DMG-32. Data-driven in silico avatar models of DMG for guiding clinical patient management

Abstract Data-driven personalization holds promise for hard-to-treat pediatric brain tumors including diffuse midline glioma (DMG) through in silico drug screening. We provide an AI-based drug response predictions platform incorporating data from DMG patients and several preclinical models. We paired newly generated and existing data from a large cohort of DMGs including DMG cells (n = 203), and post-mortem patient specimen (n = 59 primary, n = 28 metastatic from 47 DMG patients). Specifically, we generated bulk RNA and whole genome of 45 pediatric brain cancer cell lines (DMG, HGG), followed by in vitro drug screening with 119 clinically relevant therapeutics. Data were contextualized alongside patient tumor samples and public cell line data from the Childhood Cancer Model Atlas (CCMA; 158 cell lines, 1,654 drugs). Data-driven embeddings from bulk RNA-seq data informed our analysis. We benchmarked state-of-the-art drug-response-prediction pipelines for transcriptomic and drug structural embeddings building on deep learning foundation models, graph neural networks, geneset variation analysis, or highly-interpretable K-neared neighbour (k-NN) matching of molecular twins. Live-cell imaging for 14 days (Precomb, 3DTwin(R) technology) analyzed longitudinal spheroid responses in a subset of samples to capture drug efficacy and dynamics. DMG cell lines mirrored patients’ molecular profiles indicating their high fidelity, creating a DMG atlas across multi-institutional data. Models trained on CCMA identified a Random Forest regressor across GSVA scores with SMILESVec embeddings, achieving a Pearson correlation of 0.66 (±0.03) on unseen samples. We validated these predictive models on our DMG cell lines. k-NN regression performed similarly on external data, offering an interpretable alternative via molecular twin matching. Live-cell imaging revealed disparities in drug-induced cell death and proliferation patterns, highlighting the importance of longitudinal response monitoring. Data-driven in silico predictions demonstrate a promising path for personalized treatments in DMG. Modern frameworks effectively harness retrospective data but longitudinal insights into treatment dynamics remain essential to enhance traditional drug screens.