You’ve probably lived this story before: you’re working on your favorite kinase, YFK1, and are looking for selective inhibitors over YFK2.  The problem is, they’re virtually identical near the active site, save for a single residue sitting in the back pocket that your amazing structural biologist has identified.  The team thinks you have a better chance of getting selectivity for YFK1 over YFK2 with kinase inhibitor Types I½ or II, which both bind to the back pocket (Figure 1).  Is there any way to pick out the Type I½ or Type II kinase inhibitors out of the 250,000 hinge-binding compounds in your kinase-focused library?  This task sounds like a perfect application for machine learning, which a team from Boehringer Ingelheim and the Friedrich-Wilhelm-Universität recently reduced to practice.1

Type I Kinase Inhibitors with Binding Mode, Type II Kinase Inhibitor with Binding Mode, and Type I 1/2 Kinase Inhibitor with Binding Mode
Figure 1. There are three common classes of ATP-competitive kinase inhibitors: Type I inhibitors such as ceritinib bind to the hinge region in the kinase’s active, DFG-in conformation, and don’t access regions in front of the gatekeeper or the back pocket, usually due to a large gatekeeper (e.g. F, L). Type I½ inhibitors such as vemurafenib bind to the hinge and gate regions in the active, DFG-in conformation, allowed by smaller gatekeepers (e.g. T, S). Type II inhibitors such as imatinib bind to the hinge region and back pocket in the kinase’s inactive DFG-out conformation. Images adapted from Ref. 22

The authors first assembled a dataset of 1425 Type I, 394 Type I½, and 190 Type II inhibitors with binding modes confirmed by X-ray co-crystal structure in the PDB.3  They divided the inhibitors into evenly sized training and test sets, and then tested several machine learning algorithms to see whether an algorithm could distinguish the three inhibitor classes in the test sets after being trained on the training sets. The team ultimately found a model that was able to distinguish the three classes of inhibitors with remarkably high sensitivity and specificity while starting from an impressively small training set (Figure 2).  From this proof of concept, one can imagine a future where libraries might be more computationally enriched before screening (think “focused focused” libraries), leading to smaller screens and higher quality hits.

Figure 2. Receiver Operating Characteristics (ROC) curves showing that the authors’ ECFP4 fingerprint model distinguished Type I from Type II inhibitors and allosterics (A) from non-allosterics (I+I½+II), with relatively low error rates regardless of the algorithm used. Area-under-the-curves (AUC’s) for random forest (RF), support vector machine (SVM), or deep neural network (DNN) shown in the legends. A perfect model which classifies 100% of Type I inhibitors as Type I and 0% of Type II inhibitors as Type I has an AUC of 1.0. Image reproduced from Ref. 1.

Take a look at the following 8 kinase inhibitors, and try to classify them by their kinase inhibitor type (I or II, answers in the caption below).  How long did it take you?  Now imagine doing it 100,000 times to curate a library!  This illustrates how much time computational tools like machine-learning could save scientists in the future.

Figure 3. 8 Randomly selective kinase inhibitors drawn in unsuggestive formats by binding mode, with their binding modes and PDB’s: A) Nintedanib, Type II, 3C7Q B) Quizartinib, Type II, 4XUF C) Entrectinib, Type I, 5FTO D) Alectinib, Type I, 3AOX E) Tofacitinib, Type I, 3EYG F) Nilotinib, Type II, 3CS9 G) Sorafenib, Type II, 4ASD H) Dasatinib, Type I, 2GQG. See Ref. 2.

Props to the authors for putting together a highly useful dataset, developing an excellent classification algorithm, and for making both their dataset and algorithm freely available to the community.

Explore drughunter.com for more.

About the Guest Author – Cyril Bucher

Cyril hails from a little town in Switzerland and moved to the Big Apple to play college tennis. Having gotten a taste for synthesis in the Leighton lab at Columbia University, he decided to dedicate his graduate career to the assembly of marine natural products in the lab of Noah Burns at Stanford. He’s since become an enthusiastic drug hunter who is fascinated by the opportunities presented by the application of machine learning to the drug discovery process.

  1. Miljković, F., Rodríguez-Pérez, R., Bajorath, J. “Machine Learning Models for Accurate Prediction of Kinase Inhibiors with Different Binding Modes.” J. Med. Chem. 2019. doi: 10.1021/acs.jmedchem.9b00867
  2. For an excellent recent review on classifications of kinase inhibitors, see: Roskoski, R., “Classification of small molecule protein kinase inhibitors based upon the structures of their drug-enzyme complexes.” Pharm. Res. 2016, 103, 26-48. doi: 10.1016/j.phrs.2015.10.021
  3. Interestingly, they found that 3.6% of these inhibitors adopted multiple binding modes against different kinases, a feature associated with inhibitor promiscuity.