Drugon: Phase-Encoded Morgan Fingerprints Match Pretrained Chemical Transformers in Low-Data Affinity Prediction
Year: 2026
Type: preprint
Abstract: Pretrained chemical language models such as ChemBERTa and MoLFormer have become the default choice for molecular property prediction. However, they require large training sets to fine-tune effectively, and their computational footprint is substantial. We present Drugon, a compact alternative based on a Phase-Encoded Morgan fingerprint combined with a LightGBM regressor. Drugon requires approximately 15,000× fewer parameters and 40× less training time than fine-tuned transformers, while achieving higher predictive accuracy on three independent ChEMBL targets (n = 1378 to 2000). On the c-Abl dataset (CHEMBL4550), Drugon reaches a Pearson r = 0.749 under 5-fold cross-validation, compared with 0.605 for fine-tuned ChemBERTa and 0.551 for MoLFormer. Across all three targets, phase encoding provides a consistent, non-negative improvement over the classical binary Morgan fingerprint (∆r ∈ [0.000, 0.026]). Our results suggest that, in the low-data regime (n < 2000), fingerprint-based methods remain the most efficient and interpretable
Source: ChemRxiv
Author: Masoud Azizi
Institution: University of Tabriz
Language: English
Subfield: Drug Discovery
Domain: Life Sciences
Open Access status: green
OpenAlex ID: W7220374852