Attention-based hierarchical graph autoencoder for dose-specific single-cell resistance dynamics

Understanding cellular adaptation to drug concentrations is critical for modeling therapeutic resistance. Current perturbation models treat doses as categorical variables and lack biological hierarchy, limiting their ability to model continuous dose-response trajectories and capture pathway-level regulatory rewiring.

We introduce a hierarchical graph attention autoencoder that models dose-specific single-cell transcriptional responses by integrating pathway, gene, and cell hierarchies with Long Short-Term Memory (LSTM)-based dose conditioning. Applied to GSE206125 (BRCA2-deficient ovarian cancer cells exposed to escalating olaparib concentrations T1-T320), our model reconstructs multiscale embeddings while predicting resistance phenotypes. Five-fold cross-validation with held-out doses demonstrates robust dose-level generalization. When trained on T1-T160 only, the model extrapolates to held-out dose T320 with high fidelity (cosine similarity: 0.98) and produces a plausible latent space trajectory extension to T640, beyond the training range. LSTM-based encoding captures non-linear structure in the learned latent dose space, including a marked transition between T40 and T80, suggesting non-linear transcriptional dynamics. Multi-head attention mechanisms consistently prioritize resistance-associated pathways (TP53-regulated transcription, oxidative stress signaling) and key regulatory genes (CDKN2A, YBX1). Ablation studies confirm that virtual nodes, attention mechanisms, and LSTM dose encoding each contribute uniquely to reconstruction fidelity and biological coherence.

Our framework enables interpretable modeling of dose-dependent transcriptional trajectories with demonstrated extrapolation capability. While applied here to PARP inhibitor response in ovarian cancer, the architecture is designed to accommodate other dose-escalation studies as appropriately structured datasets become available.