UBio-MolFM: A Universal Molecular Foundation Model for Bio-Systems

All-atom molecular simulation serves as a quintessential computational microscope'' for understanding the machinery of life, yet it remains fundamentally limited by the trade-off between quantum-mechanical (QM) accuracy and biological scale. We present UBio-MolFM, a universal foundation model framework specifically engineered to bridge this gap. UBio-MolFM introduces three synergistic innovations: (1) UBio-Mol26, a large bio-specific dataset constructed via a multi-fidelity Two-Pronged Strategy’’ that combines systematic bottom-up enumeration with top-down sampling of native protein environments (up to 1,200 atoms); (2) E2Former-V2, a linear-scaling equivariant transformer that integrates Equivariant Axis-Aligned Sparsification (EAAS) and Long-Short Range (LSR) modeling to capture non-local physics with up to ~4x higher inference throughput in our large-system benchmarks; and (3) a Three-Stage Curriculum Learning protocol that transitions from energy initialization to energy-force consistency, with force-focused supervision to mitigate energy offsets. Rigorous benchmarking across microscopic forces and macroscopic observables – including liquid water structure, ionic solvation, and peptide folding – demonstrates that UBio-MolFM achieves ab initio-level fidelity on large, out-of-distribution biomolecular systems (up to ~1,500 atoms) and realistic MD observables. By reconciling scalability with quantum precision, UBio-MolFM provides a robust, ready-to-use tool for the next generation of computational biology.

💡 Research Summary

UBio‑MolFM is presented as a universal molecular foundation model designed to close the long‑standing “scale‑accuracy gap” between quantum‑mechanical (QM) precision and the large spatial scales required for realistic biological simulations. The authors introduce three tightly coupled innovations. First, the UBio‑Mol26 dataset is built using a “Two‑Pronged Strategy” that combines systematic bottom‑up enumeration of biochemical building blocks with top‑down sampling of native protein, DNA, and RNA environments. This results in a bio‑specific collection of up to 1 200 atoms per system and roughly 5 million samples, far exceeding the size and chemical diversity of existing datasets such as SPICE or OMol25, which are limited to small organic molecules. Second, the model architecture, E2Former‑V2, is a linear‑scaling equivariant transformer that incorporates Equivariant Axis‑Aligned Sparsification (EAAS) to prune attention maps along physically meaningful axes, and a Long‑Short Range (LSR) module that captures both local many‑body interactions and distant electrostatics. These design choices reduce memory and compute complexity to O(N) while preserving SE(3) equivariance, delivering up to ~4× higher inference throughput on large systems (1 K–100 K atoms) compared with state‑of‑the‑art equivariant models such as MACE or UMA. Third, training proceeds through a three‑stage curriculum: (1) energy‑only pretraining to establish a rough potential surface, (2) joint energy‑force consistency training to align gradients with the underlying QM forces, and (3) a force‑focused fine‑tuning stage that heavily weights force loss to eliminate systematic energy offsets that arise from multi‑fidelity data. This curriculum ensures that the model not only reproduces static energies but also tracks the temporal evolution of the potential energy surface (PES) during molecular dynamics. Benchmarking is extensive. On an out‑of‑distribution test set containing 1 300–1 500‑atom protein, DNA, and RNA fragments, UBio‑MolFM (trained on both OMol25 and UBio‑Mol26, denoted S3) achieves the lowest relative‑energy MAE (8.68 meV per 100 atoms) and force MAE (16.77 meV/Å) for protein optimizations, and competitive performance for nucleic‑acid systems. Energy‑difference tracking (|ΔE|) along the longest trajectories shows that UBio‑MolFM follows the DFT reference across multiple orders of magnitude, indicating robust PES dynamics. Downstream MD simulations further validate physical fidelity: liquid water and 0.15 M NaCl solutions reproduce experimental O–O radial distribution functions; peptide folding simulations capture solvent‑dependent open/closed conformations; RNA simulations maintain correct metal‑ion coordination. In all cases, the model yields structural observables in close agreement with high‑level QM or experimental data while running at substantially higher speed. An efficiency study demonstrates ~4× faster inference on systems up to ~1 500 atoms, making long‑timescale, solvated‑protein trajectories feasible on modest hardware. The authors acknowledge a limitation: DNA optimization shows a regression in ΔE MAE, attributed to an imbalance in nucleic‑acid coverage within UBio‑Mol26. They propose expanding top‑down sampling of nucleic‑acid environments as future work. Overall, UBio‑MolFM delivers QM‑level accuracy on biologically relevant scales, combines a scalable equivariant architecture with a principled curriculum, and provides an open‑science release of weights, code, and a representative subset of the dataset. This positions the framework as a ready‑to‑use tool for the next generation of high‑fidelity computational biology.