I saw a billboard for a Very Well Known cancer drug on US101 today and realized that ASCO’s Gastrointestinal Cancers Symposium (aka ASCO GI) was in town. Although I’m not currently working on CRC screening, it’s an interesting enough area for me to have spent a couple years working on it and I’m always curious to see what’s new in the field. CRC is also a pretty interesting tumor from a genetics and evolution perspective, as it has known predisposing variants in the germline (largely arising from DNA damage repair defects) as well as subclasses with (somatically) higher and lower mutation rates, not all of which are down to the germline factors. Those high mutation rates make it an interesting model system for studying tumor phylogeny and evolution.
Although I’m not attending the conference, I had a look at the abstracts in the program guide and thought I’d call out a handful that I found interesting for one reason or another. Note that no endorsement or vilification is implied by a mention, not least because I haven’t actually seen the posters/talks. (Though if any of the authors or others would like to share slides/pictures/poster PDFs, I’d love to see them. Get at me on Twitter.
Disclaimer: As with everything else on this blog, opinions expressed here are my own, and not those of any past, present, future, or subjunctive employers.
CellMax made a splash last year at ASCO GI 2018 with a platform talk (abstract, slides) claiming extremely high performance at detection of colorectal cancer and precancerous lesions (adenomas and carcinomas in situ), using their microfluidic CTC detection platform.
As I discuss in a separate blog post diving into their 2018 presentation, I had a lot of questions regarding their reported performance. In brief, their statistical treatment of age as a confounder, the lack of any disclosed train-test separation, and what seemed like a risky study design all led me to mentally put a big asterisk by their performance numbers. So, it’s very interesting to see new data from the platform.
CellMax’s 2019 abstract refocuses the discussion from cancer detection to cancer prevention. Specifically, the introduction highlights the low sensitivity of existing tests for adenomas (indirectly calling out Exact Sciences’ Cologuard, the “stool-based multi-analyte test”). The new data set has 627 subjects, of whom almost two-thirds (N=405) had an adenoma or a colorectal cancer. (I assume these are independent samples from those presented last year, but the abstract does not say.) The main results presented is that CTC counts were higher in pre-cancer and cancer samples (14.4 +/- 2.6 CTCs) than in healthies (1.7 +/- 0.17), and that CTC counts generally (though not always) increased with cancer stage. Based on their reported sensitivity and specificity numbers, we can compute that they were able to call 362 out of 405 precancer+cancer samples correctly, and 199 out of 222 healthy samples correctly.
I’d love to see more detail on this work, but based on the text of the abstract alone, my three concerns from last year continue to hold. Although this abstract isn’t talking about an “age-adjusted model”, my concerns about whether CTC count are a proxy for age in the sample set remain, and the concerns about sample enrollment protocol also remain given the proportion of cases vs controls in the data set. (Another interesting point: CellMax, both here and in last year’s presentation, has lumped both advanced and non-advanced adenomas together into the positive class. The Cologuard validation study (Imperiale 2014) only considered “advanced precancerous lesions” to be positive findings and put non-advanced adenomas in the negative class. If CTC count is similar between advanced and non-advanced adenomas, this analytical choice may inflate specificity compared to existing tests.)
This abstract describes the next generation of a method previously described last year, looking for RNAs derived from enterocytes shed in the stool. This year, they’ve moved from array-based RNA quantification to sequencing. The goal here is likely to build a “better Cologuard”, with the most obvious angle being to improve detection of advanced adenomas.
The method is interesting: looking at expression changes may be expected to be a more sensitive approach than looking for DNA mutations or epigenetic changes (especially if transcripts are abundant); certainly other companies (including my former employer) have looked for RNAs in the blood.
The findings look interesting but somewhat preliminary. As with CellMax’s data, I’m a little disappointed in the presentation of the numbers. Note that deciding which histologies should be considered positive or negative findings is quite complicated in CRC screening because of the large number of classifications along the adenoma-carcinoma progression and the inconsistent nomenclature between publications. (It’s especially hard for someone like me who is neither a gastroenterologist nor a pathologist. Strong clinical colleagues are priceless.)
While the abstract text states “88.5% specificity for no neoplastic findings”, this category specifically excludes benign polyps and precancerous adenomas, which would be considered negative findings in real applications. To their credit, Geneoscopy provide all the underlying counts (yay, complete data), which lets us recompute the specificity number. If benign polyps are added, specificity falls to 79%; if “precancerous adenomas” are also added to the negative set, specificity would be 71%.
Sensitivity for advanced adenomas looked good at 71%, but 1) that’s 5 out of 7, which is very small sample counts and 2) given the previous question about precancerous adenomas, it’s not clear to me whether the definition of AA used here is the same as that in previous studies. (For reference, Cologuard’s performance for “advanced precancerous lesions” was 42.4% sensitivity in Imperiale 2014, which is about the same as the 42% (16+5 / (16+5+27+2)) achieved here for precancerous + advanced adenomas.)
Cancer performance was perfect, but with only N=4 and no disclosure of stage distribution, it’s not clear how meaningful this number is. (Another detail: Geneoscopy includes stage 0 in its cancer class, whereas CellMax includes those in the pre-cancer class. Again, definitions in CRC screening are tricky.)
It sounds like this abstract may be under simultaneous review for a paper, as the conclusion states that “supplemental data will include expression from 408 seRNA transcripts”; it will certainly be interesting for many to dive into this data.
Approaches for detecting microsatellite instability (MSI) status from plasma rather than tissue aren’t new (see, e.g., this abstract from last year); detection of tumor mutational burden (TMB), a related phenomenon, in the blood was a big topic at ASCO 2018. What I appreciated about this abstract was the up-front presentation and discussion of the importance of validating performance as a function of tumor fraction. Tumor fraction is probably the single most important parameter for any blood-based biomarker (single marker or panel) seeking to measure a signal coming from the tumor itself (as my team showed last year, high tumor fraction is a confounder for multiple analytes). I found it refreshing that an abstract at a clinical meeting discussed performance at relevant tumor fraction cutoffs of 1% and 0.4%, both in simulation and with real data. Kudos.
Immunotherapy remains the hot topic in clinical treatment of cancer (as evidenced by the giant billboard on 101 that prompted this post). The biggest determinant yet identified of probability of response is the density of mutations in a tumor (tumor mutational burden), with the current belief being that more mutations = more neoantigen production = more targets to be recognized by the immune system. Mismatch repair deficient colorectal cancers generally have more mutations (because they can’t repair DNA efficiently), but MMR-proficient cancers are a big and interesting clinical target. Thus, researchers are very interested in whether other drug combinations may work well with IO agents to further boost response.
This study pairs pembro (Keytruda) with the epigenetic modifying agents 5-azacitidine (a DNA methylation inhibitor) and romidepsin (a histone deacetylase inhibitor), with the stated intuition that “HDAC/DNMT inhibition induces susceptibility to immune checkpoints and inhibit tumor growth by re-expressing silenced tumor suppressors”.
Clinical response results are not disclosed in the abstract (the abstract seems to focus on safety and adverse events, concluding that these particular combinations are “generally tolerable”), but I’m very interested to see what happens here. Daniel De Carvalho’s group at the Princess Margaret Cancer Centre has shown that the anti-cancer properties of epigenetic-modifying agents may be due to a mechanism more interesting than “re-expressing silenced tumor suppressors” sounds on the surface. In particular, in a 2015 Cell paper, they showed that 5-azacitidine treatment de-represses transcription of endogenous retroviruses (a really interesting genomic element that I’ll probably blog about in the future); this leads to the presence of intracellular double-stranded RNAs (dsRNAs), which are picked up by innate immune receptors like RIG1 and MDA5. (See also their 2016 review in OncoImmunology). Basically, these agents are immunotherapies in disguise, but bringing the innate immune system into play, not just the adaptive immune system targeting neoantigens. (I wonder if any of these viral RNAs get translated and then cause an adaptive response too.) So even though it’s not described as such in the abstract, this work is describing a dual immunotherapy approach, and it will be interesting to see how it pans out. (Also, the biology of endogenous retroviruses is just that cool and I can’t turn down an opportunity to highlight it.)
Loads of work on tumor evolution has been done on tumor tissue, often by looking at a diverse population of tumors from different individuals or different metastases within one individual, and sometimes over time from one person. It’s challenging to do the latter because biopsies are invasive. However, as liquid biopsies have moved into the clinic, it’s become easier to get multiple samples from the same patient, which allows finer temporal resolution to look at tumor evolution. This abstract looks at 116 patients with metastatic CRC who had between 3 and 12 samples drawn over about 2.5 years.
The interesting detail I saw was at the end of the results, which stated that “clonal hematopoiesis alterations that may be induced by chemotherapy, such as JAK2:V617F, were neither gained nor lost”. Clonal hematopoiesis, the accumulation of somatic mutations in hematopoietic lineages (that come to dominate the genotypes of both WBCs and non-tumor cfDNA with age) is a hot topic in liquid biopsy (GRAIL gave a talk at ASCO last year about CH) and in cancer more generally (a hot paper in Nature last year worked to distinguish benign CH from that which precedes acute myeloid leukemia to predict AML risk in healthy individuals). Even more broadly, somatic heterogeneity body-wide is an emerging area of research. It’s interesting to see multiple timescales emerging for somatic mutagenic processes, ranging from natural variation taking place with age (ARCH/age-related CH, somatic mutations in the esophagus and colon) to accelerated processes caused by chemotherapy or repair deficiencies in cancer.
A lot of hyped applications of AI in oncology and medicine focus at the discovery stage, whether it’s drug/target discovery or biomarker discovery. Broadly speaking, many of these applications focus on “problems that a human doesn’t know how to solve”, where we’d like the AI model to find signals that we cannot. There’s another class of problems, where humans could solve a problem, but it would be very labor-intensive to do so, incentivizing automation which may require statistical methods. This poster is in the latter category, focusing on matching patients into clinical trials based on their medical records.
Unfortunately, there’s almost no methodological information in the abstract, the poster itself, or indeed on Massive Bio’s website about how they are solving this problem; nor (I think?) is there a specified outcome measurement in the study of how accurate the AI model is at actually determining true trial eligibility.
I think it’s interesting to see different applications of these techniques, but absent technical details (which, to be fair, might be not find the right audience at a clinical meeting), I worry about “AI” being thrown around as a buzzword rather than as an innovative, critical component. Regardless, it’s laudable that a proper study with followup has been put together around this work, and hopefully we’ll get more details and results as time goes on.