PKP-Diffmol: A Physicochemical Knowledge-Prompt Encoding and Latent Diffusion Framework for Molecular Property...

Molecular representations that capture both chemical information and property-relevant differences are essential for reliable molecular property prediction in drug discovery and toxicity assessments. Fragment-level SMILES representation learning is effective at modeling substructure semantics, but structural reconstruction objectives alone do not explicitly encode the descriptor-derived physicochemical information needed for downstream prediction, especially in small data and scaffold-split settings. To address this limitation, we propose PKP-DiffMol, a physicochemical knowledge-prompt encoding and latent diffusion framework built upon the pretrained SMI-EDITOR encoder. PKP-DiffMol improves fragment-level molecular representations by using numerical and semantic physicochemical knowledge. The numerical physicochemical prior branch encodes RDKit descriptors as continuous quantitative priors, while the physicochemical knowledge prompt encoder transforms the same descriptors into semantic physicochemical representations by using a Transformer-based text encoder. The two complementary views are then integrated through hierarchical physicochemical knowledge fusion to construct a chemically enriched, fused latent space. In this space, a quality-controlled conditional latent diffusion module learns label-conditioned distributions, generates synthetic molecular latent representations, and retains reliable samples through a Mahalanobis distance quality gate. Experiments on seven MoleculeNet classification datasets under scaffold splitting show that PKP-DiffMol increases the mean ROC-AUC from 77.80% for the SMI-EDITOR backbone to 80.53% and obtains the best result on five of the seven datasets. The largest gains are observed on BACE (+6.91%), MUV (+4.29%), and SIDER (+3.92%), showing strong predictive performance across bioactivity prediction, virtual screening, and adverse drug reaction prediction tasks. Further analyses indicate that numerical and semantic physicochemical knowledge improves class separability and the correspondence between representation distances and RDKit descriptor distances, while Mahalanobis-filtered label-conditioned latent augmentation supports prediction-head refinement.