Thread by @VanAllenLab

A brief thread on our latest study, as well as a more general chat on biology & biomarkers for cancer immunogenomics - w/ longtime friend Dirk Schadendorf + rockstar jr faculty @dliu_ccb + more friends! @NatureMedicine @DanaFarber @broadinstitute
nature.com/articles/s4159… [1/n]

Many current biology & biomarkers blurry lines in cancer immunogenomics - e.g. our TMB melanoma work:

2014-15: Correlation btwn TMB + clinical benefit to ICB in melanoma → neoantigen biology

Though our 2015 data (ncbi.nlm.nih.gov/pubmed/26359337) suggested weak predictive effect...

...and then:

2018: TMB confounded by mutational signatures (ncbi.nlm.nih.gov/pubmed/30150660)

Here: TMB also confounded by histologic subtype

Cool biology: Yes!
Clinical biomarker: Err...probably not

(Relatedly, the rush to make potentially interesting biology into clinical biomarkers in this space is larger challenge - see also our PBRM1 work in RCC and my all caps twitter disclaimers:

https://twitter.com/VanAllenLab/status/1169665654649024512?s=20

https://twitter.com/VanAllenLab/status/1135618111087009794?s=20

https://twitter.com/VanAllenLab/status/949383664630206464?s=20

)

So, with clinical data + exomes + transcriptomes from 144 melanoma patients treated with PD1 immune checkpoint blockade, we sought to explore for new biology and attempt predictive biomarkers going beyond TMB

(S/o @Lakers for that beautiful science purple/gold @KevinOConnorNBA)

Re: biology - as we predicted in our 2018 Miao et al @NatureGenet study (ncbi.nlm.nih.gov/pubmed/30150660), this seemingly large cohort for this type of study is still underpowered for discovery, and we’re not ashamed to admit it

There are many potentially interesting findings that may point to subsequent biological evaluation - global genomic features, antigen presentation mutations, candidate novel genes, etc...

...but once again, these potential biological findings are NOT CLINICAL BIOMARKERS

Rather...to build on biology and systematically explore predictive biomarkers, we leveraged this multi-modal integrative data set (that had relevant clinical data and lots of molecular information) to examine features that may inform integrated predictive modeling

Word of caution to predictive modeling folks - in this study, simply the timing of biopsy w/r/t CTLA4 Ab exposure influenced the transcriptomes (biology!)

Had we proceeded w/o this key but often hard to discern clinical tidbit, we might have been misled

tl;dr this is hard

In building these models, it turned out some of the most fruitful ones leveraged very basic clinical data + molecular features…nothing fancy w/r/t features or computational approaches worked reasonably well

(Also, TMB was not that useful for these predictive models…)

We were especially interested in models to find patients likely to be intrinsically resistant to immunotherapy - might be useful (one day, with validation) for helping clinicians choose between IO & targeted therapies

We were also interested in exploring how single cell signatures may enable better predictive models, working with Dr. Aviv Regev + friends (ncbi.nlm.nih.gov/pubmed/30388455)

It's early, more to come!

Unfortunately, we could not find available cohorts to properly validate the full models we devised...

...but we hope this study provides a framework & momentum for the community to examine these features + others in an integrative fashion for biology & biomarkers collaboratively

Lastly, about @dliu_ccb: In our 1st mtg, I advised he join another lab bc I'd probably fail - thankfully he ignored me & totally crushed it

This was Dave's last study in my lab as he's starts his melanoma computational oncology lab @DanaFarber - a genuine privilege Dave! [fin]