WANCDR: Wasserstein Adversarial Network for Cancer Drug Response

Abstract

Predicting patient-specific drug responses from preclinical cell-line data remains challenging due to significant heterogeneity between preclinical (cell-line) and clinical (patient) gene expression profiles. In this study, we propose WANCDR, a novel adversarial neural network framework designed to improve the generalization of drug-response predictions by aligning latent representations across preclinical and clinical domains. Specifically, we introduce a domain alignment module trained adversarially, which enforces the encoder to generate domain-invariant latent embeddings. Extensive experiments conducted on preclinical (GDSC) and clinical (TCGA) datasets demonstrate that WANCDR achieves robust predictive performance on preclinical data, while substantially outperforming existing approaches in clinical generalization, particularly when classifying responses for previously unseen drugs. Qualitative analyses via UMAP visualization further validate the superior domain alignment capability of WANCDR. Collectively, these results highlight the potential of WANCDR to bridge the translational gap from preclinical insights to clinical applications.

Keywords: Graph neural network · Cancer drug response · Wasserstein Adversarial Network

Latent Alignment: UMAP Visualizations

(a) Before Training – clear domain gap; (b) DeepCDR – cluster separation is reinforced (overfits to GDSC); (c) WANCDR – reduces the gap via adversarial alignment, improving TCGA generalization.

(a) Before Training: clear gap between GDSC and TCGA — (a) Before Training — Strong domain gap between GDSC (blue) and TCGA (orange).

(b) DeepCDR: domain-specific features; overfitting to GDSC — (b) DeepCDR — Learns discriminative yet domain-specific features (overfit to GDSC).

(c) WANCDR: aligned latent space for transfer to TCGA — (c) WANCDR — Narrows the domain gap; aligned latents enable transfer to TCGA.

Why This Experiment? (Intuition)

Problem. Cell-line (GDSC) and patient (TCGA) distributions differ markedly. A model optimized only for GDSC tends to learn domain-specific features; UMAP shows separated clusters, and predictions do not transfer to TCGA.

WANCDR. We adversarially align latent representations using a Wasserstein critic with unlabeled TCGA. This reduces the source–target domain gap while preserving predictive structure, improving clinical generalization, including for unseen drug–gene pairs.

Takeaway. DeepCDR reinforces the gap (good on GDSC, weak transfer), whereas WANCDR narrows it and performs robustly on TCGA.

TCGA performance on seen vs. unseen drug–gene expression pairs (mean (std)).
Dataset	Model	AUC	ACC	Precision	Recall	F1
Seen	WANCDR	0.690(0.048)	0.660(0.029)	0.671(0.027)	0.569(0.144)	0.604(0.087)
Seen	DeepCDR	0.554(0.044)	0.595(0.045)	0.639(0.097)	0.431(0.121)	0.496(0.054)
Unseen	WANCDR	0.632(0.017)	0.622(0.015)	0.641(0.031)	0.623(0.143)	0.622(0.059)
Unseen	DeepCDR	0.534(0.040)	0.568(0.022)	0.597(0.058)	0.614(0.239)	0.573(0.107)

Comparison on preclinical (GDSC) and clinical (TCGA) datasets: WANCDR vs. DeepCDR (mean (std)).
Dataset	Model	AUC	ACC	Precision	Recall	F1
GDSC	WANCDR	0.815(0.010)	0.753(0.023)	0.284(0.019)	0.723(0.032)	0.407(0.016)
GDSC	DeepCDR	0.832(0.002)	0.762(0.017)	0.295(0.014)	0.737(0.022)	0.421(0.011)
TCGA	WANCDR	0.672(0.032)	0.644(0.024)	0.656(0.015)	0.561(0.096)	0.601(0.059)
TCGA	DeepCDR	0.550(0.030)	0.574(0.028)	0.609(0.070)	0.416(0.143)	0.496(0.054)

BibTeX

@misc{choi2025wancdr,
  author    = {HanJun Choi and Mansu Kim},
  title     = {WANCDR: Wasserstein Adversarial Network for Cancer Drug Response},
  year      = {2025},
  note      = {Manuscript}
}

Update venue/year as appropriate.