Newton Divided Difference Optimization for Fingerprint-Based Neural Virtual Screening against Avian Influenza A/H9N2

Siti Amiroch (Universitas Islam Darul 'Ulum)
Mohammad Jamhuri (Universitas Islam Negeri Maulana Malik Ibrahim Malang)
Awawin Mustana Rohmah (Universitas Islam Darul 'Ulum)
Mohammad Hamim Zajuli Al Faroby (Telkom University - Surabaya)
Chairul Anwar Nidom (Universitas Airlangga)
Reviany Vibrianita Nidom (Universitas PGRI Adi Buana Surabaya)

Journal of Applied Data Sciences
Vol 7, No 3: September 2026

Avian influenza A/H9N2 poses persistent zoonotic and veterinary threats, yet efficient computational tools for antiviral compound prioritization remain underdeveloped, particularly with respect to optimizer behavior in high-dimensional neural screening. This study proposes Newton Divided Difference (NDD) optimization, a lightweight positive diagonal curvature-aware training strategy, as a novel optimizer for fingerprint-based neural virtual screening against avian influenza A/H9N2, with the objective of evaluating its performance across aligned molecular fingerprint representations and chemically structured validation protocols. An aligned benchmark of 1,459 molecules consisting of 615 candidate active compounds and 844 decoys was represented by EState (79 features), PubChem (881 features), and Klekota–Roth (4,860 features) fingerprints, sharing identical molecule identities, labels, and split assignments. A fixed multilayer perceptron (MLP) classifier was trained with NDD and seven baseline optimizers under stratified, scaffold-key, and similarity-cluster split protocols across five repeated seeds. NDD achieved the highest descriptive ROC-AUC (Receiver Operating Characteriztic – Area Under the Curve) and PR-AUC (Precision-Recall Area Under the Curve) on Klekota–Roth fingerprints under scaffold-key and similarity-cluster protocols, and remained competitive under the stratified split with ROC-AUC of 0.9872. Architecture-sensitivity tests confirmed stable NDD performance across multiple network configurations, with ROC-AUC values ranging from 0.9876 to 0.9890. Compared with Hessian-free optimization, NDD reduced per-run runtime from approximately 50–54 seconds to approximately 8 seconds on Klekota–Roth under the same CPU-only configuration while achieving comparable ranking performance. The novelty of this work lies in the first systematic assessment of NDD for H9N2 neural virtual screening, demonstrating that positive diagonal curvature-aware scaling provides a practical, stable, and computationally efficient optimization alternative in sparse high-dimensional ligand-based screening settings, although external validation and prospective experimental confirmation remain necessary before practical antiviral prioritization.