Hi,
I am working on a project with 20 samples. 5 cancer patients. 4 repeated measures (blood) from each during treatment.
One of the samples has a very diverging expression profile. I advocate for the removal of this sample from the analysis, while other people in the project want to keep it.
I've been looking into its expression, and there are a 85 genes with normalized expression > 100 counts on the weird sample, while having almost 0 expression in the other 19 samples (usually, actual 0 reads on all, or 0 on most and up to ~10 reads in a single other sample).
Similarly, ~800 genes in this sample have >20 counts with close to 0 in the other samples. I've checked how many "sample-exclusive" genes each sample has, and the next one in list has 11 "exclusive" genes.
On a cursory look at the 85 genes, I cannot find much related to immune function. I suspect that this sample might have shed some tumor cells mixed in with the immune cells, and those were kept in the sample despite the thorough processing. A quick and dirty look at enrichR found lots of neuronal, lung, and gastrointestinal signals in the TISSUES_Curated_2025 and TISSUES_Experimental_2025 datasets.
Do you have any suggestions on what to test for the source of "contamination" inf this sample?
Thanks
PS: I am 100% convinced we have to remove this outlier, for the sanity of DEG analysis. I need to convince other people.
•
written
by
txema.heredia •
1 answer
As you add more and more samples to a project, the chance of something going wrong during sample prep, etc., increases radically. Or let me rephrase that, the chance that nothing goes wrong decreases radically.
In this case it is 1 in 20 samples. Not too bad IMHO.
I have seen many bad and unexplainable samples in my life, and so often I tried to figure them out, track things down, do secondary analyses etc apply various data salvaging operations ... I can't recall being successful
Nowadays, I am fully comfortable looking at a heatmap like yours and say: "For reasons I was unable to identify, the data for one particular sample looked way off and inconsistent with all other data. For that reason, I have removed it from further analysis."
And that's it.
That heatmap you have shows that the data is not similar to any other data you have - so that tells me the problem with it something deeper and more universal.
Log in to answer this question.
-
Ambient RNA correction when most of the ambient comes from a specific cell type of interest
written by txema.heredia •
-
Seurat - integration of metastases biopsies
written by txema.heredia •
-
Finding differences between individual patients in a time-series RNA-seq?
written by txema.heredia •
-
Tool for automatic immune cell annotation in MOUSE?
written by txema.heredia •
-
Could you please assist in identifying this cluster?
written by carolofharvest •
-
Inferring cell identity/genotype in single cells with missing information
written by txema.heredia •
-
Single cell demultiplexing using 10x and TotalSeq-B
written by txema.heredia •
-
Any reason for scRNA data of mouse lung samples to fail to capture epithelial cells?
written by 윤여가(대학원 바이오정보학) •
-
deconvolution of heterogenous cell types in RRBS data
written by rahel14350 •
-
Hemoglobin genes highly expressed in one replicate sample (RBC contamination) - what can I do?
written by terkild •
Did you check the pair-wise correlation and PCA of all the samples? If yes, then share the plots.
Also, check the alignment statistics and depth of coverage.
The pair-wise correlation looks indeed very weird and it is what makes me want to remove the sample. All 4 samples from the same patient are clustered together, except for this specific sample, which is an outlier of everything else.
Just by looking at this matrix, one could think that this blood sample was mislabeled and belongs to another patient not among the 5 sequenced here. I checked the expression of sex-specific genes and the sample's expression match the other samples from patient3, though.
The PCA looks more "reasonable". All patients cluster together. This sample appears more distant in the PC2+PC3 plot. PC1 and PC2 are both largely driven by HLA-xxx and IGKVxxx genes. Expected, as these samples are PBMCs.
Despite looking fine in this PCA, these samples where sequenced in 2 batches (p1-3, p4-5). In the original PCA using only the first batch, this sample was also a clear outlier.
I don't have access to the alignment statistics. It was done by a collaborator, I will ask them about it. Had I access to the fastq files, the first thing I would have checked would be genotyping of the sample to make sure it matches the other samples of the patient.
In the raw counts matrix, this sample had the 2nd lowest number of counts, but still nothing out of the usual.
A pathologist has suggested that this patient/sample might be having a unique immune response, and that would be interesting to look into. I agree with that, but I disagree with including it on the DGE analysis, as it will skew everything.
Also, checking the list of 85~800 sample-specific genes, I didn't see an abundance of immune-related genes, but more neuron-associated pathways.
"Interesting to look at"? But you really can't investigate one time point in one sample. It's just not practical. And how interesting is it if the later time points look like you expect?
In my reading, it is interesting only from the point of view of someone who neither fully understands the processes, nor has to expend effort to do the work ...
I second the other comment - you need to investigate further to identify why it might be an outlier. This both helps you confirm it should be removed, as well as could indicate where the error comes from for future work. Are there any technical replicates of each measure?
You can look at things like (list not exhaustive):
I've just replied the other comment.
We no replicates. Just 1 sample per patient x time. I don't have access to the raw data right now, as alignment was done by collaborators.
What I highly suspect is that this specific sample either got some type of degradation before sequencing, or it is not composed purely by PBMCs but also by circulating tumor cells/debris post-treatment.
That, or something else entirely different. Which makes me not trust the sample for DGE analysis.
Have you tried doing the DEG analysis leaving the sample as is and also by removing it. That should be an easy enough thing to do. Do you see a drastic difference? If not you could leave the sample in, if the customer insists on it. In principle, if the sample is an obvious outlier, then it should be removed.
Yes, I already did that. Including the "controversial" sample (compared to having excluded it) causes the removal of 866 DEG (717 unique) and the detection of 547 new DEG (509 unique) in the pairwise time vs time comparisons (~550 both removed and newly added when looking only at comparisons involving t3).
And when looking at the pairwise patient vs patient comparisons (DESeq2's design model is ~timepoint + patient), there are ~5,400 unique DEG removed and ~2,400 newly added.
That's why I'm looking for some argument to give strength to "this is not just differences in immune response".
I think part of your problem here is that this experiment is underpowered. If this was 5 inbred mice, 4 time points, it would be doable, but you've got 5 human patients. You need at least twice that number to really do this right.
I completely agree with you, but that is what happens with pilot studies.
I'm mostly just wondering if there is any way I can check if something went wrong with the sequencing itself, or if these weird patterns are "real" and can affect other analyses (e.g. qPCR) done on the same blood sample.
The (statistically not sound) EnrichR results matching neuron-development processes and neuronal+gastrointestinal tissues match the tumor these patients have. However, these blood samples should have had a thorough washout, getting rid of (most) circulating cell-free tumor RNA. Maybe the treatment triggered a massive "migration" of Circulating Tumor Cells in this specific patient and these are being sequenced together with the PBMCs.
If this sample was in the same pool as others then that is highly unlikely.
Is the treatment strategy dependent on results of this experiment? If it is then perhaps taking another sample would be warranted to confirm this observation.
As @Istvan said in his answer, at times there is no logical explanation for every experimental observation/sample.