Forest Fire Clustering: Iterative Label Propagation Clustering and Monte Carlo Validation For Single-cell Sequencing Analysis

With the rise of single-cell sequencing technologies, there is a growing need for robust clustering algorithms to extract deeper insights from data. Here, we introduce an intuitive and efficient clustering method, Forest Fire Clustering, for discovering and validating cell types in single-cell sequencing analysis. Compared to existing methods, our clustering algorithm makes minimum prior assumptions about the data distribution and can provide a point-wise significance value via Monte Carlo simulations for internal validation. Additionally, point-wise label entropies can highlight novel transition cell types \emph{de novo} along developmental pseudo-time manifolds. Lastly, our inductive algorithm has the ability to make robust inferences in an online-learning context. In this paper, we describe the method, provide a summary of its performance against common clustering benchmarks, and demonstrate that Forest Fire Clustering is uniquely suitable for single-cell sequencing analysis.