Machine learning-powered integration of global proteomics and ex vivo sensitivity unveils a protein signature predictive of treatment success to AML therapy: Validation in patients treated with FHD-286, a SMARCA2/4 dual inhibitor Free

Acute myeloid leukemia (AML) has a 33% 5-year survival rate, and few new therapies have made a meaningful impact on the overall prognosis. A key obstacle to developing effective treatments is the lack of accurate model systems effective at capturing the clinical response of patients. To address this, we have generated quantitative proteomics and therapeutic response data from a diverse panel of primary AML samples and used machine learning to model disease biology and therapeutic actions. These novel models allow investigators to personalize therapies based on the protein biomarkers of each patient.

To develop our predictive models, we use deep learning to integrate global proteomics, clinical data and ex-vivo sensitivity to generate an AML therapeutic sensitivity landscape that can guide therapeutic positioning. We collected 200 patient peripheral blood samples and established >90 patient-derived cell lines that capture AML heterogeneity using proprietary culturing conditions. LC-MS proteomics quantified >9,000 proteins per sample and we conducted >5,000 assays with >50 investigational and marketed AML therapies. Proteomics runs were integrated using internal reference scaling methodology, correcting for sample loading errors and batch effects, while global proteomics outlier detection ensured data consistency. Our deep learning approach integrates proteomics and dose-response data to learn the molecular function enrichments across proteomics datasets. The model is designed to be generalizable, allowing it to capture the mechanisms of sensitivity of many AML therapies that may expand beyond known biology. To enhance training and capture known biological relationships between proteins under perturbed conditions, we applied transfer learning by pretraining the architecture on publicly available proteomics datasets across several diseases. Each public dataset was reprocessed to maintain consistent normalization and data handling parameters. 2906 samples from 14 NCI-CPTAC (National Cancer Institute Clinical Proteomic Tumor Analysis Consortium) datasets, 6 internal datasets, and 3 additional public datasets across 11 indications were used for pre-training. This significantly improved model convergence, reduced overfitting, lowered overall loss, and increased model interpretability. The model was then fine-tuned on 650 AML samples to create the AML-specific sensitivity landscape.

The result is a functionally embedded latent space that represents a landscape where patients are positioned based on both disease-relevant proteomic signatures and their sensitivity to >50 monotherapies. When presented with patient proteomes only, unsupervised clustering applied to the latent space identifies patient subpopulations that are predicted to have consistent dose-responses to ex vivo treatments and similar proteomes. Therefore, we used our model to find protein biomarkers that can identify patients more likely to respond to FHD-286, an investigational SMARCA2/4 dual inhibitor (Foghorn Therapeutics). The initial set of proteins included markers of cell motility, angiogenesis, cell differentiation, apoptosis, and inflammation, including fibroblast-specific proteins, B-cell receptor signaling proteins, and differentiation antigens. This was reduced to 5 proteins with robust expression across all 200 patients in our training data. The relative abundance of these 5 protein biomarkers in pretreatment peripheral blood samples was used to predict clinical response in 7 patients with AML treated with FHD-286 (NCT04891757). Patients had received a median 3 prior lines of therapy (range 1-7), including venetoclax and hypomethylating agents (100%), high-dose chemotherapy (71%), and stem cell transplant (57%). Genetic risk status (2022 ELN) was adverse in 71% of patients, including 3 patients each with TP53 mutation and inv(3). Notably, the 5-protein signature classified the 7 patients into treatment failure, stable disease, and complete response with 100% accuracy (p=4.6×10⁻⁴). More extensive patient sample validation will be presented at the meeting. A clinical trial prospectively assigning patients to specific treatment arms based on our proteomics signature will launch in 2025. This result highlights the utility of Yatiri Bio's deep learning model to guide AML treatment decisions and improve outcomes and supports the development of a mass spectrometry-based clinical assay for patient selection in clinical trials.