SPEAR: Predicting Gene Expression from Single-Cell Chromatin Accessibility

Single-cell multiome assays enable direct measurement of chromatin accessibility and gene expression within the same cell. Still, most experimental designs remain constrained to two (and, less commonly, three) modalities per cell. This limitation motivates computational models that can predict unmeasured layers and, simultaneously, help dissect how cis-regulatory accessibility relates to transcription at gene resolution. Existing cross-modal methods often prioritize latent alignment or modality reconstruction, making it difficult to isolate the impact of model inductive bias under a shared cis-regulatory feature definition. We present SPEAR, a configuration-driven framework for gene-centric regression of single-cell gene expression from chromatin accessibility using a fixed transcription-startsite-centered representation shared across model families. Here we show that, under identical features, splits, and evaluation, model performance stratifies reproducibly across two multiome systems (mouse embryonic development and human hemogenic endothelium), with transformer encoders achieving the strongest mean test correlations (0.546 and 0.470, respectively). Per-gene performance distributions reveal substantial heterogeneity in predictability, indicating that accessibility-driven signal is concentrated in a subset of genes across contexts. Shapley value-based feature attribution further localizes predictive signal to promoter-proximal bins, with feature importance decaying with distance from the transcription start site, supporting a promoter-centered regime of cis-regulatory control within the modeled window. Together, these results provide a controlled comparison of inductive biases for chromatin- to-expression prediction and deliver analysis-ready outputs for gene-level interpretation. SPEAR is open source and publicaly available for use at https://github.com/UzunLab/SPEAR. Supplementary data are available.