Abstract Recent barcoding technologies allow reconstructing lineage trees while capturing paired single-cell RNA-sequencing (scRNA-seq) data. Such datasets provide opportunities to compare gene expression memory maintenance through lineage branching and pinpoint critical genes in these processes. Here we develop Permutation, Optimization, and Representation learning based single Cell gene Expression and Lineage ANalysis (PORCELAN) to identify lineage-informative genes or subtrees where lineage and expression are tightly coupled. We validate our method using synthetic data and apply it to recent paired lineage and scRNA-seq data of lung cancer in a mouse model and embryogenesis of mouse and C. elegans. Our method pinpoints subtrees giving rise to metastases or new cell states, and genes identified as most informative about lineage overlap with known pathways involved in lung cancer progression. Furthermore, our method highlights differences in how gene expression memory is maintained through divisions in cancer and embryogenesis, thereby providing a tool for studying cell state memory through divisions across biological systems.
Introduction
In the burgeoning field of single-cell biology, researchers can now capture two distinct "layers" of information simultaneously: what a cell is doing (its transcriptome via scRNA-seq) and where it came from (its lineage via molecular barcoding). However, standard visualization tools like UMAP or PCA are "lineage-blind"—they cluster cells by their current state but often obscure the historical relationships between them. The challenge is that lineage trees are often noisy and incomplete, and not every gene follows the "family rules." We lack robust methods to identify exactly where in a lineage tree expression changes occur and which genes are the faithful carriers of cellular memory across generations.
The PORCELAN Framework
To address these challenges, we introduce PORCELAN (Permutation, Optimization, and Representation learning based single Cell gene Expression and Lineage ANalysis). The framework employs a three-pronged approach to bridge the gap between ancestry and molecular state.
1. Representation Learning
PORCELAN utilizes an overparameterized autoencoder trained with a lineage tree-based triplet loss function. Unlike standard autoencoders, this loss forces the model to organize cells in a latent space that respects lineage distances. By selecting an anchor cell, a positive example (close in the lineage), and a negative example (distant in the lineage), the model ensures that "cousin" cells are represented closer together than "stranger" cells in the learnt representation.
2. Tree-Likeness Optimization and Permutation
The framework generalizes spatial statistics into a multivariate "tree-likeness" score. It jointly optimizes gene weights and tree edge lengths to highlight specific genes and subtrees where expression patterns are most consistent with the lineage.
To quantify "cellular memory," PORCELAN uses a permutation strategy. By shuffling cells within subtrees of varying heights and observing the decay of the tree-likeness score, the method can determine how much memory is preserved at different resolutions—from immediate siblings to distant relatives.
Datasets and Validation
The authors validated PORCELAN across several complex biological contexts:
- Synthetic Data: Generated using TedSim, simulating various cell division modes to provide a ground truth.
- Cancer: A mouse model of lung cancer (3515_Lkb1_T1) using CRISPR/Cas9-based lineage tracing.
- Development: An aggregated atlas of C. elegans embryogenesis and a mouse embryogenesis dataset.
Key Insights and Results
Continuous vs. Abrupt Evolution
A major finding was the contrast between developmental and cancerous systems. In C. elegans and mouse embryos, gene expression evolves continuously, following a "Brownian Motion" model where memory decays predictably. In cancer, however, the researchers found "jumps"—abrupt breaks in expression memory that often correspond to metastatic events.
Identifying Lineage-Informative Genes
Not all genes are informative about lineage. PORCELAN successfully filtered out noise to identify genes like Sftpc in lung cancer, which is downregulated as cells transition from a primary tumor state to a metastatic one. Genes identified as highly "tree-like" were significantly enriched for known developmental pathways (e.g., Rhox5 in mouse endoderm), demonstrating the method's ability to capture real biological signals.
Mapping Metastasis
In the lung cancer dataset, PORCELAN pinpointed subtrees giving rise to metastases and identified genes involved in the Epithelial-Mesenchymal Transition (EMT), a hallmark of cancer progression. By allowing the algorithm to "stretch" or "shrink" the edges of the lineage tree, it automatically identifies the specific divisions where the most significant biological transitions occur.
Discussion and Future Work
PORCELAN provides a robust tool for studying cell state memory through divisions across biological systems. While powerful, it currently relies on the accuracy of the input lineage tree. Future iterations could incorporate imaging data or other "omics" layers (like epigenetics) to provide a more holistic view of cellular memory. As single-cell datasets scale, further algorithmic refinement will be necessary to manage the computational cost of joint optimization.
Summary updated