In Vitro to In Vivo: Bidirectional and High-Precision Generation of In Vitro and In Vivo Neuronal Spike Data

Abstract Spontaneous neural activity is often dismissed as noise, yet predicting its future evolution remains difficult, especially across recording preparations. We study transfer between in vitro and in vivo multineuronal spike trains with a Transformer that treats spikes as sparse point processes and does not require cell-to-cell correspondence across datasets. Because binned spike trains are extremely sparse and imbalanced, likelihood-based objectives can be dominated by the majority no-spike class under domain shift. Introducing a Dice objective that maximizes spike-event overlap stabilizes training and substantially improves performance, including in cross-domain prediction and generation. Training on one domain and evaluating on another yields a quantitative predictability score and a predictability matrix. We show that (i) a performance-based embedding organizes datasets by transferability; (ii) only specific in vitro and in vivo pairs support reliable cross-domain generation, whereas mismatched pairs degrade despite strong within-domain accuracy; (iii) attention and gradient-based importance analyses reveal how the model shifts emphasis across neurons and time; and (iv) the predictability matrix guides source–target selection for generating diverse activity patterns. This framework supports systematic tests of transferability and may help prioritize datasets for translational studies.