Deep learning-driven discovery and optimization of natural LSD1 inhibitors for the treatment of Alzheimer's disease.

Abstract

Alzheimer's disease (AD) is a prevalent neurodegenerative disorder with limited effective disease-modifying treatments. Lysine-specific demethylase 1 (LSD1) has emerged as a promising target for AD therapy. However, current LSD1 inhibitors for AD still suffer from poor brain permeability, off-target toxicity, and chemical-scaffold scarcity. Herein, we developed a multimodal deep learning model (PLM-CAFT-DTA) for drug-target affinity (DTA) prediction. This model integrates ChemBERTa, ESM-2, graph attention, and cross-attention fusion to achieve high prediction precision. Using this model combined with virtual screening and molecular simulation, we identified silybin as a hit compound from a library of over 70,000 natural products. After rational modification, compound S3 was obtained with significantly improved LSD1 inhibition (IC₅₀ = 2.30 μM), approximately 7-fold more potent than the silybin. In vitro assays showed that S3 exhibited favorable neuroprotective and antioxidant activities. In APP/PS1 mice, S3 upregulated hippocampal H3K9me2, suppressed neuroinflammation and Aβ deposition, and improved cognitive function. By addressing unmet demands for AI-assisted anti-AD lead discovery, this study provides a generalized DTA tool for early-stage drug development, and identifies S3 as a novel LSD1 inhibitor with potent anti-AD efficacy.