pith. sign in

arxiv: 2602.17709 · v2 · submitted 2026-02-13 · ⚛️ physics.chem-ph · cs.AI· physics.bio-ph

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

Lin Huang , Arthur Jiang , XiaoLi Liu , Zion Wang , Jason Zhao , Chu Wang , HaoCheng Lu , ChengXiang Huang
show 3 more authors
Jiajun Cheng YiYue Du Jia Zhang
This is my paper

Pith reviewed 2026-05-15 22:45 UTC · model grok-4.3

classification ⚛️ physics.chem-ph cs.AIphysics.bio-ph
keywords molecular foundation modelequivariant transformerbiomolecular simulationab initio accuracymolecular dynamicsprotein environmentsforce field development
0
0 comments X

The pith

UBio-MolFM reaches ab initio-level accuracy on out-of-distribution biomolecular systems up to 1500 atoms using a bio-specific dataset, linear-scaling equivariant transformer, and staged training.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper presents UBio-MolFM to resolve the longstanding limit on running quantum-accurate simulations of life-scale molecules. It constructs a large dataset of biomolecular configurations through systematic enumeration and sampling from real proteins. An equivariant transformer architecture processes these with built-in handling for both local and distant interactions at high speed. A three-phase training sequence first aligns energies, then forces, then consistency between them. Benchmarks on water, ions, and peptides confirm the model reproduces microscopic forces and macroscopic behaviors at the level of direct quantum calculations even on systems far from its training distribution.

Core claim

UBio-MolFM is a universal molecular foundation model that combines the UBio-Mol26 dataset built by a two-pronged multi-fidelity strategy, the E2Former-V2 linear-scaling equivariant transformer with Equivariant Axis-Aligned Sparsification and Long-Short Range modeling, and a three-stage curriculum that moves from energy initialization through energy-force consistency to force-focused refinement. When tested on liquid water structure, ionic solvation, peptide folding, and other observables, the resulting model delivers ab initio-level fidelity on large out-of-distribution biomolecular systems up to approximately 1500 atoms while supporting realistic molecular-dynamics trajectories.

What carries the argument

E2Former-V2 linear-scaling equivariant transformer with Equivariant Axis-Aligned Sparsification and Long-Short Range modeling, trained via the three-stage curriculum on the UBio-Mol26 dataset.

If this is right

  • Enables molecular-dynamics runs of protein-scale systems that match quantum forces without the usual scale-accuracy tradeoff.
  • Reproduces structural observables such as water radial distribution functions and ion solvation shells at realistic temperatures.
  • Supports folding trajectories for peptides that align with experimental timescales and endpoints.
  • Delivers up to fourfold higher inference speed on large systems through the sparsification and long-short range design.
  • Supplies a ready-to-deploy model for computing dynamics in out-of-distribution biomolecular environments.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same three-stage protocol could be applied to other molecular domains such as materials or catalysis once an analogous multi-fidelity dataset is assembled.
  • Longer-timescale simulations of conformational changes in proteins become feasible if the model maintains stability over millions of steps.
  • Integration into existing molecular-dynamics packages would allow routine replacement of classical force fields with this higher-accuracy option for systems up to 1500 atoms.

Load-bearing premise

The multi-fidelity dataset and staged training produce a model whose quantum accuracy holds when the input shifts to real native protein environments and previously unseen large systems.

What would settle it

Direct comparison of UBio-MolFM forces and energies against new high-level ab initio reference calculations on a biomolecule of 1200-1500 atoms drawn from an environment outside the UBio-Mol26 sampling would show systematic deviation if the generalization claim fails.

Figures

Figures reproduced from arXiv: 2602.17709 by Arthur Jiang, ChengXiang Huang, Chu Wang, HaoCheng Lu, Jason Zhao, Jiajun Cheng, Jia Zhang, Lin Huang, XiaoLi Liu, YiYue Du, Zion Wang.

Figure 1
Figure 1. Figure 1: The UBio-MolFM Framework. Our approach bridges the scale-accuracy gap through three synergistic pillars: (1) Data: The UBio-Mol26 dataset, constructed via a Two￾Pronged Strategy where a bottom-up branch systematically enumerates biochemical building blocks and a top-down branch samples native environments from large protein assemblies; (2) Model: The E2Former-V2 architecture, which achieves linear memory s… view at source ↗
Figure 2
Figure 2. Figure 2: Trajectory Analysis of Potential Energy Differences. Comparison of predicted vs. DFT energy changes along the longest trajectories from each benchmark category. Absolute values of energy changes (|∆E|) are plotted on a logarithmic scale. UBio-MolFM (S3) exhibits superior alignment with ground truth fluctuations, while the S2 base consistently demonstrates the robust inductive bias of the E2Former-V2 archit… view at source ↗
Figure 3
Figure 3. Figure 3: Structural Fidelity of Liquid Water. Comparison of oxygen-oxygen radial dis￾tribution functions (O–O RDF) derived from (a) UBio-MolFM (S3) and (b) UMA-S-1p1 NV T trajectories against experimental references [14, 15]. Both models exhibit excellent agreement with experiment; UBio-MolFM (S3) accurately reproduces the primary and secondary hydra￾tion shells, demonstrating the physical robustness of the learned… view at source ↗
Figure 4
Figure 4. Figure 4: Hydration Structure of 0.15 mol/L NaCl Solution. Radial distribution func￾tions (RDF) for Na–O, Cl–O, and Na–Cl pairs from a 200 ps NV T simulation. The correspond￾ing coordination numbers (CN) for the first hydration shells of Na+ and Cl− are marked [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Environmental Dependence of Cyclosporine A Conformations. (a) In the aqueous trajectory, intramolecular H-bond distances for key residues remain large, indicating that the initial open state is stably maintained. (b) This stability is driven by consistent hydration, as shown by the high occupancy of hydrogen bonds between the peptide and solvent oxygens. (c) In contrast, the vacuum trajectory shows stable … view at source ↗
Figure 6
Figure 6. Figure 6: Mg2+ Coordination to RNA Phosphate (1L2X). Comparison of geometric distributions for Mg–O distances and angles against experimental benchmarks [21]. Amber99 captures coordination topology but exhibits rigid angles (∼160◦ ) and underestimates Mg–O distances. UMA mischaracterizes Mg–phosphate interactions with overestimated distances and skewed angles. UBio-MolFM (S3) demonstrates superior fidelity, reproduc… view at source ↗
Figure 7
Figure 7. Figure 7: Data Generation Pipeline. (1) Assembly: Solutes and solvents are packed using Packmol and parameterized with AmberTools. (2) Relaxation: Steric clashes are removed via OpenMM minimization. (3) Calculation: High-fidelity DFT labels are generated using GPU4PySCF, employing both direct optimization and MD sampling strategies to maximize structural diversity. 4.1.4 Data Analysis and Visualization With the 17 m… view at source ↗
Figure 8
Figure 8. Figure 8: Statistical overview of UBio-Mol26. (a) Histogram of atom counts highlighting the extensive coverage of large-scale biological systems. (b) Distribution of configurations across different biological categories. • Proteins: This category forms the backbone of our dataset, constructed through a dual approach. It includes exhaustively enumerated solvated tripeptides to capture fundamen￾tal backbone dynamics (… view at source ↗
Figure 9
Figure 9. Figure 9: t-SNE comparison of UBio-Mol26 and OMol25. Each panel shows the distribu￾tion of 1M sampled configurations in the t-SNE-reduced feature space, categorized by biological domain and basis set (def2-SVP vs def2-TZVPD). The gray background represents OMol25, highlighting the complementary coverage of UBio-Mol26 in the macromolecular regime. Local Chemical Environments. To explain the chemical basis of this str… view at source ↗
Figure 10
Figure 10. Figure 10: Distribution of Carbon chemical environments. Comparison of functional group frequencies for one million Carbon atoms sampled from OMol25 (left) and UBio-Mol26 subsets (def2-SVP, middle; def2-TZVPD, right). UBio-Mol26 exhibits a significantly higher proportion of methylene and amide groups, consistent with its emphasis on proteins and lipids, whereas OMol25 is enriched in aromatic groups typical of small … view at source ↗
Figure 11
Figure 11. Figure 11: Pair distance distributions for C, N, O, H. Comparison of interatomic distance distributions across OMol25 and UBio-Mol26 subsets. UBio-Mol26 captures significantly longer￾range structural information essential for biological macromolecules. 21 [PITH_FULL_IMAGE:figures/full_fig_p021_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Elemental distribution across the periodic table. Heatmaps showing the fre￾quency of elements included in the def2-SVP and def2-TZVPD subsets of UBio-Mol26, demon￾strating comprehensive coverage of biologically relevant elements and trace ions. 4.2 Model Architecture We first summarize the representation, architectural, and computational principles underlying our approach. Our objective is to construct an… view at source ↗
Figure 13
Figure 13. Figure 13: E2Former-V2 key components. We propose Equivariant Axis-Aligned Spar￾sification (EAAS) to simplify tensor couplings through SO(2) re-indexing and utilize a fused on-the-fly equivariant kernel for memory-efficient, edge-free attention computation. Equivariant Axis-Aligned Sparsification (EAAS) To resolve the arithmetic bottleneck of dense tensor products, we introduce EAAS, an algebraic reduction that conv… view at source ↗
Figure 14
Figure 14. Figure 14: Atom-balanced batching and Atom-centric message passing improve efficiency by [PITH_FULL_IMAGE:figures/full_fig_p029_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Structural Distribution of UBio-Protein26 5M. The dataset provides a com￾prehensive sampling of protein conformational space and local chemical environments. 5.2 Code and Inference The implementation of the E2Former-V2 architecture is already available at: https://github.com/IQuestLab/UBio-MolFM Standardized inference scripts for UBio-MolFM, along with integrated modules for running molecular dynamics (MD… view at source ↗
read the original abstract

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.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 1 minor

Summary. The paper presents UBio-MolFM, a universal foundation model for biomolecular systems. It introduces the UBio-Mol26 dataset (constructed via bottom-up enumeration and top-down sampling of native protein environments up to 1,200 atoms), the E2Former-V2 linear-scaling equivariant transformer incorporating Equivariant Axis-Aligned Sparsification (EAAS) and Long-Short Range (LSR) modeling, and a three-stage curriculum learning protocol (energy initialization to energy-force consistency with force-focused supervision). Benchmarking on microscopic forces and macroscopic observables (liquid water structure, ionic solvation, peptide folding) is claimed to demonstrate ab initio-level fidelity on large out-of-distribution systems up to ~1,500 atoms.

Significance. If the central claims hold, the work would be significant for enabling scalable, quantum-accurate all-atom simulations of biological systems beyond the reach of direct QM methods, potentially providing a practical tool for studying protein dynamics and environments at realistic scales.

major comments (1)
  1. [Abstract] Abstract: The claim of achieving 'ab initio-level fidelity on large, out-of-distribution biomolecular systems (up to ~1,500 atoms)' rests on macroscopic observables (liquid water structure, ionic solvation, peptide folding). These observables are insensitive to small force errors and can be reproduced by classical or semi-empirical potentials, so they do not confirm the asserted microscopic QM force accuracy for OOD cases where direct DFT/wavefunction validation is intractable.
minor comments (1)
  1. [Abstract] Abstract: The phrase 'ab initio-level fidelity' requires a precise quantitative definition (e.g., specific error thresholds relative to a reference QM method such as DFT or CCSD(T)) to allow readers to assess the strength of the benchmarking results.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for their constructive feedback. We address the major comment below and indicate the revisions we will incorporate.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The claim of achieving 'ab initio-level fidelity on large, out-of-distribution biomolecular systems (up to ~1,500 atoms)' rests on macroscopic observables (liquid water structure, ionic solvation, peptide folding). These observables are insensitive to small force errors and can be reproduced by classical or semi-empirical potentials, so they do not confirm the asserted microscopic QM force accuracy for OOD cases where direct DFT/wavefunction validation is intractable.

    Authors: We agree that direct microscopic QM force validation is intractable for the largest OOD systems (~1,500 atoms) and that macroscopic observables serve as an indirect proxy. The manuscript already reports microscopic force benchmarks on smaller OOD systems (where QM is feasible) and employs multi-fidelity training plus force-focused curriculum learning to target microscopic accuracy. To address the concern, we will revise the abstract to explicitly distinguish direct microscopic validation (where computationally tractable) from macroscopic validation for large OOD cases, and we will add a dedicated limitations paragraph discussing the sensitivity of macroscopic observables along with additional quantitative force-error metrics from our validation sets. revision: partial

Circularity Check

0 steps flagged

No significant circularity; claims rest on empirical benchmarking of trained model

full rationale

The paper describes construction of UBio-Mol26 via multi-fidelity sampling, the E2Former-V2 architecture with EAAS and LSR, and a three-stage curriculum from energy initialization to force supervision. It then reports benchmarking on microscopic forces plus macroscopic observables (water structure, solvation, peptide folding) for OOD systems up to ~1500 atoms. No quoted equations, self-citations, or uniqueness theorems reduce any performance claim to a fitted input or self-definition by construction. The validation observables are distinct from the training targets and are presented as external checks, making the derivation self-contained under the given criteria.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review provides no explicit free parameters, axioms, or invented entities; full text would be required to enumerate fitted hyperparameters, architectural assumptions, or new postulated components.

pith-pipeline@v0.9.0 · 5604 in / 1204 out tokens · 54190 ms · 2026-05-15T22:45:08.789819+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

48 extracted references · 48 canonical work pages

  1. [1]

    Dotson, Raimondas Galvelis, John E

    Peter Eastman, Pavan Kumar Behara, David L. Dotson, Raimondas Galvelis, John E. Herr, Josh T. Wills, and John D. Chodera. SPICE, a dataset of drug-like molecules and peptides for training machine learning potentials.Scientific Data, 10(1):11, 2023

    work page 2023

  2. [2]

    Levine, Muhammed Shuaibi, Evan Walter Clark Spotte-Smith, Michael G

    David S. Levine, Muhammad Shuaibi, Evan W. C. Spotte-Smith, Michael G. Taylor, Muhammad R. Hasyim, Kyle Michel, Ilyes Batatia, G´ abor Cs´ anyi, Mark Dzamba, Pe- ter Eastman, Nathan C. Frey, Xiang Fu, Vardan Gharakhanyan, Aditi S. Krishnapriyan, Joshua A. Rackers, Saro Raja, Adeel Rizvi, Andrew S. Rosen, Zachary Ulissi, Santiago Var- gas, C. Lawrence Zitn...

    work page arXiv 2025

  3. [3]

    Scalable machine learning force fields for macromolecular systems through long-range aware message passing.arXiv preprint arXiv:2601.03774, 2026

    Chu Wang, Lin Huang, Xinran Wei, Tao Qin, Arthur Jiang, Lixue Cheng, and Jia Zhang. Scalable machine learning force fields for macromolecular systems through long-range aware message passing.arXiv preprint arXiv:2601.03774, 2026

    work page arXiv 2026

  4. [4]

    Kovacs, Gregor Simm, Christoph Ortner, and G´ abor Cs´ anyi

    Ilyes Batatia, David P. Kovacs, Gregor Simm, Christoph Ortner, and G´ abor Cs´ anyi. MACE: Higher order equivariant message passing neural networks for fast and accurate force fields.Advances in Neural Information Processing Systems, 35:11423–11436, 2022

    work page 2022

  5. [5]

    Mailoa, Mordechai Kornbluth, Nicola Molinari, Tess E

    Simon Batzner, Albert Musaelian, Lixin Sun, Mario Geiger, Jonathan P. Mailoa, Mordechai Kornbluth, Nicola Molinari, Tess E. Smidt, and Boris Kozinsky. E(3)-equivariant graph neural networks for data-efficient and accurate interatomic potentials.Nature Communi- cations, 13(1):2453, 2022

    work page 2022

  6. [6]

    Lawrence Zitnick

    Saro Passaro and C. Lawrence Zitnick. Reducing so(3) convolutions to so(2) for efficient equivariant gnns. InProceedings of the 40th International Conference on Machine Learning, volume 202, pages 27420–27438. PMLR, 2023

    work page 2023

  7. [8]

    Unke, Martin St¨ ohr, Stefan Ganscha, Thomas Unterthiner, Hartmut Maennel, Sergey Kashubin, Daniel Ahlin, Michael Gastegger, Muhammad Shuaibi, Benjamin Neb- gen, et al

    Oliver T. Unke, Martin St¨ ohr, Stefan Ganscha, Thomas Unterthiner, Hartmut Maennel, Sergey Kashubin, Daniel Ahlin, Michael Gastegger, Muhammad Shuaibi, Benjamin Neb- gen, et al. Biomolecular dynamics with machine-learned quantum-mechanical force fields trained on diverse chemical fragments.Science Advances, 10(4):eadn4397, 2024

    work page 2024

  8. [9]

    E2former-v2: On-the-fly equivariant attention with linear activation memory.arXiv preprint arXiv:2601.16622, 2026

    Lin Huang, Chengxiang Huang, Ziang Wang, Yiyue Du, Chu Wang, Haocheng Lu, Yunyang Li, Xiaoli Liu, Arthur Jiang, and Jia Zhang. E2former-v2: On-the-fly equivariant attention with linear activation memory.arXiv preprint arXiv:2601.16622, 2026

    work page arXiv 2026

  9. [10]

    E2former: An efficient and equivariant transformer with linear-scaling tensor products, 2025

    Yunyang Li, Lin Huang, Zhihao Ding, Chu Wang, Xinran Wei, Han Yang, Zun Wang, Chang Liu, Yu Shi, Peiran Jin, et al. E2former: An efficient and equivariant transformer with linear-scaling tensor products, 2025

    work page 2025

  10. [11]

    Junmei Wang, Piotr Cieplak, and Peter A Kollman. How well does a restrained electro- static potential (resp) model perform in calculating conformational energies of organic and biological molecules?Journal of Computational Chemistry, 21(12):1049–1074, 2000

    work page 2000

  11. [12]

    Mihaly Varadi, Stephen Anyango, Mandar Deshpande, Sreenath Nair, Cindy Natassia, Galabina Yordanova, David Yuan, Oana Stroe, Gemma Wood, Agathe Leyrou, et al. AlphaFold Protein Structure Database: massively expanding the structural coverage of protein-sequence space with high-accuracy models.Nucleic Acids Research, 50(D1):D439– D444, 2022. 33 April 14, 20...

    work page 2022

  12. [13]

    D. A. Case et al.AmberTools 23. University of California, San Francisco, 2023

    work page 2023

  13. [14]

    The structure of water around the compressibility minimum.The Journal of Chemical Physics, 141(21):214507, 2014

    LB Skinner, CJ Benmore, JC Neuefeind, and JB Parise. The structure of water around the compressibility minimum.The Journal of Chemical Physics, 141(21):214507, 2014

    work page 2014

  14. [15]

    Ab initio elec- tronic structure of liquid water.Physical Review Letters, 117(18):186401, 2016

    Wei Chen, Francesco Ambrosio, Giacomo Miceli, and Alfredo Pasquarello. Ab initio elec- tronic structure of liquid water.Physical Review Letters, 117(18):186401, 2016

    work page 2016

  15. [16]

    Revisiting the hydration structure of aqueous na+.The Journal of Chemical Physics, 146(8):084504, 2017

    Mizied Galib, Marcel D Baer, Lawrie B Skinner, Christopher J Mundy, Thomas Huth- welker, Gregory K Schenter, Chris J Benmore, Niranjan Govind, and John L Fulton. Revisiting the hydration structure of aqueous na+.The Journal of Chemical Physics, 146(8):084504, 2017

    work page 2017

  16. [17]

    Hyerim Hwang, Yong Chan Cho, Sooheyong Lee, Yun-Hee Lee, S Kim, Y Kim, W Jo, P Duchstein, D Zahn, and Geun Woo Lee. Hydration breaking and chemical ordering in a levitated nacl solution droplet beyond the metastable zone width limit: evidence for the early stage of two-step nucleation.Chemical Science, 12(1):179–187, 2021

    work page 2021

  17. [18]

    Linking electronic and molecular structure: insight into aqueous chloride solvation.Physical Chemistry Chemical Physics, 15(31):13169–13183, 2013

    Ling Ge, Leonardo Bernasconi, and Patricia Hunt. Linking electronic and molecular structure: insight into aqueous chloride solvation.Physical Chemistry Chemical Physics, 15(31):13169–13183, 2013

    work page 2013

  18. [19]

    Local order in aqueous nacl solutions and pure water: X-ray scattering and molecular dynamics simulations study.The Journal of Physical Chemistry B, 110(46):23515–23523, 2006

    Salah Bouazizi, Salah Nasr, Nejmeddine Jaˆ ıdane, and Marie-Claire Bellissent-Funel. Local order in aqueous nacl solutions and pure water: X-ray scattering and molecular dynamics simulations study.The Journal of Physical Chemistry B, 110(46):23515–23523, 2006

    work page 2006

  19. [20]

    Metal ions and flexibility in a viral rna pseudoknot at atomic resolution.Proceedings of the National Academy of Sciences, 99(7):4302–4307, 2002

    Martin Egli, George Minasov, Li Su, and Alexander Rich. Metal ions and flexibility in a viral rna pseudoknot at atomic resolution.Proceedings of the National Academy of Sciences, 99(7):4302–4307, 2002

    work page 2002

  20. [21]

    Zheng, I.G

    H. Zheng, I.G. Shabalin, K.B. Handing, J.M. Bujnicki, and W. Minor. Principles of ion binding to rna inferred from the analysis of a 1.55 a resolution bacterial ribosome structure - part i: Mg2+.Nucleic Acids Research, 43(7):3780–3794, 2015

    work page 2015

  21. [22]

    Uma: A family of universal models for atoms

    Brandon M Wood, Misko Dzamba, Xiang Fu, Meng Gao, Muhammed Shuaibi, Luis Barroso-Luque, Kareem Abdelmaqsoud, Vahe Gharakhanyan, John R Kitchin, Daniel S Levine, et al. Uma: A family of universal models for atoms.arXiv preprint arXiv:2506.23971, 2025

    work page arXiv 2025

  22. [23]

    arXiv preprint arXiv:2502.12147 (2025)

    Xiang Fu, Brandon M Wood, Luis Barroso-Luque, Daniel S Levine, Meng Gao, Misko Dzamba, and C Lawrence Zitnick. Learning smooth and expressive interatomic potentials for physical property prediction.arXiv preprint arXiv:2502.12147, 2025

    work page arXiv 2025

  23. [24]

    Mace-off: Short-range transferable machine learning force fields for organic molecules.Journal of the American Chemical Society, 147(21):17598–17611, 2025

    D´ avid P´ eter Kov´ acs, J Harry Moore, Nicholas J Browning, Ilyes Batatia, Joshua T Horton, Yixuan Pu, Venkat Kapil, William C Witt, Ioan-Bogdan Magd˘ au, Daniel J Cole, and G´ abor Cs´ anyi. Mace-off: Short-range transferable machine learning force fields for organic molecules.Journal of the American Chemical Society, 147(21):17598–17611, 2025

    work page 2025

  24. [25]

    E2former: An efficient and equivariant transformer with linear-scaling tensor products, 2025

    Yunyang Li, Lin Huang, Zhihao Ding, Chu Wang, Xinran Wei, Han Yang, Zun Wang, Chang Liu, Yu Shi, Peiran Jin, Tao Qin, Mark Gerstein, and Jia Zhang. E2former: An efficient and equivariant transformer with linear-scaling tensor products, 2025

    work page 2025

  25. [26]

    PubChemQC project: a large-scale first-principles electronic structure database for data-driven chemistry.Journal of Chemical Information and Modeling, 57(6):1300–1308, 2017

    Maho Nakata and Tomomi Shimazaki. PubChemQC project: a large-scale first-principles electronic structure database for data-driven chemistry.Journal of Chemical Information and Modeling, 57(6):1300–1308, 2017. 34 April 14, 2026UBio-MolFM Technical Report

    work page 2017

  26. [27]

    Optimization of parameters for semiempirical methods v: Modification of nddo approximations and application to 70 elements.Journal of Molecular Modeling, 13:1173–1213, 2007

    James JP Stewart. Optimization of parameters for semiempirical methods v: Modification of nddo approximations and application to 70 elements.Journal of Molecular Modeling, 13:1173–1213, 2007

    work page 2007

  27. [28]

    Density-functional thermochemistry

    Axel D Becke. Density-functional thermochemistry. iii. the role of exact exchange.The Journal of chemical physics, 98(7):5648–5652, 1993

    work page 1993

  28. [29]

    Spice 2.0.1, 2024

    Peter Eastman, Pavan Kumar Behara, David Dotson, Raimondas Galvelis, John Herr, Josh Horton, Yuezhi Mao, John Chodera, Benjamin Pritchard, Yuanqing Wang, Gianni De Fabritiis, and Thomas Markland. Spice 2.0.1, 2024

    work page 2024

  29. [30]

    Narbe Mardirossian and Martin Head-Gordon.ωB97M-V: A combinatorially optimized, range-separated hybrid, meta-GGA density functional with VV10 nonlocal correlation.The Journal of Chemical Physics, 144(21):214110, 2016

    work page 2016

  30. [31]

    Balanced basis sets of split valence, triple zeta valence and quadruple zeta valence quality for h to rn: Design and assessment of accuracy

    Florian Weigend and Reinhart Ahlrichs. Balanced basis sets of split valence, triple zeta valence and quadruple zeta valence quality for h to rn: Design and assessment of accuracy. Physical Chemistry Chemical Physics, 7(18):3297–3305, 2005

    work page 2005

  31. [32]

    Stefan Grimme, Jens Antony, Stephan Ehrlich, and Helge Krieg. A consistent and accurate ab initio parametrization of density functional dispersion correction (DFT-D) for the 94 elements H-Pu.The Journal of Chemical Physics, 132(15):154104, 2010

    work page 2010

  32. [33]

    Birgin, and Jos´ e Mario Mart´ ınez

    Leandro Mart´ ınez, Ricardo Andrade, Ernesto G. Birgin, and Jos´ e Mario Mart´ ınez. Pack- mol: A package for building initial configurations for molecular dynamics simulations. Journal of Computational Chemistry, 30(13):2157–2164, 2009

    work page 2009

  33. [34]

    Chodera, Robert T

    Peter Eastman, Jason Swails, John D. Chodera, Robert T. McGibbon, Yutong Zhao, Kyle A. Beauchamp, Lee-Ping Wang, Andrew C. Simmonett, Matthew P. Felty, Michael E. Morozov, et al. OpenMM 7: Rapid development of high performance algorithms for molec- ular dynamics.PLoS Computational Biology, 13(7):e1005659, 2017

    work page 2017

  34. [35]

    Denning, Thomas B

    Naveen Michaud-Agrawal, Elizabeth J. Denning, Thomas B. Woolf, and Oliver Beckstein. MDAnalysis: a Python package for the rapid analysis of molecular dynamics simulations. Journal of Computational Chemistry, 32(10):2319–2327, 2011

    work page 2011

  35. [36]

    Introducing gpu-acceleration into the python-based simulations of chemistry framework, 2024

    Rui Li, Qiming Sun, Xing Zhang, and Garnet Kin-Lic Chan. Introducing gpu-acceleration into the python-based simulations of chemistry framework, 2024

    work page 2024

  36. [37]

    Castelli, Rune Christensen, Marcin Du lak, Jesper Friis, Michael N

    Ask Hjorth Larsen, Jens Jørgen Mortensen, Jakob Blomqvist, Ivano E. Castelli, Rune Christensen, Marcin Du lak, Jesper Friis, Michael N. Groves, Bjørk Hammer, Cory Hargus, et al. The atomic simulation environment—a Python library for working with atoms. Journal of Physics: Condensed Matter, 29(27):273002, 2017

    work page 2017

  37. [38]

    Schnet: A continuous-filter convolu- tional neural network for modeling quantum interactions.Advances in Neural Information Processing Systems, 30, 2017

    Kristof Sch¨ utt, Pieter-Jan Kindermans, Huziel Enoc Sauceda Felix, Stefan Chmiela, Alexandre Tkatchenko, and Klaus-Robert M¨ uller. Schnet: A continuous-filter convolu- tional neural network for modeling quantum interactions.Advances in Neural Information Processing Systems, 30, 2017

    work page 2017

  38. [39]

    Directional message passing for molecular graphs

    Johannes Gasteiger, Janek Groß, and Stephan G¨ unnemann. Directional message passing for molecular graphs. InInternational Conference on Learning Representations, 2020

    work page 2020

  39. [40]

    Owen, Mordechai Kornbluth, and Boris Kozinsky

    Albert Musaelian, Simon Batzner, Anders Johansson, Lixin Sun, Cameron J. Owen, Mordechai Kornbluth, and Boris Kozinsky. Learning local equivariant representations for large-scale atomistic dynamics.Nature Communications, 14(1):579, 2023. 35 April 14, 2026UBio-MolFM Technical Report

    work page 2023

  40. [41]

    Equiformer: Equivariant graph attention transformer for 3d atomistic graphs

    Yi-Lun Liao and Tess Smidt. Equiformer: Equivariant graph attention transformer for 3d atomistic graphs. InInternational Conference on Learning Representations, 2023

    work page 2023

  41. [42]

    Equiformerv2: Improved equivariant transformer for scaling to higher-degree representations

    Yi-Lun Liao, Brandon Wood, Abhishek Das, and Tess Smidt. Equiformerv2: Im- proved equivariant transformer for scaling to higher-degree representations.arXiv preprint arXiv:2306.12059, 2023

    work page arXiv 2023

  42. [43]

    Uni-mol: A universal 3d molecular representation learning framework

    Gengmo Zhou, Zhifeng Gao, Qiankun Ding, Hang Zheng, Hongteng Xu, Zhewei Wei, Lin- feng Zhang, and Guolin Ke. Uni-mol: A universal 3d molecular representation learning framework. InInternational Conference on Learning Representations, 2023

    work page 2023

  43. [44]

    Do transformers really perform badly for graph representation? In Advances in Neural Information Processing Systems, volume 34, pages 28877–28888, 2021

    Chengxuan Ying, Tianle Cai, Shengjie Luo, Shuxin Zheng, Guolin Ke, Di He, Yanming Chen, and Tie-Yan Liu. Do transformers really perform badly for graph representation? In Advances in Neural Information Processing Systems, volume 34, pages 28877–28888, 2021

    work page 2021

  44. [45]

    DPA-2: A large atomic model as a multi-task learner, 2023

    Duo Zhang et al. DPA-2: A large atomic model as a multi-task learner, 2023

    work page 2023

  45. [46]

    Harry Moore, Nicholas J

    D´ avid P´ eter Kov´ acs, J. Harry Moore, Nicholas J. Browning, Ilyes Batatia, Joshua T. Horton, Venkat Kapil, William C. Witt, Ioan-Bogdan Magd` au, Daniel J. Cole, and G´ abor Cs´ anyi. Mace-off23: Transferable machine learning force fields for organic molecules, 2023

    work page 2023

  46. [47]

    Smith, Berk Hess, and Erik Lindahl

    Mark James Abraham, Teemu Murtola, Roland Schulz, Szil´ ard P´ all, Jeremy C. Smith, Berk Hess, and Erik Lindahl. GROMACS: High performance molecular simulations through multi-level parallelism from laptops to supercomputers.SoftwareX, 1:19–25, 2015

    work page 2015

  47. [48]

    Fast parallel algorithms for short-range molecular dynamics.Journal of Computational Physics, 117(1):1–19, 1995

    Steve Plimpton. Fast parallel algorithms for short-range molecular dynamics.Journal of Computational Physics, 117(1):1–19, 1995

    work page 1995

  48. [49]

    Phillips, Rosemary Braun, Wei Wang, James Gumbart, Emad Tajkhorshid, Eliza- beth Villa, Christophe Chipot, Robert B

    James C. Phillips, Rosemary Braun, Wei Wang, James Gumbart, Emad Tajkhorshid, Eliza- beth Villa, Christophe Chipot, Robert B. Skeel, Laxmikant Kal´ e, and Klaus Schulten. Scal- able molecular dynamics with NAMD.Journal of Computational Chemistry, 26(16):1781– 1802, 2005. 36

    work page 2005