TxPert: Using Multiple Knowledge Graphs for Prediction of Transcriptomic Perturbation Effects
Researchers have introduced TxPert, a novel latent-transfer-based deep learning method designed to predict transcriptomic perturbation effects with high accuracy. Published in Nature Biotechnology, this model addresses the prohibitive costs and limitations of exhaustively exploring genetic perturbations in wet labs. TxPert utilizes multiple knowledge graphs encoding gene-product relationships, combining data from biological databases and high-throughput screens to capture complementary biological information. The model demonstrates state-of-the-art performance, achieving prediction accuracy for unseen single perturbations that approaches split-half experimental reproducibility. Furthermore, it significantly outperforms existing methods by increasing Pearson correlation for double unseen perturbations and cross-cell line predictions by 8–25%. By enabling reliable in silico simulations in out-of-distribution settings, TxPert facilitates more targeted confirmatory screening, potentially accelerating drug discovery and reducing reliance on extensive exploratory laboratory experiments. This advancement represents a significant step in applying machine learning to understand disease mechanisms and design effective therapeutic interventions through precise modeling of cellular responses.
Wire timeline
TxPert: Using Multiple Knowledge Graphs for Prediction of Transcriptomic Perturbation Effects
Researchers have introduced TxPert, a novel latent-transfer-based deep learning method designed to predict transcriptomic perturbation effects with high accuracy. Published in Nature Biotechnology, this model addresses the prohibitive costs and limitations of exhaustively exploring genetic perturbations in wet labs. TxPert utilizes multiple knowledge graphs encoding gene-product relationships, combining data from biological databases and high-throughput screens to capture complementary biological information. The model demonstrates state-of-the-art performance, achieving prediction accuracy for unseen single perturbations that approaches split-half experimental reproducibility. Furthermore, it significantly outperforms existing methods by increasing Pearson correlation for double unseen perturbations and cross-cell line predictions by 8–25%. By enabling reliable in silico simulations in out-of-distribution settings, TxPert facilitates more targeted confirmatory screening, potentially accelerating drug discovery and reducing reliance on extensive exploratory laboratory experiments. This advancement represents a significant step in applying machine learning to understand disease mechanisms and design effective therapeutic interventions through precise modeling of cellular responses.
Nature Biotechnology