Banca de QUALIFICAÇÃO: DANIEL HENRIQUE FERREIRA GOMES

Uma banca de QUALIFICAÇÃO de DOUTORADO foi cadastrada pelo programa.
STUDENT : DANIEL HENRIQUE FERREIRA GOMES
DATE: 28/05/2026
TIME: 09:00
LOCAL: Google Meet
TITLE:

Integration of germline and somatic variants with machine learning models for the classification of gastric cancer in young patients

KEY WORDS:

Early-onset gastric cancer; Machine learning; Germline variants; Somatic variants; Exome sequencing

PAGES: 54
BIG AREA: Ciências Biológicas
AREA: Biologia Geral
SUMMARY:

Early-onset gastric adenocarcinoma (EOGC) represents a clinically and genetically relevant subgroup, characterized by predominantly diffuse histology, aggressive behavior, and the frequent absence of classical environmental risk factors, indicating a strong influence of hereditary predisposition. In Brazil, the North and Northeast regions concentrate the highest gastric cancer mortality rates in populations historically underrepresented in major international genomic consortia, which hampers the identification of population-specific predisposing variants. This study proposed the integration of germline and somatic variants, obtained through whole exome sequencing, with machine learning models for the classification of early-onset gastric adenocarcinoma in Brazilian populations. The cohort comprised 375 individuals, including 232 cases — 95 from the state of Pará, 13 from Rio Grande do Norte, and 124 from Korean public databases — and 143 controls without cancer history. Germline variants were identified using DeepVariant and annotated through a customized pipeline integrating functional, clinical, and population databases. High functional impact germline variants were defined using DTreePred as the primary pathogenicity predictor, replacing MetaSVM. Fifteen classifiers were evaluated under three feature representation strategies: aggregation by functional gene lists, z-score normalization, and binary encoding of individual variants. For somatic analysis, a regionally representative Panel of Normals was constructed and GATK-Mutect2 was applied in a tumor-only approach; annotation and filtering steps are currently ongoing. Models based on individual variants achieved an accuracy of 0.97 and an AUC of 1.00 for Random Forest and ExtraTrees, while the gene list approach favored the Bagging model with an accuracy of 0.96 and an AUC of 0.99. The results demonstrate that machine learning models trained on high functional impact germline variants can discriminate EOGC patients from healthy controls with high accuracy in underrepresented Brazilian populations, providing a replicable computational basis for the molecular diagnosis of hereditary gastric cancer predisposition.

COMMITTEE MEMBERS:
Presidente - 2170415 - Jorge Estefano de Santana Souza
Interna - 1365498 - BEATRIZ STRANSKY FERREIRA
Interno - 3884005 - PATRICK CESAR ALVES TERREMATTE