Year: 2026
Type: preprint
Abstract: Predicting molecular properties directly from string representations avoids the computational cost of explicit 3D structure generation, yet the relative merits of different representations remain poorly understood. We systematically evaluated SMILES (Simplified Molecular Input Line Entry System) and SELFIES (SELF-referencIng Embedded Strings) molecular representations across 50 transformer model configurations on seven properties---five classification tasks (molecular dimension, ring count, chirality, symmetry planes, point group) and two regression tasks (planar fit error, ring plane angles)---derived from 3.9 million density functional theory-optimized structures. SMILES consistently outperforms SELFIES across all properties; model depth improves performance on globally complex properties without overfitting; and multi-task learning benefits classification but degrades regression, with the best configurations achieving combined scores above 0.85 on a normalized seven-task benchmark. Encoder-only transformer models (2, 4, and 8 encoder blocks; d\textsubscript{model}) were trained using PyTorch on NVIDIA A100 GPUs (Leonardo cluster, CINECA). SMILES strings were generated with RDKit; SELFIES representations with the SELFIES library. Molecular geometries originate from the Molecule3D benchmark, computed at the B3LYP/6-31G* level of theory via PubChemQC. Point group symmetries were computed using the SOFI library. Three architectural variants were compared: single-task, parallel multi-task, and hierarchical multi-task with cross-attention between property prediction heads.
Source: Research Square
Authors: Rokas Garbacauskas, Stepas Toliautas
Institution: Vilnius University
Language: English
Subfield: Drug Discovery
Domain: Life Sciences
Open Access status: green
Funder: Nvidia
OpenAlex ID: W7236712247
