Perbandingan Model Machine Learning untuk Virtual Screening Senyawa Kandidat Pengganti AGP (Antibiotic Growth Promoter)

Abstract

Penggunaan Antibiotic Growth Promoter (AGP) dalam industri unggas telah dibatasi karena kontribusinya terhadap meningkatnya resistensi antibiotik, sehingga diperlukan upaya pencarian kandidat antibiotik alternatif yang aman dan efektif. Penelitian ini bertujuan membangun dan membandingkan performa model klasifi-kasi Support Vector Machine (SVM), C5.0, dan Extreme Gradient Boosting (XGBoost) dalam merekomendasikan kandidat senyawa antibiotik alternatif AGP, serta mengidentifikasi fitur potensial yang berperan dalam prediksi aktivitas senya-wa. Pendekatan virtual screening dilakukan secara in silico menggunakan dataset yang terdiri atas 2.335 senyawa, dengan 120 senyawa aktif dan 2.215 senyawa inaktif. Sebanyak 261 deskriptor molekul diekstraksi dari representasi SMILES (Simplified Molecular Input Line Entry System) menggunakan paket rcdk, diikuti penanganan ketidakseimbangan data dengan ADASYN dan seleksi fitur berbasis Near Zero Variance. Model dilatih menggunakan skema 5-fold cross-validation dan dievaluasi menggunakan metrik akurasi, recall, precision, F1-Score, dan PR-AUC. Hasil menunjukkan bahwa ketiga model mencapai akurasi tinggi pada data uji (93,62-96,58%), namun sensitivitas terhadap kelas minoritas masih terbatas dengan nilai recall 0,30-0,35. Di antara ketiga model, XGBoost menunjukkan performa pal-ing stabil dengan nilai PR-AUC tertinggi sebesar 0,52. Analisis fitur menunjukkan bahwa deskriptor topological dan hybrid berperan dominan dalam proses klasifikasi yang menegaskan pentingnya representasi struktur molekul dalam prediksi aktivitas antibakteri.

The use of Antibiotic Growth Promoters (AGPs) in poultry production has been restricted due to their association with antimicrobial resistance, highlighting the need to identify safe and effective alternative compounds. This study aims to develop and compare the performance of Support Vector Machine (SVM), C5.0, and Extreme Gradient Boosting (XGBoost) classification models for recommending candidate compounds as AGP alternatives, as well as to identify potential features associated with compound activity. An in silico virtual screening approach was applied to a dataset comprising 2.335 compounds, including 120 active and 2.215 inactive compounds. A total of 261 molecular descriptors were extracted from SMILES representations using the rcdk package, followed by class imbalance handling with the ADASYN method and feature selection based on Near Zero Variance. Models were trained using a 5-fold cross-validation scheme and evaluated using accuracy, recall, precision, F1-Score, and PR-AUC metrics. The results show that all models achieved high test accuracy (93,62-96,58%), but exhibited limited sensitivity to the minority class (recall 0,30-0,35). Among the evaluated models, XGBoost demonstrated the most stable performance with the highest PR-AUC value of 0,52. Feature analysis indicates that topological and hybrid descriptors play an important role in predicting antibacterial activity.