000.Source: §4 · p. 5
• Transfer experiments use 977 shared genes or expand the target input and output to 2,000 genes by adding non-linear layers for fine-tuning.The expanded set combines the 977 L1000 genes with 1,023 highly variable sci-Plex3 genes; all layers, including the added ones, are trained during fine-tuning.Source: §4 · p. 5

5 Experiments

• The evaluation measures counterfactual predictive performance on all genes and the 50 most differentially expressed genes, alongside disentanglement of latent representations.Separate classifiers assess whether the basal state retains drug or cell-line information; models are filtered using disentanglement thresholds.Source: §5 · pp. 5–6

5.1 Comparing chemCPA against existing methods on unseen drug-covariate combinations

• ChemCPA outperforms scGen and CPA on unseen drug–covariate combinations, while its base version also exceeds both methods.The comparison covers 1 µM and 10 µM; the authors attribute the strongest performance to the perturbation network together with pretraining.Source: Table 1 · p. 6, §5.1 · p. 7

5.2 Using chemCPA to predict single-cell responses for unseen drugs

• On the shared gene set, pretrained chemCPA consistently outperforms the baseline and base model across drugs and three cell lines.The discussion reports that predictions improve over the baseline at high dosages, with pretraining especially advantageous across all genes.Source: §5.2 · pp. 7–8, Table 2 · p. 7
• On the extended gene set, pretrained chemCPA retains an advantage, while the base model is only slightly better than the baseline.The extension adds 1,023 highly variable genes to the 977 shared genes, making it more difficult for pretrained models to use learned bulk expressions directly.Source: §5.2 · pp. 7–8, Table 3 · p. 8

5.3 Measure uncertainty on the drug embedding

• The uncertainty score for unseen drugs correlates with accurate prediction of perturbed cells, combining MoA confidence with expected perturbation distinctness.The measure uses the entropy of pathway predictions from neighbouring drugs and inverse distance to neighbours.Source: §5.3 · pp. 9–10, Table 4 · p. 9

• chemCPA predicts cellular gene-expression responses for unseen drug perturbations by encoding drug molecular structures.Source: §6 · p. 10
• chemCPA outperforms CPA and scGen on shared tasks and extends generalisation to unseen drugs.Source: §6 · p. 10
• Pretraining on bulk HTSs improves chemCPA’s single-cell generalisation, even when source and target gene sets differ.Source: §6 · p. 10
• chemCPA’s uncertainty measurement correlates with its ability to generalise to unseen drugs.Source: §6 · p. 10

• On L1000 DEGs, the weave model achieved an r2-score of 65 ± 8, while JT-VAE and GROVER outperformed GCN.Source: Table 5 · p. 12, §A.1 · p. 12
• With identical genes, fine-tuned GROVER and RDKit models consistently beat the baseline and non-pretrained RDKit model, while fine-tuned JT-VAE beat the baseline but trailed its non-pretrained version.Source: §A.1 · p. 12, Table 6 · p. 13
• On the extended-gene sci-Plex3 transfer task, pretrained chemCPA with RDKit performed best; the authors attribute this partly to its distinct histone-deacetylation-drug cluster.Source: §A.1 · p. 12, Table 7 · p. 13

• For each test combination, chemCPA combines its drug, dose, and cell-line latent attribute state with basal states from control observations.Source: §A.3 · p. 13
• The method averages predicted and real measurements per gene, producing two n-dimensional vectors, where n is 977 or 2000 genes.Source: §A.3 · p. 13
• An r2 score between those vectors yields one score per test combination.Source: §A.3 · p. 13

A.4 Additional information on the L1000 experiment

• The L1000 appendix points to the RDKit parameter sweep, best encoder configurations, and best-run performance tables.Source: §A.4 · p. 13, Table 10 · p. 15, Table 11 · p. 15

A.5 Additional information on the sci-Plex3 experiments

• The sci-Plex3 figures compare pretrained and non-pretrained chemCPA using GROVER or JT-VAE across complete or extended gene sets, with additional drug-wise comparisons.Source: Figure 6 · p. 14, Figure 7 · p. 16, Figure 8 · p. 16

A.6 Additional information on the uncertainty score

• The uncertainty appendix shows the uncertainty score against improvement over baseline r2 and includes drug-embedding and drug-wise comparisons.Source: Figure 10 · p. 16, Figure 13 · p. 18, Figure 15 · p. 19