# DE analysis with model.SCVI: which lfc indicates gene up-/down-regulation?

**URL:** <https://discourse.scverse.org/t/de-analysis-with-model-scvi-which-lfc-indicates-gene-up-down-regulation/729>\
**Category:** scvi-tools\
**Tags:** diff-exp, scvi\
**Created:** [September 6, 2022, 6:13pm UTC](https://discourse.scverse.org/t/de-analysis-with-model-scvi-which-lfc-indicates-gene-up-down-regulation/729 "2022-09-06T18:13:16Z")\
**Posts on this page:** 3\
**Page:** 1

<div class="post-metadata">

**Author:** ![Sam](https://avatars.discourse-cdn.com/v4/letter/s/5f9b8f/32.png) [@Sam](https://discourse.scverse.org/u/Sam)\
**Post date:** [September 6, 2022, 6:13pm UTC](https://discourse.scverse.org/t/de-analysis-with-model-scvi-which-lfc-indicates-gene-up-down-regulation/729/1 "2022-09-06T18:13:16Z")

</div>

Hello!

Thank you for the wonderful scVI-tools! I am new to scRNA-seq data analysis. Briefly, I used scanpy to concatenate four samples (two replicates from two conditions each, all 10X datasets), performed preprocessing, trained scVI model, computed neighbors with scVI’s latent representation of the data, then computed UMAP and clustered with leiden method. Subsequently, I computed differential expression - a) between the two conditions (global), as well as b) between the cells from the two conditions for each leiden cluster.

I have the following questions regarding scvi.model.SCVI.differential\_expression() output:

1. The output provided the following logFoldChange columns: lfc\_mean, lfc\_median, lfc\_std, lfc\_min, lfc\_max.  
For the genes that are DE (is\_de\_fdr\_0.05=True), how can I tell if the gene expression goes higher or lower? Should I just use “lfc\_mean” for this purpose?

2. Almost always lfc\_min is negative and lfc\_max is positive, whereas the lfc\_mean may be positive or negative. So the gene expression is lower in some cells (of a condition, for example), and higher in others? And the lfc\_mean is providing the direction of change averaged over all the cells? Do I understand it correctly?

3. Sometimes the “raw\_mean1”, “raw\_mean2”, “raw\_normalized\_mean1”, and “raw\_normalized\_mean2” are all zero, yet the genes are DE nonetheless. Is that okay?

You can find a snippet of the output with selected columns below.

Thank you,  
Sam

```
lfc_mean lfc_median lfc_std lfc_min lfc_max raw_mean1 raw_mean2 raw_normalized_mean1 raw_normalized_mean2 is_de_fdr_0.05
gene                                                                                                                                                        
si:dkey-208b23.5 0.479600 0.439014 6.084463 -14.936892 15.690244 0.000000 0.000000 0.000000 0.000000 True
ENSDARG00000024602 0.679054 0.649739 7.324370 -18.278358 17.349012 0.000000 0.000000 0.000000 0.000000 True
drc1 0.540901 0.562465 6.162925 -17.544102 15.701396 0.000000 0.000000 0.000000 0.000000 True
esrp1 0.558089 0.535105 6.800211 -18.745285 16.212980 0.000000 0.000000 0.000000 0.000000 True
gfi1ab 0.561095 0.544334 6.464262 -19.063986 15.834004 0.000000 0.000000 0.000000 0.000000 True
zgc:103559 0.497405 0.481609 5.919106 -16.137840 14.522301 0.000000 0.000000 0.000000 0.000000 True
napsa 0.440470 0.389575 6.322702 -16.612000 14.962067 0.000000 0.000000 0.000000 0.000000 True
myzap 0.483889 0.502490 5.984181 -16.628532 14.994525 0.000000 0.000000 0.000000 0.000000 True
CR354435.1 0.520383 0.471522 6.401921 -14.583641 15.076799 0.000000 0.000000 0.000000 0.000000 True
dnah12 0.454225 0.424026 5.315034 -13.046480 13.362019 0.000000 0.000000 0.000000 0.000000 True
fut9b 0.456071 0.466859 6.132898 -17.295963 15.211704 0.000000 0.000000 0.000000 0.000000 True
si:ch211-113a14.22-1 0.463434 0.436661 5.820310 -13.541477 13.315475 0.000000 0.000000 0.000000 0.000000 True
gpr184 0.514271 0.445153 4.942262 -11.320700 12.297648 0.000000 0.000000 0.000000 0.000000 True
cyp4v8 0.279370 0.257111 5.054646 -13.266130 12.379461 0.000000 0.000000 0.000000 0.000000 True
si:ch1073-228h2.2 0.470916 0.417501 6.080115 -14.715637 14.257307 0.000000 0.000000 0.000000 0.000000 True
rgs13 0.548236 0.522250 6.027261 -18.523365 14.883591 0.000000 0.000000 0.000000 0.000000 True
si:dkeyp-73d8.6 0.471345 0.417540 5.409026 -13.083514 13.284245 0.000000 0.006431 0.000000 0.019727 True
zgc:64051 0.522500 0.487912 5.545477 -15.674596 13.653400 0.000000 0.000000 0.000000 0.000000 True
srd5a2b 0.524634 0.413060 5.227385 -13.321875 12.613957 0.003236 0.003215 0.027218 0.037259 True
si:dkey-192g7.3 0.387664 0.402621 4.858399 -12.267364 11.662878 0.000000 0.006431 0.000000 0.053546 True
tent5ab 0.634892 0.472449 6.047723 -13.824921 13.967886 0.000000 0.000000 0.000000 0.000000 True
gpd1a 0.591813 0.543583 5.842104 -14.230432 14.385897 0.000000 0.000000 0.000000 0.000000 True
CU984600.2 0.505643 0.457640 5.578200 -13.251217 13.337477 0.000000 0.000000 0.000000 0.000000 True
dbh 0.651233 0.666539 4.201104 -11.775505 12.334496 0.223301 0.115756 1.375563 0.451738 True
ano6 0.523729 0.489827 5.218942 -14.128026 13.381668 0.000000 0.000000 0.000000 0.000000 True
zgc:162331 0.432170 0.369040 4.980867 -11.879374 12.364733 0.000000 0.000000 0.000000 0.000000 True
oip5 0.330404 0.330091 4.209458 -10.174562 10.167267 0.000000 0.000000 0.000000 0.000000 True
ctrb1 -0.249836 -0.376842 4.786657 -13.227513 29.091158 0.009709 0.000000 0.669568 0.000000 True
mc3r 0.409198 0.240824 5.219241 -12.307832 12.104459 0.000000 0.006431 0.000000 0.042580 True
ebf3b 0.559580 0.609398 3.631861 -10.932887 9.725893 0.000000 0.000000 0.000000 0.000000 True
tmx3a 0.322164 0.324352 3.622740 -8.779639 8.504380 0.003236 0.000000 0.024187 0.000000 True
si:dkey-101k6.5 0.314722 0.264050 4.238637 -9.610153 9.698996 0.000000 0.000000 0.000000 0.000000 True
c16h2orf66 0.542751 0.535967 3.325066 -8.829274 9.562749 0.000000 0.003215 0.000000 0.021196 True
CABZ01057928.1 0.310531 0.245671 4.498628 -10.107723 9.703278 0.000000 0.006431 0.000000 0.032090 True
ENSDARG00000075540 0.398474 0.312719 4.399720 -10.870052 11.148718 0.000000 0.000000 0.000000 0.000000 True
sh2d4ba 0.240636 0.170192 3.310970 -9.403920 10.003074 0.012945 0.000000 0.090937 0.000000 True
poc1b 0.642688 0.513015 5.234146 -11.732223 12.384102 0.000000 0.003215 0.000000 0.206118 True
si:dkey-57n24.6 0.218827 0.166036 3.520045 -10.438939 9.694427 0.000000 0.003215 0.000000 0.027960 True
tbx2a 0.679166 0.644877 3.437314 -8.924354 9.783237 0.000000 0.000000 0.000000 0.000000 True
colec12 0.290175 0.205803 3.577260 -9.736406 20.785942 0.003236 0.000000 0.022090 0.000000 True
si:ch211-284k5.2 0.199817 0.094964 3.240186 -10.024623 10.127933 0.009709 0.012862 0.081583 0.117200 True
col6a3-1 -0.412441 -0.498374 3.123330 -9.520643 13.368073 0.000000 0.000000 0.000000 0.000000 True
zgc:113054 0.256077 0.233528 3.476520 -10.191195 9.285287 0.003236 0.000000 0.028264 0.000000 True
urp1 -0.115161 -0.199985 4.026823 -12.646875 24.443443 0.339806 0.009646 16.555485 0.373933 True
calml4a 0.233648 0.158408 3.125443 -9.542439 9.866495 0.000000 0.003215 0.000000 0.047707 True
CR848032.2 -0.457329 -0.593560 3.813791 -11.415813 25.767803 0.006472 0.012862 0.120338 0.678105 True
cass4 0.341258 0.263085 3.951136 -9.347710 9.992567 0.000000 0.000000 0.000000 0.000000 True
```

---

<div class="post-metadata">

**Author:** ![valehvpa](https://avatars.discourse-cdn.com/v4/letter/v/8c91f0/32.png) [@valehvpa](https://discourse.scverse.org/u/valehvpa)\
**Post date:** [September 6, 2022, 6:25pm UTC](https://discourse.scverse.org/t/de-analysis-with-model-scvi-which-lfc-indicates-gene-up-down-regulation/729/2 "2022-09-06T18:25:47Z")

</div>

@PierreBoyeau would be the best person to assist!

---

<div class="post-metadata">

**Author:** ![Valentine\_Svensson](https://yyz1.discourse-cdn.com/flex035/user_avatar/discourse.scverse.org/valentine_svensson/32/10_2.png) [@Valentine\_Svensson](https://discourse.scverse.org/u/Valentine_Svensson)\
**Post date:** [September 15, 2022, 3:27pm UTC](https://discourse.scverse.org/t/de-analysis-with-model-scvi-which-lfc-indicates-gene-up-down-regulation/729/3 "2022-09-15T15:27:05Z")

</div>

Hi Sam,

scVI uses a Bayesian framework for analysis. Very briefly, what this means is that based on data, an entity such as fold change has a range of possible values. Abstractly, there is a conditional probability distribution P([\text{Fold change}] \, |\, [\text{Data}]), which scVI estimates through various sampling strategies.

The `lfc_min` and `lfc_max` are the smallest and largest samples of log fold changes that were observed during the sampling process. They reflect samples from the distribution P([\text{Fold change}] \, |\, [\text{Data}]), and not observed cells or data.

The columns `lfc_mean` and `lfc_median` reports different measures of ‘most typical’ values of the distribution P([\text{Fold change}] \, |\, [\text{Data}]). Unfortunately it is somewhat ambiguous which notion of ‘average’ is most appropriate.

Personally, I prefer median over mean, but I don’t have a clear motivation for why.

Regarding direction, the columns `comparison`, `group1`, and `group2` indicates what is being compared to what. The fold changes are always in the form of `group1` / `group2`. So a positive log fold change means that expression is higher in `group1` than `group2`, and a negative fold change means that expression is higher in `group2` than in `group1`.

Hope this is helpful!  
/Valentine