Transfer Learning-Derived Molecular Descriptors Enable Reliable Property Prediction under Extreme Data Scarcity

Reliable prediction of experimental physicochemical properties remains difficult when only a few measured data points are available. Here, we developed FP200, a fixed 200-dimensional molecular descriptor extracted from the graph-level representation of a message-passing neural network pretrained on QM9 quantum-chemical data using DFT-optimized coordinates and atomic species. FP200 was evaluated for ten experimental physicochemical properties and compared with eleven established descriptor sets, including group-contribution-based descriptors, molecular fingerprints, RDKit descriptors, and direct use of QM9-calculated properties. With only 25 model-construction samples, FP200 achieved the lowest mean absolute error for seven of the ten properties; for the remaining three, its performance was statistically comparable to the best descriptor. FP200 also remained competitive as the model-construction data size increased, indicating that quantum-chemistry-informed learned descriptors can be effective even when downstream experimental data are severely limited. Ablation analyses confirmed the importance of QM9 pretraining and fixed feature extraction for obtaining stable small-data performance. Principal component analysis, principal-component compression, and Shapley-value-based feature-attribution analyses suggested that FP200 acts as a redundant yet broadly utilized representation, with predictive information distributed across correlated descriptor components. Generalization and applicability-domain analyses further showed that descriptor-space distance can serve as a practical indicator of prediction reliability. These findings support FP200 as a feature-based pretrained descriptor for small-data physicochemical property prediction within the evaluated QM9-derived chemical space and its applicability domain.