donovanr

Yes! Units are up to you (with some constraints)! Physicists do this in practice a lot: https://en.wikipedia.org/wiki/Natural_units

Yes, this. A lot of work in this field is missing from that timeline. Just circa 2010-2020, Les Loew's VCell 3D PDE approaches, Faeder et al.'s BioNetGen / ODE work, Luthey-Schulten Shulten's grid based cell models, the Pittsburgh supercomputing center's 3D monte-carlo MCell, the image-based deep learning models at the Allen Institute for Cell Science...

It's nice to see the idea of virtual cells make a comeback now, though the meaning seems to have shifted to transciptomics-based transformer / gpu-powered models (which have issues[0]), it's a fun field / problem, but I think it will make better progress if we take advantage of all the varied computational work that has come before.

[0] Benchmarking Transcriptomics Foundation Models for Perturbation Analysis : one PCA still rules them all https://arxiv.org/abs/2410.13956

it's uncreative/tired but at least to the point. Too many papers are confused / opaque agglomerations of a year's worth of research shoehorned into a paper. At least with these you can fairly easily assess whether the claim is supported or not.

This rings _so_ true.

While I was in grad school I got bored sitting behind a computer all day, and my wife and I decided to build a tiny house on a trailer as a way of venting our pent-up DIY urges. We'd just build it in our spare time. LOL.

We started in the late summer of 2013, with a trailer and no plans and a stack of construction books from the library.

Cut to spring 2016, having spent every single weekend and most evenings since (in zero degree winters and brutal Pittsburgh summers) sweating and swearing, really pushing the "divorce cabin" line, and having legitimate discussions late at night about the benefits of burning it all to the ground, where my wife, eight month pregnant, is trying to finish the trim work before I submit my dissertation and we tow it across the country.

The way the article captures the "not knowing what we were getting into" / tiny things that delay you to death / stressed out / losing friends / doing absolutely nothing else with your life / so over budget it hurts / final elation at success is absolutely perfect.

We only made it across the finish line because living in Pittsburgh on a grad student stipend is actually, well, livable, and I could do that while my wife worked pretty much full time on our housing boondoggle.

The main learning experience coming out of it was that you should absolutely pay how ever many thousands of dollars it costs for a good set of plans from someone who's done this before. Learning smaller tasks like framing and roofing etc is easy. Stitching it all together into a plan that you're arguing about because neither of you have any idea what you're doing, all while you're wasting precious daylight is _hard_. We would have finished at least a year sooner if we just had plans to follow.

All that said, building a place to live in was super rewarding (as others have said) type II fun.

We still have it, it's beautiful, and I have not yet burned it down.

Allen Institute for Cell Science | ONSITE Seattle | Full-time | software engineer / ML / computer vision | https://alleninstitute.org/

The modeling team at the Allen institute for Cell Science is hiring for two software engineering positions -- a data generalist and a ML / computer vision specialist:

https://alleninstitute.org/what-we-do/cell-science/careers/j...

The Allen Institute for Cell Science aims to impact the entire cell science community. Our goal is to advance understanding of cell behavior in its normal, pathological, and regenerative contexts. Our multidisciplinary team will generate novel cellular reagents, data, models and databases that are informed by and open to scientists around the world. We will produce unique dynamic, visual databases and cellular models that integrate information and data across cellular and molecular sciences.

https://alleninstitute.org/what-we-do/cell-science/careers/

Allen Institute for Cell Science | software engineer / ML / computer vision | ONSITE Seattle | Full-time

The modeling team at the Allen institute for Cell Science is hiring for two software engineering positions -- a data generalist and a ML / computer vision specialist:

https://alleninstitutecellscience.hrmdirect.com/employment/j... https://alleninstitutecellscience.hrmdirect.com/employment/j...

ONSITE, Seattle

The Allen Institute for Cell Science aims to impact the entire cell science community. Our goal is to advance understanding of cell behavior in its normal, pathological, and regenerative contexts. Our multidisciplinary team will generate novel cellular reagents, data, models and databases that are informed by and open to scientists around the world. We will produce unique dynamic, visual databases and cellular models that integrate information and data across cellular and molecular sciences.

https://alleninstitute.org/what-we-do/cell-science/careers/

Very nice. It might be worth worth weighting each data point in proportion to the population it represents.

1. All of the above. The label free tool in particular gives you such a big free lunch at the microscope that the combination of it and good visualization has the potential to massively impact research workflows.

2. We are very interested in how cells change as they divide, differentiate, age, are perturbed by their environment, etc. We study cells in culture right now -- getting good images of in vitro cells from multicellular organisms is way harder. So yes it would absolutely be useful. I don't know if we're going to tackle it ourselves, but one of our core missions is to lay the groundwork for the community to take our tools and run with them -- it's a big win for us if we can bring previously unfeasible research within the realm of the possible.

3. I am a Bayesian at heart, so modeling uncertainty is something that I'm always thinking about. It's high on my list of priorities to do something along these lines.

4. Image similarity is a hard problem. At the end of the day, metrics only get you so far and the proof is in the pudding. Unfortunately there is no ground-truth data to test against -- the probabilistic model was constructed exactly because we can't measure where everything is all at the same time. Some things we do to convince ourselves that we're on track is to see that the variation in the imputed predictions and the actual data are statistically similar, and to see if experts are confounded in differentiating the outputs of our models from actual data. You can read more here https://www.biorxiv.org/content/early/2017/12/21/238378

That's at least close to true. It might be better to think of it as a smoothed "segmentation" of a 3D image, i.e. some algorithm decided what pixels are officially part of e.g. the mitochondria set, and outlined them. That could be based on level sets or seeded watershed or whatever else works well.

There are some alternate visualizations here http://www.allencell.org/3d-cell-viewer.html of data that came off of our microscopes that we also use to visualize our models in house but wasn't;t included in the video. It's hard to visualize varying density 3D data in 2D -- there's no one good way to do it, especially on the fly over the web -- but if you have any feedback about what would be more informative / easier to understand, let us know.

Characterizing cellular variation is exactly what we're interested in, e.g when and why are the mitts clustered around the nuclear vs not. Lots of images in our data have them packed around the nucleus -- you can look at the localizations from our microscopes here (select the Tom20 tag): http://www.allencell.org/3d-cell-viewer.html here. You can also grab the raw data (including bright field images e.g. what you would "see" in the microscope) here http://www.allencell.org/data-downloading.html#DownloadImage...

not sure about specific plans but it's definitely something we're interested in

just saw this; I'm part of the computational modeling team that worked on this -- can try to field any questions or find more qualified people to do so.

would be happy to, but I don't see your email there -- mine's in my profile (I think!) if you'd like to get in touch

Some feedback on the quiz:

- a few of the questions were very good, and either spoke to key high level concepts, or were specific while being language agnostic. (e.g which one of these layers wouldn't you need, why wouldn't this type of classifier work on this data).

- too many of the questions were hyper-focused on the minutiae of word embeddings, tensor flow syntax, SQL queries, and recommender schemes.

- many of the questions were constructed vaguely enough that "I don't know" would be the technically correct answer even though I don't think that was what you were going for.

metadata: recent PhD with serious grad courses in ML and working in DL/CV for the past year using a non-tensorflow framework (PyTorch).

specious extrapolations aside, the plot itself is deceptive -- exponential growth in a log-linear plot should be a straight line, not one accelerating upward super-exponentially.

Sedgewick's 1978 paper[0] on implementing quicksort has some interesting hand optimizations of the assembly code -- loop rotating, unrolling, etc. I wonder if modern compilers do the same?

[0] http://penguin.ewu.edu/cscd300/Topic/AdvSorting/Sedgewick.pd...

Zee's QFT in a Nutshell is a very readable, high-level view of what's going on in QFT. Griffiths' book is his one text I haven't read, but I love his others so much I can't help but second the recommendation.

Right, but it's the 'every once in a while" part that I wonder about, i.e. the variance of that distribution. Maybe you could argue that since the Maxwell-Boltzmann distribution is narrower the lower its mean, if kT is small then it's an exponentially tiny effect.

I've always been curious: I understand that h.nu from cell phone radiation isn't big enough to ionize (say) DNA, but since the molecules in our body are in thermal equilibrium, can't h.nu + kT get it done every once in a while?

[edited because using an asterisk to denote multiplication was a bad idea]

I enjoyed reading this as written rather than as (I assume) intended: "The enormity of defining any job they want can be overwhelming..."

These results seem unconvincing without any error bars / confidence intervals, or even a mention of the number of replicates performed (though the downloadable data set seems to indicate that each test condition was performed only once?)

Have to rep for my favorite math text that pretty much no one else has heard of: Ray Mayer's Introduction to Analysis [1]. It's rigorous, compellingly readable, and accessible to college freshmen. It was never published commercially, but decades of students where I went to school have been delighted by it, tortured by it, or both.

[1] http://people.reed.edu/~mayer/math112.html/index.html

I work at the cellular/molecular scales in a computational biology program; musculoskeletal stuff can often find a home in bioengineering departments. Not sure how much insight I can offer, but feel free to shoot me an email, my contact info is here: http://www.pitt.edu/~donovanr/

I am a 5th year computational biology student, and have yet to take any real coursework in theoretical biology, despite the fact that most of what I do could be broadly classified as exactly that.

Theoretical biology seems to be the black sheep of both the wet-lab biology and simulation communities, a situation that I wish were different, but honestly it can be difficult to make a compelling case for the role/yield of theory in biology beyond a few simple models and natural selection etc.

It's frustrating, because I think the most fruitful area we can work on is developing a better understanding / mathematical formulation of the incredibly high dimensional, nonlinear, stochastic spaces that biological systems live in, but it's difficult to see the way forward because no on (that I know of) is putting time into thinking about it.

Massless GR is sometimes called "pure gravity", and is something people are still interested in (or were, last time I was paying attention) [1]. Kind of remarkable that even the base-case has so many open issues. I should re-read that Torretti book, I remember liking his style but that some of the math was over my head at the time.

[1] http://arxiv.org/abs/0706.3359

The author only encountering to 10^500 sting vacua/landscape issue in 2013 raises some serious crackpot alarm bells [0]. That was common knowledge in the community by 2005 at the latest.

http://math.ucr.edu/home/baez/crackpot.html

I don't mind profanity at all per se, but the tacit misogyny of that one does bum me out. I'm from the US, if it matters.

I'm a CompBio PhD student, and my experience is that folks in that field are much more careful with statistics than in, say, molecular biology labs, but it varies from lab to lab. My PI is exceedingly meticulous about stats -- for instance, we don't report p-values, but rather entire distributions -- but that's because our work is all in silico, so it's easy to run tons of replicate simulations. Wet lab work that's finicky should definitely be held to high statistical standards, but I don't think it's fair to presume everyone in the field guilty until proven innocent.

A lot of the low-hanging fruit, in terms of statistical significance for heritable disease, has already been picked, in the sense of a naive search for SNP X in gene Y that causes disease Z. The more people continue to look for single SNPs or even single gene causal factors, the more we learn that they are of limited use, due to the complexity and redundancy of biological interaction networks (genes/proteins/epigenetic modifications/environment/etc)

One thing that people are definitely still interested in regarding single genes or SNPS is looking for biomarkers. For instance, it's pretty easy to look at RNA abundance with RNA-seq and make a model with good discriminatory power to predict who does or doesn't have (say) breast cancer. But Those tests can be expensive, and profiling the appropriate tissue can be difficult. On the other hand, if we could find SNPs that correlate well with the RNA expression model that predicts disease, then we could just do a cheaper test/faster/easier test for the SNPs. Even better if we can validate causation independently using the emerging, though incomplete corpus of biological pathways.

Beyond that, some folks are excited about the field of functional genomics, which aims to correlate genome-level data to the structure and function of gene products (eg proteins) to get a more low-level, causal look at things.

Biological network extraction is something I'm excited about. Here, you usually want to extract pairs of features, or better yet, higher order structures from data to learn about what factors are important in disease pathways. This puts you immediately in the regime of way more potential features than data points, even if you sequence everyone in the world. I think this consideration worth thinking about when people try to hype you that big data and machine learning will just "solve" biology or medicine. L1 regularization only gets you so far. But it gets you somewhere, and can be hugely useful in suggesting new experiments.

I also think it's important to think about how it may soon be cheaper to store actual RNA/DNA and sequence it on demand, than it will be to store the data itself. DNA sequences would be a fine thing to sequence once and store, but for stuff, maybe not so much.

The etymology of stir dates back to proto-indo-european roots [1], so you might be right.

[1] http://www.etymonline.com/index.php?term=stir