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

Area of Science:

• Computational biology
• Machine learning in drug discovery
• Transcriptomics

Background:

• Drug-cell transcriptomic profiling is crucial for drug discovery but often sparse and irregular.
• Existing methods struggle to model continuous dose-time responses effectively.

Purpose of the Study:

• To introduce ExPO, an exposure-conditioned neural operator for predicting L1000 gene signatures.
• To enable continuous prediction of transcriptomic responses across arbitrary dose-time exposures.

Main Methods:

• ExPO utilizes a DeepONet architecture, integrating ChemBERTa-2 molecular embeddings and sinusoidal Fourier features.
• Training incorporates robust regression, a listwise objective for gene ranking, and pharmacologic plausibility priors.
• The model predicts gene signatures as a continuous function of compound, cell context, dose, and time.

Main Results:

• ExPO outperforms strong baselines in accuracy (MAE 0.83) and gene ranking (Spearman 0.52) on a held-out benchmark.
• The model demonstrates robustness to withheld exposures and generalizes across cell lines.
• Conformal calibration provides reliable uncertainty estimates, improving prediction accuracy when low-confidence predictions are filtered.

Conclusions:

• ExPO accurately predicts drug-cell transcriptomic signatures across continuous dose-time exposures.
• The model facilitates in-silico exploration of dose-time surfaces, enhancing drug discovery pipelines.
• ExPO offers a powerful, flexible tool for analyzing complex drug response data.