In grad school I had a subletting roommate for a while who was writing code to match some experimental data with a model. He showed me his model. It was quite literally making random combinations of various trigonometric functions, absolute value, logarithms, polynomials, exponents, etc. into equations that were like a whole page long and just wiggling them around. He was convinced that he was on a path to a revolution in understanding the functional form of his (biological) data, and I believe his research PI was onboard.
I guess "overfitted" never made it into the curriculum.
'‘A new scientific truth does not triumph by convincing its opponents and making them see the light, but rather because its opponents eventually die, and a new generation grows up that is familiar with it.’ This principle was famously laid out by German theoretical physicist Max Planck in 1950 and it turns out that he was right, according to a new study.'
https://www.chemistryworld.com/news/science-really-does-adva...
Also the story of Ignaz Semmelweis who discovered that if doctors washed their hands it reduced deaths during childbirth - but for a variety of reasons his findings were resisted.
https://www.npr.org/sections/health-shots/2015/01/12/3756639... https://www.npr.org/sections/health-shots/2015/01/12/3756639...
Point being, as awesome as science is, it's still a human enterprise, and humans are still, well, human.
Then they go on conferences and brag about it, because they have to (otr they know it's bs). Datasets are soso (you can have a look at QM9...) and for more specialized things, people generally don't bother trying to benchmark or compare their results on a common reference. It's just something new...
And with all that: even without doing fancy statistical methods without knowing too much about it, your theoretical computations might not make so much sense (at least in the sheer number which is pumped out and published)...
Well, that's a new acronym for me. I wonder where it came from. Apparently it's "on the real". Sounds like AAVE?
> AAVE
OK.
Technically, we call that a "neural network". Or "AI".
Yes. It just turns out it's a particular human, whose analysis is very very dumb.
People have figured that out long ago [1] (I know the author of that paper lately turned somewhat controversial, but that doesn't change his findings). It's not very widely known in the general public. But if you understand some basic issues like p-hacking and publication bias and combine that with the knowledge that most scientific fields don't do anything about these issues, there can hardly be any doubt that a lot of research is rubbish.
[1] https://journals.plos.org/plosmedicine/article?id=10.1371/jo...
After the 11th nested 'if' statement, I upped the request to a case of beer. I'm not certain he ever got the code working.
To the larger point, scientists are not programmers. They got into their programs to do research. What keeps them going is not the joy of programming, but the thrill of discovery. Programming is nothing but a means to an end. One they will do the bare minimum to get working. Asking hyper stressed out grad students to also become expert coders isn't reasonable.
And yes, that means that the code is suspect at best. If you load the code on to another computer, make sure you can defenestrate that computer with ease, do not use your home device.
I could replace "programming" in your above little bit with "mathematics" and it would be just as weird.
Our modern world runs on computers and programs, just as our modern world and modern science built itself on mathematics and required many to use it. So too the new world of science may require everyone to know to program just as they know about the chemical composition of smells, or the particulars of differential equations, etc.
And I know your argument isn't "they shouldn't learn programming", but honestly since I keep seeing this same line of reasoning, I can't help but feel that is ultimately the real reasoning being espoused.
Science is getting harder, and its requirements to competently "find the exciting things" raises the bar each time. I don't see this as a bad thing. To the contrary, it means we are getting to more and more interesting and in-depth discoveries that require more than one discipline and specialty, which ultimately means more cross-functional science that has larger and deeper impacts.
So what you end up with are that great scientists that are decent programmers are the ones who can do the cutting edge science at the moment.
Again: these are tools that are means to an end. They only need to work well enough to get the researcher to that end.
A lot of what are considered essential practices by expert programmers are conventions centered around long-term productivity in programming. You can get a right answer out of a computer without following those conventions. Lots of people did back in the day before these conventions were created.
That's not to say that everybody with horrible code is getting the right answers out of it. I'm sure many people are screwing up! My point is just that ugly code does not automatically produce wrong answers just because it is ugly.
By analogy, I'm sure any carpenter would be horrified at how I built my kayak rack. But it's been holding up kayaks for 10 years and really, that's all it needs to do.
I will add that in general, statistical analysis of data is not by itself adequate for scientific theory--no matter how sophisticated the software is. You need explanatory causal mechanisms as well, which are discovered by humans through experimentation and analysis.
And you can do science very well with just the latter. Every grand scientific theory we have available to us today was created without good programming ability, or really the use of computers at all. Many were created using minimal math, for example evolution by natural selection, or plate tectonics. Even in physics, Einstein came up with relativity first, and only then went and learned the math to describe it.
I feel like the later is obvious: of course the tools aren't science, but if you want to do real work and real science, your tools are going to be crucial for establishing measurements, repeatability, and sharing how one models their hypothesis onto real world mechanics.
Likewise, the former is just the same commonly repeated thing I just argued against and my reply is the same: so what? You building a kayak is not science and is irrelevant.
Scientists can't reach a meaningful conclusion without proper use of tools. All they can do is hypthesize, which is certainly a portion of science (and many fields are in fact stuck in this exact stage, unable to get further and come to grounded conclusions), but it is not the end-all of science, and getting to the end in the modern day science means knowing to program.
Of course there are exceptions and limitations and "good enough". No one is arguing that. The argument I am refuting is those who think "tools are just tools, who cares, I just want my science". That is the poor attitude that makes no sense to me.
I'm just trying to make the point that "proper" is subjective. Software developers evaluate the quality of code according to how well it adheres to well-established coding practices, but those practices were established to address long-term issues like maintainability and security, not whether the software produces the right answer.
You can get the right answer out of software even if the code is ugly and hacky, and for a lot of scientific research, the answer is all that matters.
Sure, it would be great if we all had more time to learn how to code. Coding is important. But I'd say the onus should be on coders to build better tools and documentation so they are empowering people to do something other than code, rather than reduce everything to a coding exercise because making everything look like code means less boring documentation and UX work for coders.
I mean, biology is in fact a full on degree program and you pretty much need a PhD before you're defining an original research topic. It's not because biologists are dumber and learn slower. It's that biology is complicated and poorly understood, and it takes years to learn.
Contrast this to coding... you don't even need to go to college to launch a successful software product, and the average person can became proficient after a few years of dedicated study. However, this is a few years that biologists don't have, as their PhDs are already some of the longest time-wise to finish.
The decision to rename genomes is totally consistent with the biologists MO: if a cell won't grow in a given set of conditions, change the conditions. Sure we can CRISPR edit the genes to modify a cell to to grow in a set of conditions, but if it's usually far easier to just change the temperature or growth media than to edit a cell's DNA.
My take away is that this is more a failure of programmers and/or a failure of their managers to guide the programmers to make tools for biologists, than of biologists to learn programming. Sure, coders get paid more, but they aren't going to cure cancer or make a vaccine for covid-19 without a biologist somewhere in the equation. And I'm glad the biologists developing vaccines today are doing biology, and not held up in their degree programs learning how to code!
I would say most research, to an ever growing degree, is so heavily dependent on software that it's tough to make that claim anymore. It makes no sense to me. It's like saying Zillow doesn't need software engineers because they are in the Real Estate business, not the software business.
I mean, sort of. Some research is essentially just programming; other research can get by with nothing but excel. Regardless, it's unreasonable to ask most scientists to be expert programmers -- most aren't building libraries that need to be maintained for years. If they do code, they're usually just writing one-shot programs to solve a single problem, and nobody else is likely to look at that code anyway.
What if you want to share data with a wetlab biologist who want to explore their favorite list of genes on their own?
Not that I'm saying using excel is bad either. I use excel plenty to look at data. But scientists need to know how to use the tools that they have.
THe basic assumption I have is that when I input data into a system, it will not translate things, expecially according to ad-hoc rules from another domain, unless I explicitly ask it to do so.
It's not clear what data input sanitization would mean in this case; date support like this in Excel is deeply embedded in the product and nobody reads the documentation of Excel to learn how it works.
If you're deciding who gets a large-scale computational biology grant, and you're choosing between a senior researcher with 5000 publications with a broad scope, and a more junior researcher with 500 publications and a more compuationally focused scope, most committees choose the senior researcher. However, the senior researcher might not know anything about computers, or they may have been trained in the 70's or 80's where the problems of computing were fundamentally different.
So you get someone leading a multi-million dollar project who fundamentally knows nothing about the methods of that project. They don't know how to scope things, how to get past roadblocks, who to hire, etc.
You might occasionally run into someone who is passable - at best - with R or Python. But most of the code they might write is going to be extremely linear, and I doubt they understand software architecture or control flow at all.
I don't know any biologists who program for fun like me (currently writing a compiler in Rust).
I'd say that getting some basic data science computing skills should be more important than the silly SPSS courses they hand out. Once you have at least baseline Jupyter (or Databricks) skills you suddenly have the possibility to do actual high performance work instead of grinding for gruntwork. But at that point the question becomes: do the people involved even want that.
Most of the code I write to do biological data analysis is fairly linear. However, I also generally use a static type system and modularity to help ensure correctness.
I've perused a lot of code written by scientists, and they could certainly learn to use functions, descriptively name variables, use type systems and just aspire to write better code. I just saw a paper published in Science had to issue a revision because they found a bug in their analysis code after publication that changed most of their downstream analysis.
It one of the reasons why people end up with spreadsheets. Most of their data is giant tables of data. Excel does very well at that. It has a built in programming language that is not great but not totally terrible either. Sometimes all you need is a graph of a particular type. Paste the data in, highlight what you want, use the built in graph tools. No real coding needed. It is also a tool that is easy to mismanage if you do not know the quirks of its math.