CNN-integrated deep learning optimization of SAG-derived Smoothened agonists relevant to holoprosencephaly

Holoprosencephaly (HPE) is a severe developmental disorder associated with impaired Sonic Hedgehog (SHH) signaling, and no targeted pharmacotherapy is currently available to correct pathway insufficiency. Smoothened (SMO), a central SHH pathway transducer, provides a rational target for optimizing SAG-derived agonists. This study applied an integrated deep learning-guided in silico workflow to generate and prioritize optimized SAG derivatives. A molecular design platform integrating three-dimensional convolutional neural networks, conditional variational autoencoders, and long short-term memory-based molecular generation was used to design SAG analogues, followed by docking, binding-site validation, ADMET/toxicity profiling, pharmacophore analysis, density functional theory calculations, 500 ns molecular dynamics simulations, PCA/DCCM analysis, and MMGBSA estimation. Among eleven derivatives, AI Derivative 7 was the highest ranked compound in the docking screen (-11.396 kcal/mol) compared with SAG (-10.855 kcal/mol), with the margin relative to SAG lying within the uncertainty of the scoring function, and it showed conserved key interactions with ARG400 and ASP577, and expanded hydrophobic and electrostatic contacts. It also showed improved predicted gastrointestinal absorption, reduced lipophilicity (log Po/w 4.32 vs 5.62), lower predicted acute toxicity (LD50 900 vs 200 mg/kg), and a marginally smaller HOMO–LUMO gap for the isolated ligand (approximately 4.06 vs 4.23 eV), which was not related to binding energetics. During simulation, AI Derivative 7 reached a stable conformational state after ~ 200–250 ns with reduced fluctuations and sustained hydrogen bonding. MMGBSA gave a total binding energy of -46.56 kcal/mol for AI Derivative 7 and -46.97 kcal/mol for SAG, a difference of 0.42 kcal/mol that was not statistically distinguishable across the ten sampled snapshots (95% CI -3.96 to + 4.80 kcal/mol, p = 0.84). AI Derivative 7 showed a substantially stronger electrostatic contribution (-44.65 vs -20.75 kcal/mol), but this was offset by a correspondingly larger desolation penalty (+ 66.83 vs + 44.76 kcal/mol), so no net energetic advantage over SAG was obtained. These computational findings identify AI Derivative 7 as a prioritized SMO-targeted candidate requiring experimental validation.