Pendekatan Rekayasa Fitur pada Prediksi Interaksi Senyawa Cyanthillium cinereum sebagai Inhibitor Protein Venom Calloselasma rhodostoma

Abstract

Gigitan ular berbisa dapat menyebabkan kerusakan jaringan, gangguan pembekuan darah, kelumpuhan, dan kematian. Antivenom merupakan terapi utama, tetapi memiliki keterbatasan cakupan spesies dan kemampuan menetralisasi kerusakan lokal. Senyawa bioaktif dari tumbuhan berpotensi menjadi kandidat inhibitor terhadap protein bisa ular. Penelitian ini bertujuan mengevaluasi rekayasa fitur dan model pembelajaran mesin untuk prediksi drug-target interaction (DTI), mengidentifikasi interaksi senyawa Cyanthillium cinereum dengan protein venom Calloselasma rhodostoma dari empat populasi, serta menganalisis pasangan prioritas secara molekuler. Data pelatihan diperoleh dari Snakebite Envenoming Medicines Database dan setelah praproses menghasilkan 102 pasangan interaksi positif. Putative negative dibangkitkan menggunakan GANATCpair untuk membentuk data pelatihan seimbang. Rekayasa fitur mencakup Morgan fingerprint, representasi graf senyawa, komposisi asam amino, convolutional neural network (CNN), dan feature-wise linear modulation (FiLM). Sembilan model dievaluasi menggunakan nested cross-validation, skenario unseen, confusion matrix, dan ablation study. Model terbaik, yaitu GINE-FiLM untuk representasi senyawa dan CNN-FiLM untuk representasi protein, memberikan hasil paling konsisten. Pada nested cross-validation, model ini menghasilkan performa terbaik dengan F1-score, AUROC, dan AUPR masing-masing sebesar 0,920, 0,945, dan 0,965. Model tersebut digunakan untuk memprediksi DTI pada populasi Borneo, Jawa, Malaysia, dan Thailand. Pada masing-masing populasi, 100 pasangan dengan probabilitas tertinggi dipilih untuk molecular docking. Prediksi DTI dan molecular docking menghasilkan tujuh pasangan prioritas yang melibatkan N-Cyclohexylidene-2-carbamylcyclohex-1-enylamine (251559) dan 7-heptadecynoic acid (22600366). Skor docking paling negatif diperoleh pada pasangan protein phospholipase B-like A0A9F2NW96 dengan senyawa 251559, yaitu sebesar –9,056 kcal/mol. Untuk simulasi molecular dynamics, dipilih dua pasangan dari tujuh pasangan prioritas tersebut, yaitu A0A9F2NW96–251559 dan T2HQ57–22600366. Pemilihan dilakukan berdasarkan skor docking, relevansi biologis protein target, serta kedekatan residu pada hasil visualisasi interaksi. Simulasi selama 100 ns menunjukkan bahwa kedua kompleks mengalami penyesuaian lokal tanpa perubahan besar pada struktur global. Perhitungan MM/PBSA menghasilkan energi pengikatan total masing-masing sebesar –27,9813 dan –37,8049 kcal/mol. Hasil tersebut menunjukkan bahwa kedua pasangan memiliki afinitas pengikatan yang baik secara komputasional dan berpotensi menjadi kandidat inhibitor protein bisa ular untuk divalidasi lebih lanjut secara eksperimental.

Venomous snakebites can cause tissue damage, blood coagulation disorders, paralysis, and death. Antivenom is the primary treatment, but it has limitations in species coverage and the neutralization of local tissue damage. Plant-derived compounds may serve as potential inhibitors of snake venom proteins. This study aimed to evaluate feature engineering and machine learning models for drug-target interaction (DTI) prediction, identify interactions between Cyanthillium cinereum compounds and Calloselasma rhodostoma venom proteins from four populations, and analyze the prioritized pairs at the molecular level. The training data were obtained from the Snakebite Envenoming Medicines Database and yielded 102 positive interaction pairs after preprocessing. Putative negatives were generated using GANATCpair to construct a balanced training dataset. Feature engineering included Morgan fingerprints, compound graph representations, amino acid composition, convolutional neural networks (CNN), and feature-wise linear modulation (FiLM). Nine models were evaluated using nested cross-validation, unseen scenarios, a confusion matrix, and an ablation study. The best-performing model, combining GINE-FiLM for compounds and CNN-FiLM for proteins, provided the most consistent results. In nested cross-validation, this model achieved an F1-score of 0.920, an AUROC of 0.945, and an AUPR of 0.965. The model was then used to predict DTI in the Borneo, Java, Malaysia, and Thailand populations. For each population, the 100 pairs with the highest probabilities were selected for molecular docking. DTI prediction and molecular docking identified seven priority compound–protein pairs involving N-Cyclohexylidene-2-carbamylcyclohex-1-enylamine (251559) and 7-heptadecynoic acid (22600366). The most negative docking score was obtained for the phospholipase B-like protein A0A9F2NW96 paired with compound 251559, with a score of –9.056 kcal/mol. Among the seven pairs, A0A9F2NW96–251559 and T2HQ57–22600366 were selected for molecular dynamics simulations based on their docking scores, the biological relevance of the target proteins, and the proximity of residues observed in the interaction visualization. The 100 ns simulations showed that both complexes underwent local adjustments without major changes in global structure. MM/PBSA calculations yielded total binding energies of –27.9813 and –37.8049 kcal/mol, respectively. These findings suggest that both compound–protein pairs are promising candidates for further experimental validation as potential snake venom inhibitors.