CASCADE Predicts Gene Perturbation Effects with Patient Data.

Researchers have developed CASCADE, an agentic framework designed to predict the downstream transcriptional effects of gene perturbations. Unlike prior validation methods that merely check if predicted genes are known cancer genes, CASCADE rigorously tests whether the predicted direction of change aligns with real-world patient tumor data. It uses focal-gene copy-number amplification as a proxy for gene knockdown effects, leveraging data from TCGA patients. For genes like MYC, CASCADE demonstrated strong concordance between predicted knockdown targets and actual amplified-vs-non-amplified tumor expression across multiple cancer types, significantly outperforming permutation baselines. While its accuracy did not exceed existing public knowledge for MYC or E2F-driven biology when compared against curated gene-set baselines, its ability to predict gene-specific directions of change clearly surpassed a naive uniform guess. However, extending the validation to fifteen other genes revealed that predictive success is gene-specific. Proliferation-machinery regulators generally replicated well, whereas lineage-identity transcription factors and a specific cyclin-D paralog consistently failed. Separately, an LLM-based agent for grounding natural-language requests into CASCADE's tool calls achieved moderate exact match accuracy, with challenges in resolving ambiguous perturbation types.