AI-derived constrained conditional model for screening marker genes through integrated high-throughput transcriptome big data

Currently, a large volume of high-throughput transcriptome data are accessible through various public databases. This brings urgent need to integrate transcriptome data sourced from numerous independent studies to extract common characteristics to precisely address specific research questions. Referring to this, we innovatively designed a model framework that leverages integrated transcriptome data with declarative constraints. Here the constraint is used to incorporate the AI-derived prior knowledge (genes in certain biological pathways, functions or user-specified terms) into the model and drive the model prediction/decision to satisfy these constraints. Distinguishing from existing models or methods, this framework implies tailed non-parametric bootstrapping algorithm to generate millions of samples of p-value of independent analyses to avoid the normal procedure of integrating data with distinct distribution styles and difference from users’ requirements. The model was applied in 5 tumor and 5 non-tumor case studies using 81 downloaded transcriptome datasets with 10,647 samples. High percentage of the selected genes were accordant with published work, co-occurrence results and gene ontology results. Experimental validations including transwell invasion/migration confirmed the identified genes associated with cancer progression in prostate, liver, and endometrial cancer case studies. Therefore our model was effective in extracting marker genes under user-specified conditions and were possessing significance in further understanding specified situations.