Lung Cancer Genomics

A large lung cancer genomics study has been making a big splash. Using SNP microarrays to look for changes in the copy number of genes across the genome, the group looked at a large batch of lung adenocarcinoma samples. Note: the paper will require a Nature subscription, but the supplementary materials are available to all.

As with most such studies, there was some serious sample attrition. They started with 528 tumor samples, of which 371 gave high-quality data. 242 of these had matched normal tissue samples. All of the samples were snap-frozen, meaning the surgeon cut it out and the sample was immediately frozen in liquid nitrogen.

The sub-morphology of the samples is surprisingly murky; much of the text focuses on Non-Small Cell Lung Cancer (NSCLC), the most common form of lung adenocarcinoma, but the descriptions of the samples do not rule out other forms.

After hybridizing these to arrays, a new algorithm called GISTIC, whose full description is apparently in press, was used to identify genomic regions which were either deficient or amplified in multiple samples.

Many changes were found, which is no surprise given that cancer tends to hash the genome. Some of these changes are huge: 26 recurrent events involving alteration of at least half a chromosome arm. Others are more focused.

One confounding factor is that no tumor sample is homogeneous, and in particular there is some contamination with normal cells. These cells contribute DNA to the analysis and in particular make it more difficult to detect Loss-of-Heterozygosity (LOH), in which a region is at normal copy number but both copies are the same, such as both carrying the same mutated tumor suppressor.

Seven recurrent focal deletions were identified, two of which cover the known tumor suppressors CDKN2A and CDKN2B, inhibitors of the cell cycle regulatory cyclin-dependent kinases. The corresponding kinases were found in recurrently amplified regions; an neat but evil symmetry. Tumor suppressors PTEN and RB1 were found also in recurrent deletions. The remaining recurrent deletions hit genes not well characterized as tumor suppressors. One hits the phosphatase PTPRD -- the first time such deletions have been found in primary clinical specimens. Another hits PDE4D, a gene known to be active in airway cells. A third takes out a gene of unknown function, AUTS2.

In order to gain further evidence that these deletions are not simply epiphenomena of genomic instability, targeted sequencing was used to look for point mutants. Only PTRPD yielded point mutants from tumor samples, several of which are predicted to disable the enzymatic function of this gene's product.

On the amplification side, 24 recurrent amplifications were observed. Three cover known bad actors: EGFR (target of Iressa, Tarceva, Erbitux, etc), KRAS and ERBB2 (aka HER2, the target for Herceptin). Another amplification covers TERT, a component of the telomerase enzyme which is required for cellular immortality, a hallmark of cancer. Another amplification covers VEGFA, a driver of angiogenesis and part of the system targeted by drugs such as Avastin. Other amplifications, as mentioned above, target cell cycle regulation: CDK4, CDK6 and CCND1.

The most common amplification has gotten a lot of press, as it covered a gene not previously implicated in lung cancer: NKX2-1. A neighboring gene (MBIP2) was present in all but one of the amplifications, and so NKX2-1 was focused on. Fluorescent In Situ Hybridization (FISH), a technique which can resolve amplification on a cell-by-cell basis in a tissue sample, confirmed the frequent amplification of NKX2-1 specifically in tumor cells. Resequencing of NKX2-1, however, failed to reveal any point mutations in the tumor samples. RNAi in lung cancer cell lines with NKX2-1 amplification showed a reduction of a commonly-used tumor-likeness measure (anchorage-independent growth). This effect was not seen in a cell line with undetectable NKX2-1 expression, nor was it detected when MBIP2 was knocked down. Previous knockout mouse data has pointed to a key role for NKX2-1 in lung cell development. The protein product is a transcription factor, and the amplification of lineage-specific transcription factors has been observed in other tumors.

What will the clinical impact of this research be? None of the targetable genes which were amplified are novel, so this will nudge interest further along (such as in using Herceptin in select lung cancers), but not radically change things. Transcription factors in general have no history of being targeted with drugs, so it is unlikely that anything will come rapidly from the NKX2-1 observations. On the other hand, there will probably be a lot of work to try to characterize how NKX2-1 drives tumor development, such as to identify downstream pathways.

At least some of the press coverage has remarked on the price tag for this work & the surrounding controversy over the Cancer Genome Project that this represents. The claimed figure is $1 million, which does not seem at all outrageous given the large number of microarrays used (over one thousand, if I'm adding the right numbers) -- a few hundred dollars per microarray for the chip and processing is not unreasonable, and the study did a bunch more (analysis, sequencing, RNAi). If such a study were to be repeated at todays prices in the next 5 big cancer killers (breast, ovarian, prostate, pancreatic, colon), it means another $5M not spent on other approaches. In particular, the debate centers around whether the focus should be on more functional approaches rather than genomics surveys. As fond as I am of genomics approaches, it is worth pondering how else society might spend these resources.

It is also worth noting what the study didn't or couldn't find. A large number of known lung cancer relevant genes did not turn up or turned up only weakly. In particular, p53 is mutated in huge numbers of cancers but didn't really turn up here. The technique used will be blind to point mutants and also can't detect balanced translocations. Nor could it detect epigenetic silencing. If you want to chase after those, then it is more genomics -- which is probably one of the things that eats at critics, the appearance that genomics will never stop finding ways to burn money.

Weir et al. Nature Advance Publication. Characterizing the cancer genome in lung adenocarcinoma. doi:10.1038/nature06358

Imaging gene expression

In one of my first posts I commented on the challenge of obtaining samples for microarray and other biomarker work. Getting samples for microarrays is at best difficult, painful to the patient and only a little dangerous to them; in many cases the samples are simply unobtainable. Getting a broad range of samples from multiple sites, or a time series is going to be very rarely feasible.

With this backdrop, a recent paper in Nature Biotechnology is quite stunning. Indeed, it is a bit of a surprise that it didn't show up in the mother ship or Science: the paper is well written, audacious in design and shows very nice results.

Using actual liver cancer patients the paper correlates contrast-enhanced CT (aka CAT) imaging features to gene expression patterns detected by microarrays using samples from the same patients. While these patients had to go through biopsies, the approach holds out the hope of calibrating imaging assays for future use.

The imaging-microarray connections have many intriguing possibilities. Some of the linked microarray patterns have clear therapeutic associations, such as cell cycle genes and VEGF. Such an imaging approach might, with much further validation, enable appropriate selection of therapeutic agents -- such as Avastin to target VEGF.

The paper also notes the challenges that lie ahead. The choice of liver cancer was no accident: liver tumors tend to be large and well-vascularized, making them straightforward to image using CT. Some of the imaging features found are generic to tumors, but others have some degree of liver specificity. Expression program to image feature mappings may vary from tumor to tumor.

One potential side-effect of this study would be to increase biopharma interest in liver cancer. Liver cancer is a scourge outside of the Western world (perhaps driven by food-borne toxins) but is not in the top of deadly cancers in the U.S. According to some 2002 figures from the American Cancer Society, liver cancer in the U.S. is about 17K new cases and about 15K fatalities -- a horrible toll, but far less than 160K annual lung cancer deaths. One big attraction for companies is potential payoff, but another is the potential for accelerated development decisions. Being able to subset patients based matching drug mechanism to biology inferred from imaging is potentially a powerful means to do that.

Tagging Up With Protein Microarrays

Molecular Systems Biology, an open access journal, has an impressive new functional protein microarray paper. The authors identified a large number of targets for a yeast ubiquitin transferase (enzymes which transfer a protein tag, ubiquitin, onto other proteins), and the data has a good ring to it.

Some background: protein microarrays are a much more complicated subject than nucleic acid microarrays. One way to split them is by intent. Capture arrays have some sort of affinity capture reagent, most likely antibodies, on the chip surface. If properly designed, built & calibrated they represent a very highly multiplexed set of protein assays. Reverse-phase protein arrays spot fractionated, but unpure, proteins from biological samples on an array.

In contrast, functional protein microarrays attempt to represent a proteome on a chip as individually addressable spots in order to study aspects of that proteome. A number of groups have worked on functional protein microarrays, but there are a limited number of commercial sources, with perhaps the most successful being Invitrogen, which offers human and yeast arrays. If you'd like a great beach book on the subject, a new volume covers a wide array of topics, with Chapter 22 ("Evaluating Precision and Recall in Functional Protein Arrays") definitely my favorite.

Functional protein arrays present a huge challenge. In the ideal case the proteins would be produced, folded correctly and deposited on the slide in such a way that an assay can be run on every protein in parallel. This is a tall order, with lots of complications. Proteins may not fold correctly during expression or may unfold in the neighborhood of the slide surface, the post-translational state of the protein may be variable and is unlikely to capture all possible states of the protein, and the protein may not have key partners which are important for its function.

Despite these, and many other concerns, protein microarray experiments have been published describing various feats. Protein-protein interaction experiments to discover novel interactions (such as this one) or create comprehensive binding profiles (such as this one) are probably the most prevalent use, but the arrays can also be used to discover DNA binding proteins, identify novel enzymes, assay phenotypic differences of mutants, develop novel infectious disease diagnostic strategies, and identify the targets of protein kinases. [links are a mix of open access & paid access; apologies)

A wide variety of ingenious methods have been used to produce functional protein microarrays. The Invitrogen arrays are spotted from purified expressed protein and expected to bind randomly, but some other approaches ensure that the majority of protein molecules bind in a defined way. Some approaches actually synthesize the proteins in situ, and one group even deposited proteins on spots using a mass spectrometer!

Protein microarrays have had their growing pains. The amount of active protein found in a spot can vary widely. One study of protein-protein interactions failed to recover most of the known interactors of the bait protein. Since the bait is primarily a phosphoprotein binding protein, one possible explanation is that the insect-expressed human proteins were not in their correct phosphorylation state. However, poor recall of known substrates was also observed in protein kinase substrate searches run in both human and yeast (see Chapter 22 of the Predki book). Even without worrying about post-translational modification, coverage is an issue. While essentially the complete Saccharomyces proteome is available, the most extensive commercial human chip has less than 1/5th of the proteome and there are not (last I checked) commercial arrays for any other species.

The new publication wins on a bunch of scores. First, it is one of the handful of publications using such arrays which is not from one of the labs pioneering them, suggesting that they might work routinely. This publication uses the Invitrogen yeast arrays. Second, they did recover a lot of known substrates for their ubiquitinating enzyme. Third, the signals look very strong by eye, which has been the case for protein-protein interaction assays but much less so for protein kinase substrate discovery. Fourth, they batted 1.000 with novel positives from array in an independent in vitro ubiquitination assay and were able to verify that at least some of these are ubiquitinated by Rsp5 in vivo (by comparing ubiquitination in wt and Rsp5 mutant strains). Fifth, they performed a protein-protein interaction microarray assay with Rsp5 and the interaction results and ubiquitination results strongly overlapped.

Of course, I used to work at Ubiquitin Proteasome Pathway Inc (which is now touting a new drug with a new target in the pathway), and there I would have been digesting this paper until arrays danced in my dreams. Such assays offer an interesting possibility for greatly expanding our understanding of UPP players and functions -- many Ub transferases or Ub-removing proteases have no known substrates. While they have a lot of issues, functional protein microarrays are starting to make a difference in proteomics.

A good RNAi guide

Cell Cycle is an interesting little journal that publishes many papers open access. A nice little review of the statistical treatment of genome-wide RNAi is available freely.

The review focuses on noise and variance in RNAi screens, and doesn't explore some of the other key issues such as off-target effects, interferon response and appropriate cell lines. So it isn't a complete guide but a sharply focused one.

A recent Nature has a paper on genome-wide RNAi for targets increasing sensitivity to the key antitumor agent paclitaxel (alas, not free).

Genome-wide RNAi with siRNAs is a powerful technology, but it requires a pretty large investment in automation to make it work. That will slow the widespread adoption of the technology, which isn't entirely bad. In some ways, mRNA microarray technology spread too far too fast leading to many bad papers being published before the methodologies were well worked out. Of course, there are still lots of bad microarray papers being published, but you can't make the horse drink. Some bad papers have poor microarray analysis, and others are just atrocious experimental design. In the end, the technology has been besmirched, generally unfairly.

Breast Cancer Genomics

This month's Cancer Cell has a pair of papers (from the same group), plus a minireview, on breast cancer genomics.

One paper focuses on comparing 51 breast cancer cell lines to 145 breast cancer samples, using a combination of array CGH and mRNA profiling. The general notion is to identify which cell lines resemble which subsets of the actual breast cancer world. Cell lines long propagated in vitro are likely (almost assured) to have undergone evolution in the lab; this means they are not the perfect proxies for studying the disease. Array CGH is a technique for examining DNA copy number changes, which are rampant in many cancers. Its use has exploded over the last few years, with a number of interesting discoveries. It is also a useful way to fingerprint cell lines; at least one cell line was described recently as an imposter (wrong tissue type), but I can't find the paper because of the huge flood of papers a query for 'array CGH' brings up.

The second paper looks at a set of clinical samples from early breast cancer, and again uses both transcriptional profiling and aCGH. I need to really dig into this paper, but the abstract has some interesting tidbits (CNAs=copy number abberations) -- emphasis my own

It shows that the recurrent CNAs differ between tumor subtypes defined by expression pattern and that stratification of patients according to outcome can be improved by measuring both expression and copy number, especially high-level amplification. Sixty-six genes deregulated by the high-level amplifications are potential therapeutic targets.

The mini-review does highlight a key point: as impressive as this study is, no study can ever hope to be the final word. As new omics tools are developed, new studies will be desirable. Two obvious examples here: running intensive proteomics and looking in depth at alternative transcripts.

Red Alert Mr. Pseudomonas!

I finally decided that four weeks of laryngitis was perhaps too long and got myself in to the nurse practitioner, who obliged me with an antibiotic script. Our bar for using antibiotics has historically been too low, but perhaps I overshot in the other direction.

Or maybe not. A recent paper in PNAS presents the provocative thesis that low doses of antibiotics can stimulate nasty traits in pathogenic bacteria. Using a microarray and low doses of three structurally unrelated antibiotics, they detected switching on of a number of unpleasant genetic programs.

All three antibiotics induce biofilm formation; tobramycin increases bacterial motility, and tetracycline triggers expression of P. aeruginosa type III secretion system and consequently bacterial cytotoxicity. Besides their relevance in the infection process, those determinants are relevant for the ecological behavior of this bacterial species in natural, nonclinical environments, either by favoring colonization of surfaces (biofilm, motility) or for fighting against eukaryotic predators (cytotoxicity)

The authors go on to suggest that antibiotics may be important signalling molecules in natural communities. This is in contrast to the older model of antibiotics as weapons in microbial battles for dominance. It is a provocative thesis worth watching for stronger evidence. In my mind, their data still fits the weapons model -- what they see is the same sort of signalling as my blood in the ocean signals a shark. Or, a bacterial Captain Kirk detecting an unseen ship raising its shields, prompting a defense posture.