AI-Driven Virtual Screening of the IRACEMA Database for Novel Modulators of Amyloid-Oligomer Neurotoxicity

The aggregation of the amyloid-β peptide (Aβ) represents the primary pathological hallmark of Alzheimer's Disease (AD), the most prevalent neurodegenerative disorder worldwide, which currently lacks a cure. Aβ aggregation occurs via a complex pathway wherein misfolded monomers assemble into oligomers, which subsequently elongate into protofibrils until reaching saturation as insoluble fibrils. Amyloid-β oligomers (AβOs) have emerged as highly promising targets for disease-modifying therapies, as they are widely considered the most neurotoxic species. Here, we employed artificial intelligence (AI)-driven computational frameworks to explore the IRACEMA database, which comprises synthetic bioactive compounds designed by Brazilian reseachers. The virtual screening was performed using Boltz-2, which integrates structural prediction and affinity estimation at levels comparable to traditional physical methods. This deep learning tool, utilizing an architecture analogous to AlphaFold 3, was applied to predict the 3D structures of complexes formed between AβOs and the IRACEMA database compounds. Boltz-2 provides two distinct parameters for affinity calculation: affinity_pred_value and affinity_probability_binary. First, the affinity_probability_binary module was utilized to eliminate potential decoys, applying a selection threshold of > 0.5, which excluded 12 of the 520 library ligands. Subsequently, the remaining ligands were ranked according to their affinity_pred_value. These hits were then subjected to the ADMET_Risk filter within ADMET Predictor 13. Following, this in silico approach, we identified 8 potential hits from the IRACEMA database as promising ligands of AβOs. Among them, candidates IRA-131, IRA-466, and IRA-130 emerged as the top hits, with estimated IC₅₀ values in the 50 to 70 nM range. Furthermore, ligands IRA-473, IRA-474, IRA-475, and IRA-476, from a series of β-caryophyllene oxide derivatives originally developed for Trypanosoma cruzi activity, exhibited exceptionally low ADMET risk scores (ranging from 2.0 to 3.5) and estimated IC₅₀ values up to 170 nM. We adopted a dual-model predictive pipeline combining 3D structural/affinity predictions with ADMET risk profiling. This approach serves strictly as a prioritization tool, allowing us to select candidates with favorable theoretical pharmacokinetic profiles for subsequent experimental validation using biophysical and imaging techniques.