![]() ![]() ![]() There are a few neat ideas that can help. There are pretty good and cheap models such as cepC. I'm confused about saying you can't do a DFT model to compare on a large number of compounds. One nice thing is as long as you keep an eye on uncertainty, you can be selective in which new data points you acquire. I would say a few hundred at least to predict static properties, though if you want a machine-learned potential (i.e., a force field) to predict dynamics you might need millions. It depends on how dissimilar the new structures you want to predict are from your training data. How many DFT calculations (training points) do you need to parametrize a reliable NN model? These are pretty difficult to deal with since all machine learning models are smooth functions and struggle to learn large jumps (as humans sometimes do trying to rationalize SAR). Some other limiting factors apart from dataset size are the quality of the data and the sensitivity to small structural changes (activity cliffs). These methods give us a way to overcome the typically limited amount of affinity data we have for a particular target by bringing in more information. Sadly I don’t think we have gotten much better at activity prediction in the last few years, but neural networks also let us do interesting things in QSAR space such as large-scale multitask learning or even federated learning, and I think these approaches will be the standard in future. If deep learning methods beat canonical QSAR approaches depends on who you ask, but in my experience, one is almost never worse off with ChemProp instead of a fingerprint method (though one might not be as much better off as one hopes). QSAR methods in particular have benefited from a lot of optimization and seem to extract almost all the useful predictive power out of affinity data. So we have been doing “machine learning in chemistry” for decades, the only difference now is that we have a larger toolbox of models that might or might not help build better activity models. How accurate can the latest machine learning methods be in predicting the possible starting synthons for a designed novel molecule or polymer material? Neural networks also let us do interesting things in QSAR space such as large-scale multitask learning or even federated learning, and I think these approaches will be the standard in the future. Neural networks also let us do interesting things in QSAR space such as multitask learning or even federated learning, and I think these approaches will be the standard in the future.Ĭan we simulate the effect of magnetized water clusterization through machine learning tools? In these cases, they are not really competing with other methods as much as expanding the type of data we can use and using it in new ways. that don’t fit neatly into the QSAR label. ![]() But I don’t think that is the whole story – new machine learning methods are letting us solve new types of problems such as generative models, retrosynthesis prediction, massive multitask predictions, etc. Many QSAR techniques such as SVR/SVM and random forest are components of traditional/shallow machine learning, so we have been doing “machine learning in chemistry” for decades and in my experience, traditional QSAR methods are mostly competitive with deep learning approaches for affinity prediction especially. I would say “QSAR modeling” is a label for a specific type of machine learning (activity prediction based on molecular structure). ![]()
0 Comments
Leave a Reply. |
Details
AuthorWrite something about yourself. No need to be fancy, just an overview. ArchivesCategories |