ExPO: an exposure-conditioned neural operator for L1000 signature prediction

Experimental profiling of drug-cell transcriptomic responses over dose and time is sparse and irregular, complicating discovery. We present ExPO, an exposure-conditioned neural operator that predicts full L1000 (978-gene) z-score signatures for a given compound-cell context as a continuous function of exposure. ExPO ifanstantiates a DeepONet, fusing ChemBERTa-2 molecular embeddings (with LoRA adaptation) and a trunk over sinusoidal Fourier features, so signatures can be evaluated at arbitrary dose-time pairs without regridding. Training is discovery-aligned, combining robust regression with a two-list listwise objective to optimize early-rank gene ordering, plus lightweight priors (Sobolev smoothness and an optional dose-monotonicity penalty) for pharmacologic plausibility. On a scaffold-held-out CMap/L1000 benchmark, ExPO improves over strong baselines in both accuracy and ranking (MAE 0.83 vs 0.89, Spearman 0.52 vs 0.48 vs DeepCE; NDCG@50↑/↓ 0.74/0.72 vs 0.71/0.69 for CIGER), tolerates withheld exposures (ΔMAE + 0.03 for interior interpolation; + 0.07 at edges, reduced by - 0.02 with the monotonicity prior), and transfers across cell lines (LCL-O MAE 0.90 vs 0.95; ρ 0.44 vs 0.40). Quantile heads with conformal calibration yield reliable uncertainty (test PICP 0.80/0.90), and filtering the least-confident 20% reduces MAE by ~ 18%. By modeling exposure as a field rather than discrete buckets, ExPO delivers accurate, rank-faithful, and calibrated signatures for in-silico exploration of dose-time surfaces. ExPO as a standalone predictor, along with its source code and benchmark dataset, is available at https://github.com/MLBC-lab/ExPO .