LIU, H.-K. | Search bioRxiv

bioRxiv2025

DELPHAI, AI Agent for Predicting Drug Response and Resistance

Patient-derived organoids preserve critical tumor features and drug sensitivity patterns that mirror patient clinical responses, enabling single-cell RNA sequencing analysis of drug responses. Analyzing these perturbation data presents significant computational challenges in predicting cellular responses while maintaining biological interpretability. We developed DELPHAI (Deep ExplainabLe Predictive Human-organoid based AI), an AI agent that integrates single-cell perturbation prediction with mechanistic analysis using large language models. We designed a comprehensive benchmarking framework evaluating methods in both computational embedding and reconstructed gene expression spaces. Applied to glioblastoma organoids treated with temozolomide, optimal transport combined with principal component analysis outperformed baseline methods in capturing population dynamics. DELPHAI correctly identified DNA alkylation as the mechanism of action without prior drug knowledge and recommended combination therapies aligning with clinical trials. These results demonstrate DELPHAPs ability to translate single-cell perturbation data into actionable therapeutic insights, representing a significant advance toward Al-driven precision medicine in cancer treatment.

bioinformatics↗

bioRxiv2023

A multidimensional atlas of human glioblastoma organoids reveals highly coordinated molecular networks and effective drugs

Recent advances in the genomics of glioblastoma (GBM) led to the introduction of molecular neuropathology but failed to translate into treatment improvement. This is largely attributed to the genetic and phenotypic heterogeneity of GBM, which are considered the major obstacle to GBM therapy. Here, we use advanced human GBM organoid (LEGO: Laboratory Engineered Glioblastoma Organoid) and provide an unprecedented comprehensive characterization of LEGO models using single-cell transcriptome, DNA methylome, metabolome, lipidome, proteome, and phospho-proteome analysis. We discovered that genetic heterogeneity dictates functional heterogeneity across molecular layers and demonstrates that NF1 mutation drives mesenchymal signature. Most importantly, we found that glycerol lipid reprogramming is a hallmark of GBM, and several targets and drugs were discovered along this line. We also provide a genotype-based drug reference map using LEGO-based drug screen. This study provides novel human GBM models and a research path toward effective GBM therapy.

cancer biology↗