De Novo Prediction of Stem Cell Identity using Single-Cell Transcriptome Data

Identifying stem cells in tissues is critical for understanding

homeostasis, disease, regenerative medicine, and cancer

Stem cells are rare, making discovery by traditional population-based

assays very difficult

Single-cell mRNA-seq can profile tissues at single-cell resolution

but distinguishing cell types from noisy data is hard

Existing trajectory inference methods (e.g., Monocle) assume continuous temporal change and struggle

with correct tree topology in multi-lineage systems

RaceID2

improved cell type clustering algorithm

StemID

algorithm to infer lineage trees and identify stem/multipotent cells from scRNA-seq snapshots

Clustering method

k-means and k-medoids

cluster number selection

Gap statistic and Saturation of within-cluster dispersion

k-medoids improved

robustness substantially

Guided topology

all pairwise cluster links are considered as possible differentiation trajectories (no assumption about branching)

Each cell is assigned to the link with the longest projection of the vector

from its cluster medoid to the cell, onto inter-cluster links

Projection coordinate

pseudo-temporal position of the cell on that link

Links with significantly more cells than expected by chance

(vs. randomized background) are retained (p < 0.01)

Link score = 1 − (maximum gap between neighboring cells on the link)

close to 1 = dense, continuous coverage → strong differentiation trajectory evidence

Outperformed Monocle

in resolving multiple secretory branches

The StemID score combines two features:

StemID score = (median entropy − min entropy across all cell types) × number of significant links

Number of Links (Connectivity)

Low-scoring links indicate transcriptome plasticity / fate bias fluctuations

Stem/multipotent cells connect to many clusters

many links

Differentiated cells are more restricted

fewer links

Transcriptome Entropy (Shannon entropy)

Stem cells have a more uniform transcriptome — no single gene dominates

Differentiated cells (e.g., Paneth cells) express a few cell-type-specific genes at very high levels

lower entropy

Entropy had previously been applied to

study cellular differentiation

Entropy alone doesn't discriminate perfectly;

rescaling connectivity by entropy improves discrimination

Lgr5+ crypt base columnar cells (CBCs) → transit-amplifying (TA) cells

enterocytes (absorptive) or secretory cells (goblet, Paneth, enteroendocrine)

Used lineage tracing (Lgr5-CreERT2 × Rosa26-YFP)

cells collected 5 days post-label induction

Clusters

Cluster 2: Lgr5+ stem cells (also Ascl2+, Clca4+)

Clusters 1, 8: early progeny

Cluster 3: enterocytes

Clusters 4, 19: goblet cells

Clusters 5, 6: Paneth cells

Cluster 7: enteroendocrine cells

Secretory lineages branch directly

from stem cell cluster

Two Paneth differentiation routes recovered:

Via Dll1+ common precursor (cluster 1) — canonical

Direct from stem/TA cells — rare, previously described only after Paneth cell ablation

StemID score

highest for cluster 2 (Lgr5+ CBCs) ✓

Cluster 7 (Lgr5+/Clca4+)

highest StemID score

Entropy rescaling correctly distinguished mature Paneth/goblet cells

(same connectivity as stem cells but lower entropy)

StemID findings:

Highest score: cluster 1 (HSCs)

Score level correlates with degree

of multipotency

HSCs have highest, narrowest entropy distribution

uniform transcriptome not dependent on cell cycle

54/276 HSCs (20%) show lineage-specific fate biases

(likely an underestimate due to sequencing sensitivity limits)

112/276 HSCs project onto the link toward

multipotent progenitor (cluster 5)

Pseudo-temporal analysis of neutrophil branch (clusters 1→11→3→2→12)

gene expression changes smoothly, 5 co-regulated modules identified by self-organizing maps (SOMs)

StemID robust to subsampling

even with only 10 HSCs, cluster 1 still scores highest

Clusters 4 and 14 acquire the two highest StemID scores

predicted as most multipotent

Pseudo-temporal expression of FTH1 and INS (insulin)

smooth, anti-correlated gradients along the link connecting clusters 4→8→6

Suggests a continuous transition from ductal precursor

to mature β cell

Found individual cells co-expressing insulin and FTL

within ductal structures

Multipotency definition: purely functional

the cell population with the highest degree of multipotency is the stem cell candidate

Transcriptome plasticity (low-score links) ≠ differentiation trajectory

but reveals fate biases and promiscuous transcriptome states

Classical dichotomous differentiation hierarchy challenged

early fate bias already present at HSC stage

Entropy as a proxy for

Waddington landscape energy