1. HELPING CHEMISTS FIND JOBS IN A TOUGH MARKET. 2. TOWARDS A QUANTITATIVE UNDERSTANDING OF THE QUALITY OF THE CHEMISTRY JOB MARKET.
What's the job market like for chemists? Dude -- it's always bad.*
How bad is it? How the heck should I know? Quantifying the chemistry job market is what this blog is about. That, and helping chemists find jobs.
E-mail chemjobber with helpful tips, career questions or angry comments at chemjobber -at- gmail dotcom. All correspondence is kept confidential. (Didn't get an e-mail back? It's okay to try again.) Please address correspondence to "Chemjobber" or "CJ."
Voicemail/SMS: (302) 313-6257
Twitter: @chemjobber
RSS feed here
(The Blogger spam filter gets hungry sometimes, and likes to eat comments. You can e-mail me, and sometimes I can get it to cough up your comment. I am always happy to try.)
(*For the literal-minded, this is a joke. Mostly.)
There is a long history of increased automation in pharmaceutical chemistry, and a lot of that development has been praised for allowing chemists to spend more time on the thoughtful aspects of the job rather than the routine physical manipulations. Along those same lines, high throughput experimentation for reaction optimization (which is being increasingly coupled with automated technologies) has allowed chemists to generate more data in one experiment, with the rate-limiting step often being the thoughtful chemist's ability to analyze the data and infer relations to guide the next iteration of experiments. As a chemist with experience with high throughout experimentation, how do you view the application of AI/machine learning for reaction optimization to support drug discovery efforts (e.g., to optimize tough couplings at the med chem or process level)?
ReplyDeleteThis seems like it might be an attainable application of AI that doesn't seem to be discussed as much, even if it wouldn't be as impactful as a computer churning out the perfect target molecule and an ideal route for its synthesis.
In the 1940s, mathematician Norbert Wiener discussed the types of industrial revolutions. The first brought the devaluing of manual labor, and the second targeted route mental work. He foreshadowed a third revolution that would devalue human thinking. Presumably, functional AI will be the full arrival of the third stage. What does this mean for #chemjobs?
ReplyDeleteWe have seen what has happened over the last 50 years to people made obsolete by the first two steps of this process, and it's not like chemists are in a great place as it stands...
Technology is already used for repetitive tasks in chemistry, decreasing the demand for chemist labor. See the Novartis video: https://youtu.be/EwzhEAZRX5o Adding a competent AI would decrease the demand further, but good/novel ideas would still be needed.
ReplyDeleteRight. My question is specifically targeted at the employment outlook after the rise of AI for the chemists currently employed to come up with the "good/novel ideas" when, if we can be honest, execution of truly novel ideas isn't really the bulk of the job.
DeleteHow do we handle activity cliffs? How do we solve the problem of bad or inadequate data without which no AI can work? How do we train a neural net to recognize unexpected MOAs or side effects? How do we go beyond correlation to understanding in deep learning for drug discovery?
ReplyDeleteMy take on AI in drug discovery is captured by Bill Gates's quote: "We always overestimate the impact of technology in the next five years and underestimate its impact in the next fifty years."
DeleteI'd bet the progress in AI personality develops faster than the progress in AI thinking. This is because so much in AI and robotics development is driven by the sex field as well as the medical field. (personal assistants for the elderly) If this happens, it won't be the science lab jobs that get wiped out first, but the science teaching jobs because it will be quite easy for AI to transfer information; all it needs is a good deskside manner. Imagine millions of jobs in the education field go belly up because robotic teachers are just cooler and more effective in dealing with kids than the actual teachers are.
ReplyDeleteI guess I’m a lot more skeptical than most on the prospects of AI in chemistry. I’ve been working in automated chemistry for 15 years and in that time we’ve seen good advances in both the hardware and in the statistics/analysis software. The one big takeaway that we’ve learned is that for any ML or even AI application to work well, the data quality needs to be high (signal to noise with replicate runs) and large (10-100K data points at minimum). Hence the notion that one could build an AI that would scan the literature and find new reactions or synthetic routes is a bit laughable at this point as it’s very much a garbage in / garbage out issue. Although chemistry is a bit better than most biomedical fields in terms of repeatability, the overall quality and documentation isn’t really there. Just think about the difficulties repeating a published procedure for the first time...is my solvent dry enough, what about the exact order of addition and catalyst activation, what was the purity on the reported yield, is there some trick I’m missing that the authors didn’t document...? Until we address the data quality problem, and/or create training sets of good quality chemistry data and procedures from which NLP can extrapolate from the lit., I don’t see AI making rapid headway. Even some of the impressive examples from Merck and others on ultra HTE reaction optimization suffers from severe limitations in the sense that the reaction conditions at micro-scale don’t always translate exactly on scale-up (i.e. heterogeneous bases, residual oxygen, L/M ratios, activation, etc...). Thus I would argue that we need to focus more on standard experimental documentation, reagent characterization, as well as data capture/formatting. Automation coupled with HTE and statistical experimental design are certainly the best way to creating these “AI friendly” data sets, but we are still a long ways from having both the scope and depth of chemical reaction data available to do this.
ReplyDeleteOh yea...on the biology/medical side of things, forget about AI. If one is to trust the Amgen study and others that something like 50-80% of the biomed. Lit. Is not reproducible, than what hope does AI even have. Perhaps it might suggest some testable hypothesis by applying good search tools (with NLP) looking for related documents and patterns, but here the data quality problem is much more pronounced and serves as a serious barrier.
DeleteI think should be a corollary to Betteridge's law of headlines, for "How much will xx impact yy?"
ReplyDeleteAnswer: Not very much.
(there should be)
ReplyDelete