Yeast de résistance

May 07, 2013

I kickstarted my now defunct academic lab 5 ½ years ago with a simple yeast genetics experiment that I now want to repeat and scale up as an independent scientist in preparation for rare disease drug discovery. The inspiration for this experiment dates back to my grad school days, when I observed that yeast cells overdose on psychoactive drugs. At the time, many an estimable scientist told me that psychoactive overdose in yeast was non-specific, but underneath the jargon what they were really saying was that it was uninteresting.


But I couldn’t understand how that could be when yeast cells were overdosing on antidepressants like Zoloft, a selective serotonin reuptake inhibitor (SSRI) whose seemingly sole drug target — the serotonin transporter protein — isn’t even present in yeast. Single-celled yeast don’t produce serotonin let alone have a brain, so what’s at the root of Zoloft overdose? Conflating non-specific and uninteresting doesn’t make sense from a mechanistic standpoint, because there are many ways to kill a cell, but it’s a matter opinion as to whether these mechanisms are deemed worthy of study.


So in the Fall of 2007, I let genetics be the arbiter: I and Meredith Rainey McClure, my first technician at Princeton, selected for yeast mutants that are resistant to Zoloft overdose. If the molecular basis of overdose were non-specific, resistance-conferring mutations would occur in genes that are involved in drug removal or detoxification, pathways that would also confer resistance to drug overdose in general, without regard for chemical structure. On the other hand, if the mechanism of overdose were specific, resistance-conferring mutations would occur in genes that did not modify cellular responses to just any drug.


As shown in this table from my Princeton lab’s 2010 Genetics paper, we observed resistance-conferring mutations that fall into two classes:


table 1


The first class includes CUP5, TFP1 and VMA9, all components of the V-ATPase complex (below, left). Mutations in the V-ATPase affect drug response by a mechanism that was first elucidated decades ago by Christian de Duve, who discovered the waste management facility of the eukaryotic cell called the lysosome (vacuole in yeast), and after which the drug accumulation known as lysosomotropism is named. I blogged last summer about the second class, SWA2 and CHC1, the latter encoding for a protein called clathrin. Clathrin molecules assemble in to the buckyball structures (below, right) that encapsulate lipid vesicles, including the vesicles that store neurotransmitters like serotonin in the synapse:


The 14-subunit vacuolar ATPase complex (left) and an assembled clathrin coat (right)


As we argued in the Genetics paper, the pattern of Zoloft overdose-resistant mutations was consistent with specific mechanisms that modify the accumulation of drugs called cationic amphipaths, or in slightly friendlier terms, hydrophobic weak bases. So what about overdoses caused by other psychoactive cationic amphipaths? In unpublished work that we’ll be submitting soon, it turns out that the genetic basis of psychoactive drug-overdose resistance is complex in yeast, such that distinct pharmacological classes like antipsychotics and antihistamines exhibit more than one resistance mechanism.


In 2007-2008, the state of the art in mutation detection (“genotyping”) were whole-genome yeast tiling arrays from Affymetrix. (Remember them?) Those arrays cost $150 a pop, and that was with an institutional discount to Princeton researchers, who originally helped design and validate the technology, and as far as I can tell were the only people ordering these arrays. However since 2008, mutation detection has advanced to the point where one can sequence an entire 12 million base pair yeast genome 30-times over for $15 or less, so I myself could in theory do in my second bedroom in a few weeks what it once took a full-time technician six months at more than 10-times the cost. But I’m not sure my wife is too keen on the idea of an experimental Man Cave just yet, so I’m on the hunt for kosher lab space.


If I had gotten a job as an assistant professor, I would have done exactly these experiments right out of the gate. Here’s what I wrote in the research statement that was part of my application package:


I propose to generalize yeast-based genetic resistance selections that made possible Project 1 to other antidepressants, as well as to three other pharmacological classes bedeviled by molecular promiscuity: antipsychotics (both typical and atypical), antihistamines (both sedating and non-sedating), and anesthetics. Using whole-genome sequencing technologies, robotic automation and miniaturization, it will be possible to perform many single, or even combination, de novo drug-resistance selections in parallel. Follow up investigation of specific conserved cellular pathways revealed by this approach will be guided by the template established by my experience with sertraline, including the use of radiolabeled analogs to validate novel targets or target pathways.”


As I blogged earlier this year, no dice. At least one of the departments I applied had the professional courtesy to actually explain albeit in a terse, unofficial email of rejection why I didn’t make the cut: “one of the issues was the relevance of understanding how neuro drugs like antidepressants work by studying their action in non-mammalian organisms.” I obviously couldn’t disagree more strongly with this short-sighted view, which shows just how far pharmacology has drifted away from the tenets of cell biology and genetics. Eroom’s Law and the wholesale abandonment of CNS drug pipelines by Big Pharma is proof that we don’t have enough basic understanding of how existing drugs work. So how on Earth are we supposed to create new first-in-class drugs for common or rare diseases without the basic understanding that comes from studying simple genetic model organisms like yeast?


So here’s my plan. Before, I was interested in studying individual drug-resistant mutants, but as I’ll explain I now want to take a statistical approach by pooling all resistant colonies from a drug-laden agar plate, like this one:


Drug Resistance


Each of these colonies sprouts from a single spontaneous mutant. In fact, some of these colonies are clones, so the same mutation may be represented on the plate in multiple copies. And as the above Genetics table shows, en toto there are multiple genes with resistance-conferring mutations. So is there a way in one fell swoop to identify all mutations? I plan to scrape all mutants off a plate like the one above into a gene pool, literally; then extract genomic DNA and outsource sequencing using a site like Science Exchange. These pooled whole-genome datasets would yield the entire distribution of resistance-conferring mutations for each tested drug, with a given mutation present at a frequency in the pool that is proportional to the number and size of colonies harboring it.


I’d start with Zoloft as the baseline, since I have a concrete expectation of the genes that are likely to pop up. For example, I expect that V-ATPase mutants will be common across many resistance selections. The downside of this distributional approach is that I lose information like whether a mutation is recessive or dominant, or even whether resistant colonies harbor one or more than one resistance-conferring mutations. And I wouldn’t be able to study mutants individually. But using yeast strain construction techniques, I could after the fact regenerate specific mutants to test in secondary assays of cellular physiology.


While I figure out where to do these experiments, which in theory could take a few weeks if all goes without a hitch, I’d be interested in finding a computational collaborator who would help me with variant calling from pools of drug-resistant mutants. It would also be helpful to do some simulations before any physical experiments are done to get a sense of sensitivity. In other words, if there are 100 mutant colonies on a plate ranging in size, would this approach be able to detect the rarest mutant? What is the minimum genome coverage required to identify all mutants on a plate? How many independent drug-laden plate replicates would be needed to distinguish in a statistically significant way between close structural analogs, for example between a secondary and tertiary amine tricyclic?


Think of this post as an experiment in open review of a pre-experiment hypothesis. Keen to hear your thoughts!


  • André Rendeiro

    I’m really excited about this project and wish you all the luck!

    • Ethan O. Perlstein


  • Michael

    My background is more systems neuro than cellular (or pharmacology, for that matter), so I’m a little out of my depth here. So can you explain how identifying genes responsible for protecting yeast against toxic drug effects will connect to neuropharmacology?

    Is it that you are identifying previously unknown cellular interactions that may be responsible to non-specific effects of the drugs through analogous pathways in humans?

    In neuroscience, another way that people distinguish specific from non-specific effects is dose. Are the concentrations of drugs in your cultures ‘physiological’?


    • Ethan O. Perlstein

      Thanks for the comment, Michael!

      It’s been known since the 60s and 70s from elegant electron microscopy and cell physiology experiments that psychoactive hydrophobic weak base drugs accumulate in cells, especially over long-term (weeklong+) treatments. However, once the cloning revolution started in mid-70s, pharmacologists started to focus almost exclusively on drug-protein target interactions (“targetophilia”), and studies of drug-lipid target interactions languished. Antidepressants exhibit a therapeutic lag, inducing neurogenesis after chronic but not acute treatment. No one knows why it takes weeks or months before they have effects. I think part of the explanation could involve drug accumulation in acidic compartments and subsequent reprogramming of vesicular trafficking patterns, which are specializations that make brain cells brain cells.

      So the answer to your second question is yes, but I have no reason to assume that drug-lipid interactions would explain only the “non-specific” effects. Btw, I hate that word and think the entire concept of non-specific. There are varying degrees of specificity, and one must also distinguish between affinity vs avidity, which almost never happens.

      The physiological concentrations of psychoactive drugs is still in many ways an open question. Plasma/serum concentrations are much lower than effective brain or CSF concentrations, and much lower than the effects of psychoactives on model organisms. That said, accumulation over time is what it key to drug-lipid interactions, and the measurements that have been made put drugs like antidepressants in the low micromolar, exactly where we see physiological effects in model orgs. Pharmacologists tend to obsess over Kd values but those are most relevant in in vitro assays with reconstituted synapses — synaptosomes. Hardly faithful models of neuronal cell function.

      • Bill Hooker

        Regardless of “physiological or not”, do you see a dose-response: that is, if you vary the drug concentrations systematically, what happens?

        Second question: are there any existing mouse, Xenopus, drosophila or other multicellular model mutants for any of the genes you’ve identified? If so, do they tell you anything and/or would they be useful models with which to connect your work to something easier to “sell”?

        • Ethan O. Perlstein

          Yes! I’m not saying this about you but if people, specifically pharmacologists, actually read the two yeast papers we published on Zoloft action then they would know the effects are of cationic amphipaths are biphasic. That’s been seen in all many of model systems, even red blood cells.

          All of the genes we identified are evolutionarily conserved over 1 billion years, so yes. In unpublished work we showed that V-ATPase and ARF1-regulated vesiculogenesis genes also affect drug accumulation in cultured rat neurons. If I had had more $ and time I would have pursued more traditional model organisms, but setting up mouse work is highly non-trivial, as you know. The genesis of Crowd4Discovery was my search for a mouse psychopharmacologist to collaborate with in order to translate the yeast and cultured neurons studies!

          The only reason this has to be “sold” is because pharmacology is so adrift. And I seriously underestimated how difficult it would be to break through. I take responsibility for not being a good enough salesperson. But if that’s the reason I didn’t get an academic job in the departments I applied to, then is that’s an indictment on them, not the science.

  • Graham Ruby

    It’ll work, but the informatics will be a minefield of potential analytical erros esp given the low frequency with which you are anticipating to see even successful variants. It looks like you have 100-1000 colonies on that plate, depending on where you draw the line between big colonies versus small colonies (at some point of smallness, the seq error rate will catch up with the expected frequency of reads with a particular error from a small colony). I think it will work pretty much as you describe it, but I would break up the surface of the plate into 10 or so sectors for sequencing (barcoding). Easy enough to do with a scalpel, and you’ll end up with 10-100 colonies per sector – that will allow you to be confident about smaller-colony mutants. You will loose only a little bit of the power of comparing frequencies from one barcoded dataset to another, but presumably you will take some pictures of the sectors before/after you cut them, so you can get a good estimate of the total number of colonies and yeast cells per sector from the image, and get good cross-barcode comparisons by normalizing using that and the read count from each barcode, yada yada yada… The other advantage is that, since rare SNP calling from a complex dataset is similar to SNP calling in a diploid genome dataset but not as mature, you will probably have to experiment a little with different analytical techniques – much better to do on a dataset a tenth the size (one sector) that still represents the full scale of data for the experiment (the additional sectors will not have significantly different amounts of data or noise).

    • Ethan O. Perlstein

      Thank you so much, Graham! This is exactly what I was looking for in terms of feedback.

      Sectoring the plate is an elegant solution!

      • Titus Brown

        Graham said what I would have said, but better. +1. Main thing to do is generate WAY more data than you think you’re going to need.

    • Jessica Chong

      I really like the sector + barcoding idea. I suspect the rare mutation calling problem will be similar to what you see in cancer tumor sequencing (expecting multiple rare interesting variants in the pool that will be hard to detect), so you might want to take a look at that literature.

  • Andrew H

    What are the conventions for knowing if you have saturated your resistance screen? After sectoring the plate could you replica plate the resistant lines at different doses, isomers,other compounds, etc. to extract more phenotypic information about how a potential mutation within a sector and subsequently mapped may impact resistance? Good luck!