Novel Machine Learning Approaches Tailored to Protein Configuration Predict Drug Vulnerabilities in Myeloid Neoplasia

Therapy of high risk myeloid neoplasia (MN) may involve not only rationally targeted approaches but also repurposed drugs which may have direct or off target activity in a certain genetic context. One of the crucial needs in drug prediction is the rational incorporation of drug response in relation to the impact on protein level. For instance, KRAS-G12C inhibitors were successful in a strictly targeted fashion and similar restriction may apply to many drugs based on individual circumstances including actual targets or presence of comutated genes. Fulfilling the need of clinically applicable drug selection algorithms, futile treatments could be avoided and decision of therapeutic interventions could be addressed more rationally. While bioinformatic approaches may allow for sensitivity prediction without in vitro high throughput drug screens, to generate bioanalytic predictive pipelines, actual in vitro screen results need to be incorporated in training and confirmatory stages.

We hypothesize that vulnerabilities to pharmacologic agents are imprinted in genomic features and that only rational stratification of patients according to their molecular make up (amino acid changes, mutations in functional domains, protein effect) might inform unsuspected sensitivities not otherwise recognized using a basket purely molecular approach.

Thus, we used a computational model to personalize drug response prediction by integrating genomic features stratified by functional effects on protein sequence. We curated a compendium of molecular data of 3,588 MN (22/58% MDS high/low risk; 20% secondary AML) and 452 primary AML. Data collected from The BEAT AML1 were also curated based on the scope of our analysis. Targeted NGS of 40 myeloid genes was uniformly applied. Variants were curated according to their functional impact due to their topology within the gene. For instance, TP53 mutations were divided based on the occurrence in the DNA binding domain and whether they were monoallelic vs biallelic, TET2 mutations were categorized based on their impact on the catalytic vs non catalytic domain, frameshift/stop codon/splice site vs missense, RUNX1 mutations based on whether or not they localized in the RUNT domain, SF3B1 mutations in K700E vs other common hotspots (R625, H662, K666), DNMT3A mutations in R882 vs non R882 vs frameshift/stop codon/splice site. Molecular and drug screening profile (79 FDA approved drugs) from The BEAT AML was used as discovery cohort to generate the model. Once modelling algorithm was assembled, an in-house in vitro drug/improved mutational dataset was generated for validation.

Given the success of deep representation learning approaches, we trained an encoder-decoder model to generate a low dimensional representation of genomic features. Using the pretrained encoder, we trained a downstream feed-forward neural network via the observations acquired from the discovery dataset. Given the trained model, we quantified genotype-drug response associations by Pearson correlation and compared the predictors with the in house generated dataset.

Out of 79 drugs, we found at least one drug showing sensitivity patterns against each unique protein configuration or mutational combination. The examples of specific observations illustrate the personalized approach. For instance, RUNX1 frameshifts shared a sensitivity profile with EZH2 frameshifts but not missense and SET domain. FLT3 missense, IDH1-R132H, TP53 splice sites, WT1 and CEBPA frameshifts exhibited divergent sensitivity patterns. Only RUNX1 frameshifts (but not others) showed sensitivity for MEK1/2 inhibitors except for cases with comutated TP53 splice site rendering them resistant. In contrast RUNX1 RUNT domain showed unique heightened response to CDK9 inhibitors. FLT3-TKD but not ITD showed higher sensitivity to cabozantinib while NRAS-Q61 to dactolisib (PI3K/mTOR dual inhibitor). Our trained algorithm integrated this seemingly unintelligible complexity of individual genotype/phenotype responses to overcome our inability to rationally predict sensitivity/resistance responses. This algorithm was then used in our ongoing confirmatory drug screen collection that mimics real life application in a clinical setting.

In summary, incorporation of protein configuration in drug response prediction might help unveiling unsuspected vulnerability profiles in MN addicted to specific gene mutations.