Abstract
<jats:p>Boron-dipyrromethene (BODIPY) photosensitizers (PS) are important candidates for photodynamic therapy (PDT) and other light-driven applications because of their readily tunable photophysical properties. However, existing molecular design approaches are constrained by insufficient property-prediction accuracy, inefficient structure generation and limited consideration of experimental feasibility. Here, we develop an integrated molecular design framework that combines multimodal property prediction, fingerprint-based molecular diffusion, GRU–SELFIES decoding and human-in-the-loop decision-making. We further develop a Perturbation–Immediate Denoising–Evaluation (PIDE) guidance strategy that decouples molecular generation from property prediction, unlike classifier guidance, which requires a property predictor to estimate properties from noisy intermediate states during diffusion. The framework was trained using 889,033 symmetric BODIPY molecules, whose structural symmetry was introduced to reduce the expected difficulty of chemical synthesis. Compared with PaiNN, Chemprop, random forest, XGBoost and support vector machine baselines, the multimodal fusion model improved R2 by 0.6–86.0% across five regression tasks. The fingerprint diffusion model achieved a molecular validity of 96.88%, compared with 54.79% for GDSS. Complete unconditional generation from random fingerprint initialization through 200 reverse-diffusion steps and GRU–SELFIES decoding to SMILES required 0.000911 ± 0.000009 s per molecule. When benchmarked using the same device, GDSS required 4.3704 s per generated sample (approximately 4,797-fold higher). The GRUSELFIES model correctly translated 93.5% of fingerprint-derived tokens and exactly reproduced the reference SMILES for 82.1% of molecules, with 100% molecular validity and an approximately 193-fold speedup over the Transformer–SMILES decoder. Following automated QSPR-based ranking, human experts participate in every round of guided molecular optimization rather than evaluating the candidates only after the entire generation process is completed. At each round, experts assess the synthetic accessibility, commercial availability of starting materials and empirical chemical value of the proposed molecules. This iterative interaction enables overestimated out-of-distribution candidates to be identified and removed at an early stage, preventing them from being propagated into subsequent rounds of guidance and reducing the risk of investing wet-laboratory resources in poorly supported candidates.</jats:p>