Ridho, Felza (2026) Identifikasi Biomarker pada Klasifikasi HIV-1 berdasarkan Respons ART menggunakan Multimodal Metaheuristik Knowledge Graph Neural Networks. Doctoral thesis, Institut Teknologi Sepuluh Nopember.
![[thumbnail of 7002231003_Doctoral.pdf]](http://repository.its.ac.id/style/images/fileicons/text.png) |
Text
7002231003_Doctoral.pdf - Accepted Version Restricted to Repository staff only Download (18MB) | Request a copy |
Abstract
Human Immunodeficiency Virus tipe 1 (HIV‑1) masih menjadi beban kesehatan global dengan respons terhadap Antiretroviral Therapy (ART) yang sangat bervariasi, sehingga menuntut pendekatan personal berbasis biomarker. Namun, metode machine learning konvensional belum optimal dalam menangkap interaksi gen nonlinier serta struktur relasional antarpasien. Penelitian ini mengembangkan BioFL‑GNN, sebuah kerangka Graph Neural Networks (GNN) yang mengintegrasikan knowledge graph, seleksi fitur metaheuristik, dan Fast Local Convolution (FLC) untuk klasifikasi respons ART serta penemuan biomarker. Sebagai pembanding, digunakan pendekatan OCT‑GCN pada data klinis dan ConL‑HIV berbasis Transformer pada data genomik yang telah diperluas. Data ekspresi gen microarray dan data klinis diintegrasikan melalui penyelarasan pasien dan konstruksi fitur diferensial. Knowledge graph dibangun menggunakan 11 metrik jarak, termasuk Bregman divergence dan angular distance. Enam algoritma metaheuristik (Harmony Search, Ant Colony Optimization, Binary Grey Wolf Optimizer, Binary Cuckoo Search, Genetic Algorithm, dan Particle Swarm Optimization) diterapkan untuk menyeleksi gen yang informatif. Empat arsitektur GNN (GCN, GAT, GIN, GraphSAGE) yang diperkuat FLC dilatih untuk mengklasifikasikan pasien resisten dan non‑resisten. Evaluasi dilakukan pada data genomik utama (138 pasien), data klinis (1.056 rekaman), dan data genomik diperluas (406 sampel) dengan validasi eksternal. Hasil menunjukkan bahwa BioFL‑GNN mencapai akurasi 96,43% pada data genomik utama dan 98,78% pada data yang diperluas, dengan recall sempurna. Integrasi FLC meningkatkan akurasi rata‑rata 4,7%. Strategi konsensus multi‑model mengidentifikasi 26 high‑confidence biomarkers yang terlibat dalam regulasi pH, pensinyalan lipid, apoptosis, dan restriksi virus; subset 26 gen ini saja mencapai akurasi 92,86%. Pada data klinis, BioFL‑GNN bersaing ketat dengan OCT‑GCN (78,30% vs. 83,96%), sementara pada data diperluas BioFL‑GNN mengungguli ConL‑HIV secara signifikan, termasuk pada validasi eksternal (87,18% vs. 76,92%). Temuan ini menegaskan bahwa integrasi seleksi fitur metaheuristik, konstruksi knowledge graph berbasis divergensi non‑Euclidean, dan FLC‑GNN merupakan strategi yang efektif untuk klasifikasi respons ART dan penemuan biomarker, sehingga membuka jalan bagi precision medicine dalam tata laksana HIV‑1.
======================================================================================================================================
Human Immunodeficiency Virus type 1 (HIV‑1) remains a global health burden with wide variability in response to Antiretroviral Therapy (ART), necessitating a precision medicine approach through biomarker identification. Conventional methods are suboptimal in capturing nonlinear gene interactions and inter‑patient relational structures. This study developed BioFL‑GNN, a Graph Neural Network (GNN) framework that integrates knowledge graph construction, metaheuristic feature selection, and Fast Local Convolution (FLC) for ART response classification and biomarker discovery. For comparison, OCT‑GCN was applied to clinical data and the Transformer‑based ConL‑HIV to a larger genomic dataset. Microarray gene expression and clinical data were integrated via patient alignment and differential feature construction. The knowledge graph was built using 11 distance metrics, including Bregman divergence and angular distance. Six metaheuristic algorithms (Harmony Search, Ant Colony Optimization, Binary Grey Wolf Optimizer, Binary Cuckoo Search, Genetic Algorithm, and Particle Swarm Optimization) were employed to select informative genes. Four GNN architectures (GCN, GAT, GIN, GraphSAGE) augmented with an FLC layer were trained to classify patients as resistant or non‑resistant. Evaluations were conducted on a primary genomic dataset (138 patients), a clinical dataset (1,056 records), and an extended genomic dataset (406 samples) with external validation. BioFL‑GNN achieved 96.43% accuracy on the primary genomic data and 98.78% on the extended data, with perfect recall. The FLC layer improved average accuracy by 4.7%. A multi‑model consensus identified 26 high‑confidence biomarkers involved in pH regulation, lipid signaling, apoptosis, and viral restriction; the 26‑gene subset alone reached 92.86% accuracy. On clinical data, BioFL‑GNN was competitive with OCT‑GCN (78.30% vs. 83.96%), while on the extended data it significantly outperformed ConL‑HIV, including on external validation (87.18% vs. 76.92%). These findings confirm that integrating metaheuristic feature selection, non‑Euclidean divergence‑based graph construction, and an FLC‑enhanced GNN is an effective strategy for ART response classification and biomarker discovery, paving the way for precision medicine in HIV‑1 management.
| Item Type: | Thesis (Doctoral) |
|---|---|
| Uncontrolled Keywords: | Antiretroviral Therapy, Klasifikasi HIV, Knowledge Graph, Metaheuristik, Graph Neural Networks, Antiretroviral Therapy, HIV Classification, Knowledge Graph, Metaheuristics, Graph Neural Networks |
| Subjects: | Q Science > QA Mathematics > QA336 Artificial Intelligence Q Science > QA Mathematics > QA76.87 Neural networks (Computer Science) R Medicine > R Medicine (General) > R858 Deep Learning R Medicine > RB Pathology |
| Divisions: | Faculty of Science and Data Analytics (SCIENTICS) > Mathematics > 44002-(S3) PhD Thesis |
| Depositing User: | Felza Ridho |
| Date Deposited: | 10 Aug 2026 08:03 |
| Last Modified: | 10 Aug 2026 08:03 |
| URI: | http://repository.its.ac.id/id/eprint/144005 |
Actions (login required)
 |
View Item |