(284g) Structure-Guided Functional Phylogenetic Trees of Human Channel Proteins and Kinases for Drug Discovery

Kinases mediate a large number of cellular processes (such as metabolism, proliferation, and apoptosis) via catalyzing protein phosphorylation. Kinases have thus emerged as promising drug targets for oncology. However, the large size of the kinome (~500 human kinases) and influence of activation states upon drug binding makes it difficult to discern their cellu­lar modes of action. Clustering analyses across the human kinome (518 kinases) are limited to clustering them based only on sequence similarity - from a multiple sequence align­ment (MSA). ML-based structure predictors like AlphaFold2 have sup­plemented experimental PDBs and thus allows us to access 3D-structures of the entire human kinome including the less understood dark kinome. Functionality of kinases is majorly guided by structural similarity and the presence of similar binding pockets for drug accessibility. We introduce a consolidated CLASSIF (Clustering Annotated Sequence, Structure, and Information of Function) workflow to: (a) predict/use kinase PDB structures, (b) annotate kinase domain geometries, (c) identify ATP binding domains (d) align ATP binding pockets in structure as well as latent space, (e) integrate experimental drug binding affinity with sequence and structural data, and (f) construct phylogenetic trees based on these sequence-structure-function metrics. We provide a computational framework to discern the biophysical bases of drug-kinome interactions at the atomic level and extend this on to dark kinases and novel drugs alike.

While kinases are shorter proteins (<400aa) and have nearly 75% of the kinome experimentally resolved, voltage-gated ion channels constitute a far more challenging set. These channels are often ~1500 aa long single-chain proteins, with barely 5% structures resolved experimentally. These neuron-bound surface proteins are therapeutic targets for pain medications. We challenge our deep-learning-based structure-prediction tool RGN2 to first predict the structures of all ~130 human voltage-gated ion channels. Preliminary results after passing these channel structures through the CLASSIF workflow have been able to recover conformational binding modalities of known pain killer drugs like lidocaine.