@TreehouseFi

#Treehouse and $TREE

The birth of the TREE model breaks through these limitations. The TREE model is a Transformer-based graph representation learning AI model that can handle both homogeneous and heterogeneous networks, where homogeneous networks contain only genes, while heterogeneous networks include various node types such as transcription factors (TF), miRNA, and IncRNA.

▲(a) Multi-omics data collection and homogeneous/heterogeneous network construction; (b) Overall model flowchart for cancer gene prediction; (c) Gene representation learning layer of the model; (d) Multi-channel integration module.

02

TREE's 'Superpowers'

Research shows that TREE exhibits superior performance on 8 biological pan-cancer networks and 31 cancer-specific networks. Compared to 5 network-based AI methods, TREE has the best AUC and AUPR metrics, with an average AUC improvement of 5.91% and AUPR improvement of 9.87%, demonstrating the model's generalization and robustness.

At the same time, TREE also performs excellently in terms of interpretability. Mutations are crucial in cancer gene identification, and TREE has advantages in precisely locating rare mutation genes; heterogeneous information allows TREE to verify significant cancer gene regulatory mechanisms through network pathways.

▲Venn diagram of cancer candidate genes identified by all methods

After scoring all common unlabeled genes in the network, TREE recommended 57 potential cancer candidate genes, suggesting they may be associated with cancer. Subsequently, researchers tested the model's performance using the entire dataset, and the results showed that the evaluation results provided by the model were stable and consistent, indicating that TREE is a reliable tool for identifying new cancer candidate genes.