The UBio team at the Institute of Innovation Research released UBio-MolFM v1.5, a machine learning interatomic potential for biomolecular dynamics simulations with quantum chemical accuracy. It offers near-classical force field speed and quantum-level precision without retuning per system, demonstrated through two case studies on ion channels and drug molecules.
SynthePulse Insight · AI deep reading
Version 1 · 1 source
The UBio team at ZhiZhi Innovation Research Institute releases UBio-MolFM v1.5, simulating biomolecules with quantum-chemical accuracy, running 180K atoms on a single GPU, with 8.3x speedup and 45% error reduction.
Molecular dynamics simulations have long faced the dilemma of balancing accuracy and scale: quantum chemistry methods (e.g., DFT) are accurate but computationally expensive, typically limited to a few hundred atoms; classical force fields can simulate hundreds of thousands to millions of atoms but rely on approximate parameters, with errors that are difficult to assess a priori.
UBio-MolFM aims to satisfy both simultaneously, simulating biomolecules with quantum-chemical accuracy, with inference speed close to classical force fields, and without the need for re-parameterization for new systems.
In KcsA potassium channel simulations (108,964 atoms, including a realistic lipid bilayer and physiological saline), across five independent trajectories from UBio-MolFM, adjacent potassium ions deep in the filter formed direct contact without water, with an average distance of 3.32 Å, maintained throughout all five trajectories; in contrast, this contact never appeared (0%) across ten trajectories from two classical force fields.
In cyclosporine A conformational free energy simulations, UBio-MolFM resolved a funnel-shaped landscape with a 5.51 kcal/mol drop, with a transmembrane conformation energy cost of about 3.5 kcal/mol; classical force fields showed only a 1.51 kcal/mol drop, nearly flat. Model predictions were validated against DFT single-point calculations on 100 decorrelated conformations, with an energy deviation of 0.44 kcal/mol, smaller than the disagreement among three DFT reference methods (minimum 1.18).
These cases indicate that limitations of classical force fields may go beyond quantitative errors, leading to qualitative misjudgments.
The architecture employs E2Former-V2 equivariant Transformer, satisfying rotational equivariance, non-covalent distance receptive fields, and single-GPU hardware efficiency. Through SO(3)→SO(2) basis transformation, node-centric Wigner-6j decomposition, and Triton streaming attention, near-linear computational complexity is achieved. The backbone stacks four layers with mixed cutoff radii: three layers of 5Å short-range attention covering all atoms, and a fourth layer at 8Å excluding hydrogen, yielding an effective receptive field of about 23Å.
Training data comprises approximately 160 million quantum chemistry labels, from OMol25 (about 140 million) and self-built UBio-Mol26 (about 19 million). UBio-Mol26 enriches methylene and amide groups, with atom-pair distances extending beyond 5-6Å, covering biomolecular features. About 64,000 periodic condensed-phase configurations are used to calibrate thermodynamic properties such as water density.
Three-stage curriculum learning totals about 1000 GPU-days: the first stage uses independent force heads, the second switches to automatic differentiation (F=-∇E), and the third introduces biomolecular data, supervising only force vectors to eliminate DFT systematic energy offsets.
On the baseline model training distribution (OMol-Bio-10k), UMA-S-1p2 achieves the lowest error, with UBio-MolFM not exceeding it but in the top tier. On held-out biomolecular fragments and ultra-large-scale scenarios, UBio-MolFM force errors are 40%-45% lower than the best baseline, with reductions of 21%-64% across various chemical environments.
Under the same test protocol (standard cubic water box, single 141GB H20), standard mode UBio-MolFM processes about 37K atoms/s, supporting up to ~65,000 atoms; activation recomputation mode supports 120K atoms, with throughput of 25.1 vs. 3.0 thousand atoms/s, an 8.3x speedup. Memory growth slopes: UBio-MolFM uses 0.73GB per thousand atoms (activation recomputation) or 1.97GB (standard), while UMA standard mode uses 17.5GB, nearly 9x higher.
By slope calculation, a single GPU theoretically fits 193K atoms, with 180K atoms tested; an 8-GPU node scales to 1.44 million atoms.
Model weights, code, and technical report are open-sourced; links are in the original article.
UBio-MolFM is intended for systems outside the baseline training distribution: those requiring explicit solvent and biomolecular environments, up to million-atom scale. For small-molecule systems with a few hundred atoms, UMA-S-1p2 still holds an accuracy advantage.
This article's information primarily comes from a submitted article by QbitAI, a second-hand source. All performance data and case results are from that article and have not been independently verified. The article does not mention third-party independent comparisons with other models, nor does it provide full technical report details.
UBio-MolFM v1.5 achieves breakthroughs in accuracy and scale, but it should be noted that its performance data comes from the team's own report, and practical application effects await independent replication.
Primary report
Primary source