HILCPS Lab · IIT Madras · Dec 2023 – May 2025
Inductive graph-transformer frameworks for scRNA-seq that recover cell trajectories in gliomas, with position and structure encodings and multiscale embeddings that preserve lineage and cell-to-cell communication. The visualisation below follows one sampled cell neighbourhood from graph to lineage.
01 The scaffold
Every cell is a node in a kNN graph built over expression space. In the middle sits one attention head of the graph transformer: a query cell surrounded by its sampled neighbourhood, each neighbour carrying a positional or structural encoding. On the right, an empty UMAP panel waits with its pseudotime scale.
02 Inference
An inductive sampler rings the query cell, softmax attention weighs each neighbour, and the learned embeddings scatter onto the UMAP manifold ordered by pseudotime. A principal curve threads the cells and forks at a branch point, separating the neoplastic lineage from the non-neoplastic one.
The pipeline drawn above: a kNN cell graph over scRNA-seq profiles, inductive neighbourhood sampling around a query cell, and a graph-transformer attention head with softmax-weighted edges, α = softmax(qᵀk/√d), augmented with Laplacian positional encodings and structural encodings. The learned multiscale embeddings project to a UMAP manifold, and a branching principal curve recovers pseudotime, splitting a neoplastic from a non-neoplastic lineage.
The first scroll lays out the scaffold. The second animates the inference: the sampled subgraph, the attention weights, and the trajectory threading the manifold, which is what preserving lineage structure across sparse atlases looks like in practice.