We have made progress, and published corresponding open-access manuscripts and software packages, on each one of the Aims of the project.
Signal Filtering. We developed two methods, probabilistic and spectral-based. Specifically, we developed a kernel-based method for probabilistically filtering known (measured or computationally reconstructed) signals from single-cell data to expose over-shadowed signals, which are originally harder to detect (Piran and Nitzan, Nature Communications 2024). We demonstrated that our method, SiFT, could enhance the signal of the circadian rhythm within single-cell data of liver lobules by filtering the signal of spatial zonation. In a different setting, SiFT was demonstrated to filter the healthy background signal from COVID-19 single-cell patient data to expose disease-related cells and dynamic signatures.
Additionally, we developed a spectral computational method that uses topological priors to decouple, enhance and filter different classes of biological processes (e.g. periodic or linear) in single-cell data using spectral template matching (Karin, Bornfeld and Nitzan, Nature Biotechnology, 2023). We demonstrated the use of our method, scPrisma, for the analysis of diverse processes, such as the circadian rhythm and spatial zonation, diurnal cycle in Chlamydomonas, and circadian rhythm in the suprachiasmatic nucleus in the brain.
Signal disentanglement. We developed a generative deep learning approach, biolord, to simultaneously disentangle multiple biological signals from single-cell data (Piran et al., Nature Biotechnology, 2024). The computational framework is tailored for multiple types of biological data, including single-cell RNA-sequencing. Beyond disentanglement, biolord allows for counterfactual predictions – effectively virtually shifting cells across space, time and conditions to predict cellular responses to perturbations and sample unseen cellular states. The framework was shown to reveal gene expression programs associated with liver infection, and in a different context, to predict cellular responses to unseen drugs and genetic perturbations.
For both filtering and disentanglement of signals in single-cell data, having integrated, high-quality annotations of measured cells is key. To advance this goal, we developed an optimal transport-based method for annotation transfer across biological datasets (Mages, Moriel, Avraham-Davidi, et al., Nature Biotechnology, 2023), and proposed an algorithm, based on iterative gene weight updates, to improve the robustness of computational reconstruction of single-cell data (Sheng, Barak and Nitzan, Bioinformatics, 2023).
Towards the goal of deriving biological universality classes for processes occurring in various cellular populations under different conditions, we have developed physics-informed machine learning frameworks, for the identification and classification of different dynamical regimes and bifurcations within snapshot measurements, as well as for the recovery of underlying equations or mechanisms driving such dynamics (Ricci et al., ICLR, 2023; Moriel, Ricci and Nitzan, ICLR, 2024). These works allowed us, for example, to recover distinct dynamical regimes corresponding to proliferation vs. differentiation in pancreas development from single-cell RNA-sequencing data, without the use of prior (gene-based) knowledge.
Towards identifying design principles guiding the low-dimensional structure and distributions of cells in expression space, we demonstrated how division of labor theory predicts distinct, mechanism-dependent spatial patterns. By optimizing collective cellular task performance under trade-offs, we found that distinguishable expression patterns can emerge from cell-cell interactions versus instructive signals, and demonstrated these results across multiple tissues and biological contexts (Adler et al., Cell Reports, 2023).
