Cell typing metrics

Announcements

• Exam handed out at end
• policy on dropping a exam grade

BBKNN:

• explain again key methodological step: each cell gets kNN of cells in each batch
• side effect: k = 3 -> each cell has 3 * number of batches nearest neighbors
• expression estimates and PCs are not altered, only the kNN

Harmony

• Harmony is an alternative to BBKNN
• Paper on Pubmed
• Python package

How it works

Iterate: - soft kmeans clustering (does k value matter?) - within a cluster, compute batch specific adjustment to push cells in a batch to overall cluster centroid - PC values of cells updated based on ^ and soft cluster membership

Other things

• Performance compared to BBKNN seems similar
• When batches are different times of a process, this method seems to produce knn graphs -> embeddings that are more continous
  • However, they seem to need to be restricted to cells of a shared lineage

PySingleCellNet

• What happens when a classification result is not definitive?
• Rationale for 'random' class and for catgorization of prediciton results into 'singlular', etc

HW3

• https://compscbio.github.io/cscb2026/hws/hw_3.html

  • Describe the data
  • Gene set enrichment analysis (GSEA)
  • GSEApy
  • Example application with pySCN wrapper to GSEAPY

• Performance metrics and parameter sweeps

  • https://compscbio.github.io/cscb2026/notebooks/classification_metrics.html