Background: Lysine-specific demethylase-1 (LSD1/KDM1A) erases mono- and dimethyl marks on histone H3 and is over-expressed in prostate, breast, and lung tumours, making it an increasingly attractive epigenetic target for anticancer therapy. Although the natural polyphenol curcumin inhibits LSD1 only weakly, its modular scaffold lends itself to systematic optimisation through computational chemistry. Methods: A previously published study compiled chemical structures and Ellman-assay IC₅₀ values for nineteen curcumin analogues. After three-dimensional geometry optimisation (MM⁺ → PM3), 4,885 Dragon descriptors were calculated; variance, missing-value, and multicollinearity filters (r ≥ 0.95) reduced the pool to 763. Stepwise selection retained four informative descriptors—P_VSA_s_5, JGI8, H2s, and SpPosA_A—reflecting polar surface area, eighth-order topological charge, spatial polarizability, and positive fragment surface. A radial-basis-function artificial neural network (4-6-1 architecture) was trained on thirteen compounds, internally validated by leave-one-out cross-validation, and externally evaluated on a three-compound test set. Results: The model reproduced experimental potency with R² = 0.999, Q² = 0.9996, and MAE = 0.11 log units; external prediction yielded R²test = 0.928. Sensitivity analysis identified P_VSA_s_5 as the dominant contributor, indicating that enlarging the polar surface area in specific atomic states enhances enzyme binding, while the JGI8 descriptor underscored the importance of charge distribution. Conclusions: This compact, rigorously validated QSAR model offers accurate, mechanism-based predictions and actionable design rules—optimal polar surface area, favourable charge topology, and judicious halogenation—for next-generation LSD1 inhibitors, thereby accelerating epigenetic drug discovery pipelines.