MET kinase inhibitor types and resistance mutations screened against a nearly comprehensive library of kinase domain substitutions.

(A) Crystal structure of the ATP-bound MET kinase domain (3DKC) overlaid with type Ia (crizotinib, 2WGJ), type Ib (savolitinib, 6SDE), type II (merestinib, 4EEV), type I½ (AMG-458, 5T3Q), and type III inhibitors (tivantinib, 3RHK). (B) Pocket view of ATP and each inhibitor type bound to the active site of the MET kinase domain with the respective inhibitor and crystal structures from panel A. (C) 2D chemical structures of each inhibitor screened against the site saturation mutagenesis library of the MET kinase domain, with each experimentally determined IC50 values displayed for Ba/F3 cells stably expressing the wild-type MET ICD in a TPR-fusion background. (D) Dose-response curves for each inhibitor against the wild-type MET intracellular domain expressed in a TPR-fusion in the Ba/F3 cell line. (E) Schematics of the full-length and exon 14 skipped MET receptor alongside the TPR-fusion constructs with the full-length and exon 14 skipped intracellular domain, displaying four mechanisms of oncogenic activity: point mutations, exon 14 skipping, constitutive activity through domain fusions, and inhibitor resistance mutations. (F) Experimental workflow for defining the mutational landscape of the wild-type TPR-MET and exon 14 skipped TPR-METΔEx14 intracellular domain against 11 ATP-competitive inhibitors in Ba/F3, interleukin-3 (IL-3) withdrawn pooled competition assay.

Structural inhibitor classification and dose-response determination.

(A) Type I (crizotinib, 2WGJ; tepotinib, 4R1V; capmatinib; savolitinib, 6SDE; NVP-BVU972, 3QTI) and type II (merestinib, 4EEV; cabozantinib; glesatinib) inhibitor-bound MET kinase domain structures globally aligned. Hinge (gray) and G1163 (represented as a sphere) are highlighted to show the kinase domain solvent-front relative to each inhibitor. Inhibitors lacking experimental structures (capmatinib, cabozantinib, glumetinib, and glesatinib) were docked onto a representative type I (PDB 2WGJ) and type II (4EEV) structure through AutoDock Vina (Eberheart et al., 2021; Trott et al., 2010). (B) Solvent-front and G1163 highlighted relative to the ATP-bound kinase domain crystal structure (3DKC) and all inhibitors screened. (C) Dose-response curves for each inhibitor against the TPR-fusion MET and METΔEx14 intracellular domains, stably expressed in Ba/F3 cells.

Correlation analysis of the MET kinase domain site saturation mutagenesis library across replicates and conditions.

Replicate correlation analysis for each inhibitor for both the TPR-fusion MET background scores with Enrich2.

Correlation analysis of the METΔEx14 kinase domain site saturation mutagenesis library across replicates and conditions.

Replicate correlation analysis for each inhibitor for the TPR-fusion METΔEx14 background score with Enrich2.

Fitness landscapes of the MET kinase domain against a panel of 11 inhibitors.

Heatmap for the DMSO control condition and all inhibitor fitness scores from Rosace, subtracted from DMSO for >99% of MET kinase domain variants in the full intracellular domain background in the context of the TPR-fusion.

Mutational landscape of the MET kinase domain under 11 ATP-competitive inhibitor selection.

(A) Distributions of all variants (wild-type synonymous, early stop, and missense) for each condition in the wild-type TPR-MET kinase domain, scored with Rosace and normalized to the growth rate of the DMSO control population. (B) Correlation plots for all mutational fitness scores for each drug against DMSO, fitted with a linear regression and Pearson’s R value displayed. (C) Heatmap showing the Pearson’s R correlation for each condition against each other, annotated by condition and inhibitor type. Correlations are colored according to a scale bar from gray to blue (low to high correlation). (D) Crystal structure of the tivantinib-bound MET kinase domain (PDB 3RHK) overlaid with the ATP-bound kinase domain (PDB 3DKC), with tivantinib-stabilizing residues and overlapping density of tivantinib (orange) and ATP (purple) highlighted. (E) Dose responses of crizotinib and tivantinib tested against stable Ba/F3 cells expressing the wild-type intracellular domain of MET fused to TPR, tested in the presence and absence of interleukin-3 (IL-3).

Mutational landscape of the METΔEx14 kinase domain under 11 ATP-competitive inhibitor selection.

(A) Distributions of all variants (wild-type synonymous, early stop, and missense) for each condition, scored with Rosace and normalized to the growth rate of the DMSO control population. (B) Correlation plots for all mutational fitness scores for each drug against DMSO, fitted with a linear regression and Pearson’s R value displayed. (C) Heatmap showing the Pearson’s R correlation for each condition against each other, annotated by condition and inhibitor type. Correlations are colored according to a scale bar from gray to blue (low to high correlation).

Novel resistance mutations identified and mapped for crizotinib.

(A) Heatmap of crizotinib fitness scores subtracted from DMSO, scaled from loss-of-function (red) to gain-of-function (blue), with the wild-type protein sequence, secondary structure, kinase domain residue position, and mutational substitution annotated. Wild-type synonymous substitutions are outlined in green. (B) Resistance positions mapped onto the crizotinib-bound, MET crystal structure (PDB 2WGJ). Positions are labeled and colors are scaled according to the average score for the resistance mutations at each site. (C) 2D protein-drug interactions between crizotinib and the MET kinase domain (PDB 2WGJ) with pocket residues and polar and pi interactions annotated. Schematic generated through PoseEdit (Diedrich et al., 2023) (https://proteins.plus/). (D) Condensed crizotinib heatmap displaying direct drug-protein interacting and non-direct resistance position. (E) Crizotinib binding site and pocket residues displayed with resistance positions highlighted (pink) and the wild-type residue and inhibitor interactions shown (PDB 2WGJ). (F) Resistance mutations modeled for I1084H, V1092I, Y1159R, M1211Y, and N1167 relative to ATP (PDB 3DKC).

Resistance mutations mapped onto experimental and docked kinase domain structures for type I and type II inhibitors.

(A-F) Resistance positions and average resistance mutational score mapped onto representative crystal structures (tepotinib, 4R1V; merestinib, 4EEV) and labeled (type I, pink; type II, blue). Inhibitors lacking experimental structures (capmatinib, cabozantinib, glumetinib, and glesatinib) were docked onto a representative type I (PDB 2WGJ) and type II (4EEV) structure through AutoDock Vina (Eberheart et al., 2021; Trott et al., 2010). (G) Heatmaps of each resistance position structurally modeled for each drug with fitness scores scaled from loss-of-function (red) to gain-of-function (blue). Wild-type synonymous substitutions are outlined in green.

Novel resistance mutations and “hotspots” identified for MET inhibitor types.

(A) Collapsed heatmap of common resistance positions along the kinase domain, with the wild-type protein sequence and secondary structure annotated. Each tile represents a sum of counts for statistically filtered resistance mutations across all inhibitors for type I (pink), type II (blue), and the type I½ inhibitor AMG-458 (green), with the scale bar reflecting counts of resistance mutations across respective inhibitor types. (B-D) Expanded heatmap showing each resistance position and the counts for each specific resistance mutation across all inhibitor types type I (pink), type II (blue), and the type I½ inhibitor AMG-458 (green). Wild-type sequence and variant change are annotated. (E-F) Average frequency of resistance mutations for each mapped on to a representative type I (crizotinib, 2WGJ) and type II (merestinib, 4EEV) crystal structure, alongside the type I½, AMG-458 structure (5T3Q), with associated scale bars. Individual positions with high resistance mutation frequencies are annotated on each structure, with a zoom-in of the bound inhibitor and surrounding resistance sites. (G) Venn diagram showing mutations shared among type I (pink), type II (light blue), and type I½ (green). (H) Structurally mapped (PDB 2WGJ) resistance positions shared among type I, II, I½ (blue-gray), type I and II (purple), type I and I½ (dusty rose), type II and I½ (teal) inhibitors.

Statistically filtered resistance mutations for grouped type I, type II, and type I½ inhibitors for MET.

(A-C) Heatmaps of the sum of resistance mutations grouped for type I (pink), type II (blue), and type I½ (green) for MET.

MET kinase domain differential sensitivities revealed for type I and type II inhibitors.

(A) Heatmap showing Pearson correlation values for all combinations of screened type I and type II inhibitors. Correlations were determined from DMSO subtracted fitness scores (B) Correlation plot correlation plot of DMSO subtracted fitness scores for crizotinib and cabozantinib. Mutations with differential scores are highlighted for type I (pink) and type II (blue). (C) Average scores of mutations with differential sensitivities within inhibitor pairs mapped and annotated in respective crystal structures (crizotinib, 2WGJ; cabozantinib, docked into 4EEV). Positions that are gain-of-function for type I but loss-of-function in type II are highlighted in pink, whereas positions that are gain-of-function for type II but loss-of-function in type I are highlighted in blue. (D) Dose-response curves for crizotinib and cabozantinib in Ba/F3 cells expressing TPR-MET (full MET intracellular domain) harboring mutations at Y1093K and L1195M.

Cross-comparison of type I and type II inhibitor pairs.

Scatter plots of each type II inhibitor fitness scores (cabozantinib, glesatinib analog, merestinib; axis in blue) against each type I inhibitor (crizotinib, capmatinib, tepotinib, glumetinib, savolitinib, NVP-BVU972; axis in pink).

Cross-comparison analysis of inhibitors within the same type.

Scatter plots of each inhibitor pair within the type II group (cabozantinib, glesatinib analog, merestinib; axes in blue) and within the type I group (crizotinib, capmatinib, tepotinib, glumetinib, savolitinib, NVP-BVU972; axes in pink).

Inhibitor-bound variant fitness predicted from a machine learning model trained on the MET DMS dataset.

(A) Model architecture outlining the information flow and inputs for model training, validation, fitness predictions, and prediction tests. (B) Improvement in correlation between experimental and predicted fitness for each inhibitor with usage of different kinds of features. (C) Cross-validation trends between the baseline ESM model and the model with all features incorporated. (D) Scatter plots of predictions versus experimental fitness scores of the baseline ESM model (top) compared to the model with all features (bottom), with a dashed cross-graph line in red displayed. (E) Residue-level analysis of feature significance in fitness predictions (ESM, stability, distance, conformation, all features). The Rosace experimental score is shown as a red line. (F) Residues with improved predictions mapped on a crizotinib-bound MET kinase domain (PDB 2WGJ). Predicted resistance mutations (dark purple) modeled relative to the wild-type residue (pink).

Distribution and visualization of features used in the XGBoost machine learning models.

(A) Distribution of ESM LLR vs. Experimental fitness (top) and ΔΔG vs. Experimental fitness (bottom). (B) Distribution of all features (except ESM LLR and ΔΔG) extracted and used for the XGBoost models. The features that were incorporated in the best performing model are shown in yellow. The red dashed lines within each distribution show the edges of bins used to bin the feature values. (C) ΔΔΔG calculated from predicted ΔΔG Type I (PDB 2WGJ) (left) and Type II (PDB 4EEV) (right) MET kinase structure by subtracting type II ΔΔG from type I ΔΔG. The key regions showing difference in conformation between type I and II structures are the DFG motif (purple) and aC helix (teal). (D) Calculation of “residue to ATP” distance feature for residue D1228 in ATP bound MET Kinase structure (3DKC) is shown. Centroid of the ATP molecule is shown as a pink sphere. (E) Example of ΔVolume feature calculation using the difference between the volume of Asp and Cys. (F) Ensemble of MET kinase domain crystal structures aligned and RMSF of a given residue (D1228 in this example). (G) The shortest distance between the inhibitor and a mutation calculated from the Umol predicted variant-inhibitor structure. (H) The binding pocket of crizotinib in the predicted Umol structure. Pocket volume, hydrophobicity score, polarity score and RF score are calculated from this binding site. (I) Residue RMSD feature is described by the Umol predicted structure of variant D1228C (pink) superposed onto the wild-type reference structure (PDB 2WGJ, gray) and RMSD between D1228 in the wild-type structure and 1228C in the variant structure. (J) Ligand RMSD feature The Umol predicted structure of variant D1228C (pink) superposed onto the wild-type reference structure (PDB 2WGJ, gray) and RMSD between crizotinib in the wild-type structure (pink) and in the variant structure (blue).