Supplementary material
Supplementary Figures
Figure S1:
Perturbation responses and predicted cell state distributions by D-SPIN; net-
work inference of D-SPIN is further improved with double perturbations; related to Fig.
2
(A) On the single-gene knockdown and activation perturbations, the response vectors by D-SPIN
agree with the actual perturbation except for the Gata2 knockdown. This is potentially due to Gata2
being a transient regulator in the network and does not express in the differentiated states. (B) D-
SPIN achieves further improved accuracy with 96 more double gene perturbations. With extra double
perturbations, the top 10, 15, and 20 accuracies are 1, 0.9, and 0.805. (C) The D-SPIN model predicts
the cell-fate distribution generated through perturbations to the underlying network.
1
.
CC-BY-NC 4.0 International license
available under a
was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprint (which
this version posted May 20, 2023.
;
https://doi.org/10.1101/2023.04.19.537364
doi:
bioRxiv preprint
Figure S2:
Exams of gene program discretization; distribution of responding program
number of perturbations; D-SPIN reconstruction of cell cycle states; Coherence of macro-
molecular complex; related to Fig. 3
(A) Heatmap of gene expression and discretized gene pro-
gram level on two examples: control samples of non-targeting guide RNAs and knockdown of ribosome
large or small subunits. The expression of each gene is normalized by its maximum expression and
capped at 0.8 for visualization. (B) The number of responding gene programs of each perturbation
shows an exponential distribution. Perturbations that influence a large number of perturbations are
relatively rare. (C) D-SPIN reconstructs the distribution of cell cycle progression with a 6-node sub-
network taken from the inferred network. The response vector of the subnetwork is inferred by the
mean gene program expression on control samples conditioned on the network. Samples from the
control population and D-SPIN model samples are embedded together on UMAP, and we annotate
and order the cell cycle state of each Leiden cluster by the gene program expression. (D) The cell cycle
state distribution of D-SPIN is similar to data distribution with an 11% mean error. (E) Regulatory
coherence of macromolecular complex subunits of the 46 filtered molecular complexes in the CORUM
database.
2
.
CC-BY-NC 4.0 International license
available under a
was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprint (which
this version posted May 20, 2023.
;
https://doi.org/10.1101/2023.04.19.537364
doi:
bioRxiv preprint
Figure S3:
Timecourse of T-cell mediated immune activation; cell typing of the drug pro-
filing experiments; related to Fig. 5
(A) CD3/CD28 antibody does not activate monocytes. Im-
munostaining of primary human monocytes cultured in CD3/CD28 antibody (25uL/mL cells), IFNG
(5ng/mL), GMCSF (5ng/mL), MCSF (5ng/mL), or TNFA (5ng/mL) reveals that GBP1 is only ex-
pressed in IFNG-activated macrophages. (B) UMAP rendering of the timecourse of T-cell mediated
immune activation for a series of example time points. The immune population gradually moves from
the resting state to the activated state. (C) The dynamics of signaling gene expression in each cell
type show active communication between different cell types. Expression is normalized by the mean
across the timecourse. (D) The dynamics of gene markers for immune activation in each major cell
type. The color-shaded range is 10 and 90 percentile of gene expression. (E) The Z-score of cell typing
gene marker expression. Gene markers are identified by differential expression analysis of each Leiden
cluster. The Z-score is computed in reference to all the cells in the same major cell type.
3
.
CC-BY-NC 4.0 International license
available under a
was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprint (which
this version posted May 20, 2023.
;
https://doi.org/10.1101/2023.04.19.537364
doi:
bioRxiv preprint
Figure S4:
Examples of gene program discretization and gene program expression on the
UMAP; related to Fig. 5
(A) Heatmap of gene expression and discretized gene program level
on three examples: activated control population, resting control population, and antibody activation
with Halcinonide treatment. The expression of each gene is normalized by its maximum expression
and capped at 0.8 for visualization. (B) The expression level of each gene program is rendered on the
UMAP embedding. Each cell is colored by its expression level on the gene program. The expression
level is normalized by the 95 percentile for visualization only.
4
.
CC-BY-NC 4.0 International license
available under a
was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprint (which
this version posted May 20, 2023.
;
https://doi.org/10.1101/2023.04.19.537364
doi:
bioRxiv preprint
Figure S5:
D-SPIN identifies phenotypical classes of drug categories; related to Fig. 6
(A) Heatmap of relative response vector inferred by D-SPIN. D-SPIN identifies 71 effective drugs on
the immune population and groups the drugs into 7 phenotypical classes. (B) UMAP embedding of
example drugs from each class. Drug categories exhibit different signatures of immune cell population
change.
5
.
CC-BY-NC 4.0 International license
available under a
was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprint (which
this version posted May 20, 2023.
;
https://doi.org/10.1101/2023.04.19.537364
doi:
bioRxiv preprint