Diagnosis of thyroid neoplasm using support vector machine algorithms based on platelet RNA-seq

Objective To assess the capacity of support vector machine (SVM) algorithms that are developed based on platelet RNA-seq data in identifying thyroid neoplasm patients and differentiating patients with thyroid adenomas, papillary thyroid cancer and metastasized papillary thyroid cancer. Methods Platelets were collected and isolated from 109 patients and 63 healthy controls. RNA-seq was performed to find transcripts with differential levels. Genes corresponding to these altered transcripts were identified using R packages. All samples were subsampled into a training set and a validation set. Two SVM algorithms were developed and trained with the training set, using the genes with differential transcript levels (GDTLs) as classifiers, and validated with the validation set. GO and KEGG pathway enrichment analysis were performed using the R package clusterProfiler. Results We detected 765 GDTLs (442 up-regulated and 323 down-regulated) in platelets of patients and healthy controls. The algorithm identifying thyroid neoplasm patients achieved an accuracy of 97%, with an AUC (area under curve) of 0.998. The other algorithm differentiating patients with multiclass thyroid neoplasms had an average accuracy of 80.5%. GO analysis showed that GDTLs were strongly involved in biological processes such as neutrophil degranulation, neutrophil activation, autophagy and regulation of multi-organism process. KEGG pathway enrichment analysis revealed that GDTLs were mainly enriched in NOD-like receptor signaling pathway and pathways in endocytosis, osteoclast differentiation, human cytomegalovirus infection and tuberculosis. Conclusion Our results indicated that the combination of SVM algorithms and platelet RNA-seq data allowed for thyroid neoplasm diagnostics and multiclass thyroid neoplasm classification.