Transcriptional immune suppression and upregulation of double stranded DNA damage and repair repertoires in ecDNA-containing tumors

Reviewer #1 (Public Review):

Recently discovered extrachromosomal DNA (ecDNA) provides an alternative non-chromosomal means for oncogene amplification and a potent substrate for selective evolution of tumors. The current work aims to identify key genes whose expression distinguishes ecDNA+ and ecDNA- tumors and the associated processes to shed light on the biological mechanisms underlying ecDNA genesis and their oncogenic effects. While this is clearly an important question, the analysis and the evidence supporting the claims are weak. The specific machine learning approach seems unnecessarily convoluted, insufficiently justified and explained, and the language used by the authors conflates correlation with causality. This work points to specific GO processes associated (up and down) with ecDNA+ tumors, many of which are expected but some seem intriguing, such as association with DSB pathways. My specific comments are listed below.

A. The claim of identifying genes required to 'maintain' ecDNA+ status is not justified - predictive features are not necessarily causal.
B. The methods and procedures to identify the key genes is hyperparameterized and convoluted and casts doubt on the robustness of the findings given the size and heterogeneity of the data.
a. In the first two paragraphs of Boruta Analysis Methods section, authors describe an iterative procedure where in each iteration, a binomial p-value is computed for each gene based on number of iterations thus far in which the gene was selected (higher GINI index than max of shadow features). But then in the third paragraph they simply perform Random Forest in 200 random 80% of samples and pick a gene if it is selected in at least 10/200. It is ultimately not clear what was done. Why 10/200? Also "the probability that a gene is a "hit" or "non-hit" in each iteration is 0.5" is unclear. That probability is of a gene achieving GINI index higher than the max of shadow features. How can it be 0.5?
b. The approach of combining genes with clusters is arbitrary. Why not start with clusters and evaluate each cluster (using some gene set summary score) for their ability to discriminate? Ultimately, one needs additional information to disambiguate correlated genes (i.e. in a co-expression cluster) in terms of causality.
c. The cross-validation procedure is not clear at all. There is a mention of 80-20 split but exactly how/if the evaluation is done on the 20% is muddled. The way precision-recall procedure is also a bit convoluted - why not simply use the area under the PR curve?
d. The claim is that Boruta genes are different from differentially expressed genes but the differential expression seems to be estimated without regards to cancer type, which would certainly be highly biased and misleading. Why not do a simple regression of gene expression by ecDNA status, cancer type and select the genes that show significant coefficient for ecDNA status?
C. After identifying key features (which the authors inappropriate imply to be causal) they perform a series of enrichment/correlative analysis.
a. It is known that ecDNA status associates with poor survival, and so are cell cycle related signal. Then the association between Boruta genes and those processes is entirely expected. Is it not? The same goes for downregulation of immune processes.
b. The association with DSB specifically is interesting. Further analysis or discussion of why this should be would strengthen the work.
c. On page 15, second paragraph, when providing the up versus down CorEx genes, please also provide up versus down for non-CorEx genes as well to get a sense of magnitude.
d. The finding that Boruta genes are associated with high mutation burden is intriguing because in general mutation burden is associated with better survival and immunotherapy response. This counter-intuitive result should be scrutinized more to strengthen the work.
e. On page 17 "12 of the 47 genes not specifically enriching any known GO biological Process" is confusing. How can individual gene enrich for a GO process?

Reviewer #2 (Public Review):

In their manuscript entitled "Transcriptional immune suppression and upregulation of double stranded DNA damage and repair repertoires in ecDNA-containing tumors" Lin et al. describe an important study on the transcriptional programs associated with the presence of extrachromosomal DNA in a cohort of 870 cancers of different origin. The authors find that compared to cancers lacking such amplifications, ecDNA+ cancers express higher levels of DNA damage repair-associated genes, but lower levels of immune-related gene programs.

This work is very timely and its findings have the potential to be very impactful, as the transcriptional context differences between ecDNA+ and ecDNA- cancers are currently largely unknown. The observation that immune programs are downregulated in ecDNA+ cancers may initiate new preclinical and translational studies that impact the way ecDNA+ cancers are treated in the future. Thus, this study has important theoretical implications that have the potential to substantially advance our understanding of ecDNA+ cancers.

Strengths
The authors provide compelling evidence for their conclusions based on large patient datasets. The methods they used and analyses are rigorous.

Weaknesses
The biological interpretation of the data remains observational. The direct implication of these genes in ecDNA(+) tumors is not tested experimentally.

Reviewer #3 (Public Review):

Summary:
Using a combination of approaches, including automated feature selection and hierarchical clustering, the author identified a set of genes persistently associated with extrachromosomal DNA (ecDNA) presence across cancer types. The authors further validated the gene set identified using gene ontology enrichment analysis and identified that upregulated genes in extrachromosomal DNA-containing tumors are enriched in biological processes like DNA damage and cell proliferation, whereas downregulated genes are enriched in immune response processes.

Major comments:
1. The authors presented a solid comparative analysis of ecDNA-containing and ecDNA-free tumors. An established automated feature selection approach, Boruta, was used to select differentially expressed genes (DEG) in ecDNA(+) and ecDNA(-) TCGA tumor samples, and the iterative selection process and two-tier multiple hypothesis testing ensured the selection of reliable DEGs. The author showed that the DEG selected using Boruta has stronger predictive power than genes with top log-fold changes.

2. The author performed a thorough interpretation of the findings with GO enrichment analysis of biological processes enriched in the identified DEG set, and presented interesting findings, including the enrichment in DNA damage process among the genes upregulated in ecDNA(+) tumors.

3. Overall, the authors achieved their aims with solid data mining and analysis approaches applied to public data tumor data sets.

4. While it may not be the scope of this study, it will be interesting to at least have some justification for choosing Boruta over other feature selection methods, such as Recursive Feature Elimination (RFE) and backward stepwise selection.

5. The authors showed that DESEQ-selected DEGs with top log-fold changes have less strong predictive power and speculated that this may be due to the fact that genes with top log-fold changes (LFC) are confined only to a small subset of samples. It will be interesting to select DEGs with top log-fold changes after first partitioning the tumor samples. For example, randomly partition the tumor samples, identify the DEGs with top LFC, combine the DEGs identified from each partition, then evaluate the predictive power of these DEGs against the Boruta-selected DEGs.

6. While the authors showed that the presence of mutations was not able to classify ecDNA(+) and (-) tumor samples, it will be interesting to see if variant allele frequencies of the genes containing these mutations have predictive power.