tag:blogger.com,1999:blog-9633767.post1594181627785279637..comments2024-03-25T09:11:17.877-07:00Comments on The Curious Wavefunction: Will CADD ever become relevant to drug discovery?Wavefunctionhttp://www.blogger.com/profile/14993805391653267639noreply@blogger.comBlogger5125tag:blogger.com,1999:blog-9633767.post-67620548188736672202019-01-05T11:48:12.543-08:002019-01-05T11:48:12.543-08:00Agree that machine learning has been oversold...ho...Agree that machine learning has been oversold...however also agree with the post author that it is a powerful technique as long as we recognize its limitations and restrict it to its proper domain of applicability, and it is -- HAS to be -- part of the future of CADD. I think that just as overhyping ML is dangerous, so too is the kind of skepticism you encounter among (some) medicinal chemists who are skeptical at least partly because they don't know what it is. I have worked with synthetic/medicinal chemists who were deeply skeptical of all this "machine-learning stuff", who usually had only a very vague notion of what it was, but would at the same time make arguments about which compounds to make based on imagined correlations from a simple plot of logP vs activity that to be generous had an R^2 of perhaps 0.2 or 0.3 at the most. There is danger in overhyping a useful tool; there is also a danger in rejecting it because I'm Not Sure What It Is And Anyway It's Not How We Do Things Around HereAnonymousnoreply@blogger.comtag:blogger.com,1999:blog-9633767.post-27206926163763814202018-12-14T13:29:56.241-08:002018-12-14T13:29:56.241-08:00Interesting article, thanks for posting. Feels ove...Interesting article, thanks for posting. Feels overly negative to me if I'm being honest. One thing to highlight is there are a number of recent oncology agents where comp chem has been key to their discovery. Similar to the Pete's comment, maybe I've just been fortunate to work at places where comp chem was an embedded and valued part of a discovery team. Just need to be open and honest about what comp chem can and can't do. Out of the chemistry specialisms comp chem feels to have more employment opportunities than most. Finally, I think we need to be careful not to oversell AI/ML or we'll go back to the days of hype and over promise. Great discussing though! Anonymousnoreply@blogger.comtag:blogger.com,1999:blog-9633767.post-77739685063012293932018-12-13T16:09:54.604-08:002018-12-13T16:09:54.604-08:00Hi Ash, here are some thoughts on the subject.
I ...Hi Ash, here are some thoughts on the subject.<br /><br />I don’t think that it’s useful to compare CADD with synthetic chemistry. I would draw a distinction between medicinal chemistry (deciding what gets made and/or tested) and synthetic chemistry (making what needs to be made) even when the same person performs both roles. CADD is ultimately part of medicinal chemistry and one mistake that CADD scientists sometimes make is to treat medicinal chemists as (purely) synthetic chemists. Towards the end of my time at AZ, we introduced design teams which included medicinal chemists, synthetic chemists, CADD scientists, physical chemists and DMPK scientists. <br /><br />In my experience, CADD methods tend to have more impact in lead identification than in lead optimization. Lead optimization often focuses on solving specific problems and a CADD scientist with a strong background in physical chemistry can often provide useful input. However, this problem solving is not the same as trying to suggest a development candidate. It’s also worth pointing out that drug design is incremental and it is generally easier to predict the effects of relatively small (i.e. within structural series) changes than it is to predict activity or properties directly from molecular structure. The difficulty for QSAR modelers is that QSAR is a data-hungry way to make predictions. My view is that a big part of drug design is assembling the data needed by the project team as efficiently as possible and this is more Design of Experiments than drug design.<br /><br />Another problem that CADD scientists encounter is that CADD capabilities have traditionally been oversold. These days, it seems that any regression or classification model can be described as ML and can therefore be touted as AI. A lot of Kool-Aid is getting drunk.Peter Kennyhttps://www.blogger.com/profile/12180360326821860667noreply@blogger.comtag:blogger.com,1999:blog-9633767.post-22666745661280359182018-12-13T08:37:18.512-08:002018-12-13T08:37:18.512-08:00This is somewhat true even in fields where CAD has...This is somewhat true even in fields where CAD has a proven track record, ie, semiconductors. The issues are similar. The predictive power is only with respect to the action of dopants and crystal structure, and everything else is fairly obvious generalities. As a colleague of mine used to joke, TCAD is excellent at solving yesterday's problem. I don't think AI/ML is really CAD. And I don't think the size of the training sets will reach critical mass in the next 10 years. Commercial data will likely remain silo'd which hinders NN training. That does mean that people should not be trying. NNs was a backwater field until 2012 and now registration for NIPS sells out in 10 minutes and people use bots to nab slots. Anonymousnoreply@blogger.comtag:blogger.com,1999:blog-9633767.post-61654054274452377802018-12-11T20:54:48.985-08:002018-12-11T20:54:48.985-08:00Rutherford Ernest physics background won Nobel Pri...Rutherford Ernest physics background won Nobel Prize in chemistry. Dan Schectman metallurgists won NL in chemistry. V. Ramakrishnan , physicist won Nobel in chemistry. This year F Arnold , aerospace and mechanical engineering background then PhD in Biophysical chemistry. Has any chemistry background undergrad won a Nobel in physics? Anonymousnoreply@blogger.com