Hacker News Reader

Comment by praccu | Hacker News Reader

As an alternative perspective: I worked in a lab that was relying heavily on intuition and heuristics for exploratory synthesis. By digitizing all of the lab's data and doing some straightforward machine learning, we were able to sufficiently improve the efficiency of the lab, compared to using intuition, and help the chemists develop a deeper understanding of the system they were studying:

http://www.nature.com/news/computer-gleans-chemical-insight-...

So, I guess my question is: to what degree is this reliance on intuition really just a lack of good data driven methods?

View on HN

from original article: "One of the most striking features of chemistry as a science is that very palpable properties like color, smell, taste and elemental state are directly connected to molecular structure."

If the lab notebook doesn't contain that information, no amount of machine learning will find it. There's also intuition in subtle things that don't necessarily make it into the lab notebook (did I shake the extractor vigorously or just a little bit?). Or when running on autopilot, maybe you didn't do something that you wrote down you did - that makes a difference.

If you could automate lab note taking, that would be excellent.

Moreover, this 'principles based search' in the subthread article is about crystallization, which is notorious for being a tractable solution based on high-throughput methodologies (and thus has massive parallelization as a worked out strategy). Synthesis, by contrast, is a low-n experience. When machine learning techniques get good at classification by solving MNIST by looking at 3-5 samples of each letter, then maybe I'll be more optimistic.

>> One of the most striking features of chemistry as a science is that very palpable properties like color, smell, taste and elemental state are directly connected to molecular structure.

That doesn't sound so striking. Smell and taste are both direct physical assessments of molecular structure. There's no model where they could fail to be related to molecular structure.

Yeah that's in the Op's linked article, but would be unaddressed by the sub thread linked article's technique.

https://arxiv.org/abs/1507.02672

Ladder networks get 1.62% error rate on MNIST with 50 labeled samples

In other words, having that good intuition heuristic matters less when you can brute-search a wider range of options. Of course, no one likes being told that idiotically testing many many options is better than "intelligence". I suspect we programmers are going to be OK with this idea, until we're replaced too. :-)

Brute-search is not idiotic testing. It is many times much faster to automate the search and then to move on to the next level/parameter (where you may get to start exploring again).

it might not be that easy. What happens when each iteration takes 30 minutes of setup, 4 hours -> overnight reaction time, 20 minutes of concentration, and a 30 minute silica gel column?

You parallelize and pipeline.

Interns?

Brute-force search is idiotic testing, almost by definition. The smart approach is sparse matrices, optimized either by staring at mountains of data (possible computer-assisted), or by gut feeling. Oftentimes the search space is so large (protein crystallization involves at least buffer, precipitant, pH, concentration) that you need to look at promising spots in the search space.

brute force is "naive" testing. it's almost never better (I say this as the person who ran exacycle, the world's largest brute force supercomputer).

There will still be cases where this isn't possible. In the chemistry case, let's say you look at the structure of your starting material and your product and notice that it has a functional group that your catalyst can be tethered to with a particular reach length and chirality to reach the reaction site... Unless the computer "understands" these concepts building on some other set of principles, it won't be able to creatively generate solutions to those sorts of problems. In some cases, the "n" of directly analogous situations may be in the low single digits or even zero.

Reply on news.ycombinator.com