Predicting Neuronal Projections from Gene Expression under Incomplete Detection

A central challenge in neuroscience is to connect what a neuron is molecularly to where it sends its axons in the brain. Predicting neuronal projections from gene expression could make this connection computationally, extending sparse anatomical tracing to the large cell populations captured by molecular brain atlases. Joint measurements of gene expression and projections make this goal increasingly accessible, but assays miss existing projections and measure different subsets of genes and projection targets. As a result, an assay zero can mean either that a projection is absent or that it was simply missed, so models trained directly on assay calls can learn detectability rather than the underlying probability of projection. We introduce Gene2Wire, a multi-target positive-unlabeled model that separates projection probability from assay detectability while sharing information across incompletely observed targets. A paired training subset estimates detection sensitivity, and a calibrated likelihood uses these estimates to learn projection probabilities from incomplete readouts. Gene2Wire further combines molecular effects shared across projection targets with target-specific effects, allowing related targets to borrow statistical strength without forcing distinct projection patterns to follow the same molecular rule. We show theoretically how errors in detection calibration propagate to projection estimates and how this shared-plus-specific parameterization exploits cross-target structure without restricting the predictor to be low rank. Controlled simulations show that Gene2Wire adapts to target relationships ranging from distinct molecular effects to shared predictive structure. Across five real projection–transcriptomic assay panels, Gene2Wire outperforms previous state-of-the-art models in both projection ranking and probability estimation across varying gene coverage, target coverage, and detection sensitivity.