{"id":"b3d5e559-7903-4b0f-98cc-5be1b6c37188","arxiv_id":"2611.06237","paper_version":1,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"SO3LR Tauri achieves consistent near-quantum mechanical accuracy across biomolecular classes and 85 elements using a fast neural network backbone with explicit long-range physics, without system-specific fitting.","lead":"SO3LR Tauri is a family of machine-learned force fields trained on 148 million molecular structures across 85 elements, designed to simulate proteins, nucleic acids, lipids, carbohydrates, and reactive chemistry with near-quantum accuracy. It aims to replace the fragmented set of classical force fields with a single, consistent model that runs on a single GPU for tens of millions of atoms.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek/deepseek-v4-flash","headline":"The 47% MRE in liquid self-diffusion is attributed to nuclear quantum effects without verification; if the model's intrinsic error dominates, the central claim of near-QM accuracy for dynamics is undermined.","rationale":"The reader identified the omission of nuclear quantum effects as an unverified assumption, which overlaps with this concern. However, the reader's weakest assumption also flagged the multi-head training scheme and mixed reference data. I focus specifically on the liquid diffusion error attribution because it is the single most concrete, quantifiable, and impactful weakness that directly tests the central claim of 'near-quantum-mechanical accuracy' in dynamics. The paper's open code and data allow this test to be performed independently, and the concern does not require re-deriving theorems or questioning published results. The test is computationally feasible (~400 GPU-hours per approach) and would settle whether the dynamical limitation is fundamental or correctable by including NQE. I recommend UNCHANGED because the paper is transparent about limitations, and the concern does not invalidate the static benchmarks or the framework's general promise; it merely demands stronger evidence for the dynamical claim. The verdict remains ACCEPT but with a higher correctness risk noted if the test reveals intrinsic errors.","tokens_in":67007,"tokens_out":3748,"duration_ms":51499,"concrete_test":"Perform classical MD and path-integral MD (PIMD) simulations of liquid water (900 molecules) using SO3LR-T-M at 300 K and 1 atm for 20 ns per replica (four replicas). Compute self-diffusion coefficients via the Einstein relation with Yeh–Hummer correction. Compare the classical and PIMD diffusion coefficients with the experimental value (230 Å²/ns). If the PIMD value agrees with experiment to within 20% while the classical value deviates by >40%, the NQE attribution is supported. If both deviate by >30%, the model has intrinsic dynamics errors. As a lightweight alternative, estimate the NQE contribution using a well-calibrated quantum water model (e.g., q-TIP4P/F) at the same thermodynamic point and subtract it from the SO3LR-T-M classical result; if the residual deviation exceeds 25%, the model error is dominant.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's strongest dynamical validation claim rests on liquid water and organic solvent self-diffusion. The 47% mean relative error (MRE) in diffusion is attributed to nuclear quantum effects (NQE) with a brief statement that 'NQE may affect the self-diffusion coefficient strongly by up to a factor of two' (main text, Liquid Properties), citing general references rather than a quantitative test on this specific model. No simulation or analysis is provided to confirm that SO3LR-T-M's classical dynamics would become accurate if NQE were included. If the actual model error dominates (e.g., due to overstructured hydrogen-bond networks that suppress diffusion beyond NQE corrections), then the claim that the model recovers 'experimental densities and diffusion' for water and organic solvents—a central pillar for biomolecular applicability—would be misleading. Given that classical force fields for water (e.g., OPLS4) achieve 28% MRE in diffusion without invoking NQE, a 47% MRE that is hand-waved to NQE weakens the argument that the model provides near-quantum-mechanical accuracy for dynamics. This is the most load-bearing concern because if the model's dynamics are systematically unreliable, its use for protein folding, ligand binding, and transport phenomena—the intended applications—is severely limited, regardless of static benchmark performance.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek/deepseek-v4-flash","summary":"The paper presents SO3LR Tauri, a family of machine-learned force fields that augment a SO3krates neural network with explicit short-range repulsion (NLH), electrostatics (PME), and environment-dependent dispersion, trained on ~148 million structures across 85 elements from mixed DFT levels (PBE0+MBD-NL and ωB97M-V). The models are evaluated on a comprehensive suite of static benchmarks (interaction energies on SAPT10k, QUID, JSCH-2005, PLA15; conformational energetics; dipoles), reactive chemistry (organocatalytic cycle, benzoin condensation, cobalamin methyl transfer), and molecular dynamics (pure-liquid densities and diffusion, ion hydration, pKa values, proteins, lipids, nucleic acids, metal-containing systems). The central claim is that SO3LR Tauri is a step toward a single, universally applicable force field that can simulate biomolecular systems with near-quantum-mechanical accuracy without system-specific fitting, while scaling to tens of millions of atoms on one GPU.","tokens_in":67484,"tokens_out":1782,"duration_ms":24907,"significance":"If the claimed accuracy holds, this work represents a major advance in the unification of biomolecular force fields. The combination of a fast equivariant backbone with physics-based long-range terms trained on diverse data covering 85 elements is novel in scope. The static benchmarks, especially the sub-1 kcal/mol interaction-energy errors on SAPT10k and QUID, and the correct stereoselectivity in the aldol reaction, are impressive. The demonstration that a single pretrained model can simultaneously describe proteins, nucleic acids, lipids, carbohydrates, ions, and reactive chemistry without reparametrization is a significant step toward the long-standing goal of a universal MLFF. However, the dynamical validation shows clear limitations (47% MRE in self-diffusion, over-structured water), and the attribution of these errors to nuclear quantum effects is asserted but not tested, which tempers the significance until these issues are resolved.","major_comments":[{"comment":"The 47% mean relative error in self-diffusion for the 16 neat solvents is the largest quantitative discrepancy in the dynamics benchmarks. The paper attributes this to nuclear quantum effects (NQE) with a brief statement that NQE 'may affect the self-diffusion coefficient strongly by up to a factor of two,' citing general references without any test specific to SO3LR-T-M. Since classical force fields (e.g., OPLS4) achieve 28% MRE without invoking NQE, the possibility that the model's intrinsic error dominates is real. The authors should either provide a quantitative estimate of the NQE contribution for this specific model (e.g., via a path-integral simulation on a subset, or by comparing to experimental isotope effects) or add a more cautious interpretation that acknowledges the unresolved error. This is load-bearing for the central claim that dynamics reach 'near-quantum-mechanical accuracy.'","section":"§Liquid Properties, Fig. 4B","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The water oxygen–oxygen radial distribution function shows a first-peak height of 3.05 against an experimental 2.58. This over-structuring is consistent with suppressed diffusion (119 vs 230 Å²/ns). The authors attribute this to residual PBE0 bias corrected only by late-stage fine-tuning. The connection between structural over-structuring and the 47% diffusion error should be discussed quantitatively: what fraction of the diffusion error can be explained by the over-structured H-bond network alone, and how much remains for NQE? This would make the NQE discussion more concrete.","section":"§Water properties, Fig. S11","authors_acknowledge":null,"authors_acknowledge_p":null}],"minor_comments":[{"comment":"The radar chart in Fig. 1B is visually appealing but the axes are not numerically labelled; consider adding values or at least defining the scale in the caption.","section":"§TOC graphic","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The sentence 'The neural network takes the atomic positions R, atomic numbers Z, total charge Q, and spin multiplicity S' could be clarified: are Q and S supplied as scalar inputs to the first layer or as global feature conditioning? A brief sentence on how they enter the backbone would help.","section":"§Model Architecture","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"Table S3 reports sampling weights relative to OPoly26 but does not define the weight for OPoly26 itself. Clarify that OPoly26 has weight 1.0.","section":"§Training data, Table S3","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The sentence 'Only the aldol saddles were found blindly; benzoin and cobalamin were refined from published structures' (line after Fig. 3) is important for reproducibility. Consider moving it to the main text rather than the figure caption, and explicitly state whether the benzoin TS search was initialized from the published geometry or re-discovered de novo.","section":"§Reactive chemistry","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The IRC convergence tolerance is given as 'force tolerance 0.1 eV/Å' but the main text references energies in kcal/mol. For consistency, consider reporting in a single unit or providing conversion.","section":"§Supplementary §S4.1","authors_acknowledge":null,"authors_acknowledge_p":null}],"recommendation":"minor_revision"},"author_rebuttal":null,"desk_editor":{"model":"deepseek/deepseek-v4-flash","letter":"This is a serious engineering effort. The team has assembled 148M structures across 85 elements, trained through a multi-head scheme that mixes PBE0+MBD-NL and ωB97M-V, and released three model sizes with open code and data. The static benchmarks are genuinely strong: SAPT10k MAE of 0.45 kcal/mol, QUID 1.24, GlcMal conformer ordering with ρ=0.89, and dipoles that approach CCSD quality. The pKa calculations (MAE 0.98 pK units) and the demonstration of reactive saddle points—particularly the blind aldol cycle—are real steps forward. The coverage of metal-organic systems, even with higher errors, shows ambition that goes beyond the usual biomolecular scope.\n\nThe soft spot is where the paper will be used: dynamics. Liquid densities have a 6.3% MRE, and the 47% MRE in self-diffusion is handwaved to nuclear quantum effects (NQE) with a single citation but zero verification. The water oxygen-oxygen RDF peak is 3.05 versus experimental 2.58—over-structured, which directly suppresses diffusion. The stress test is fair: if the model's intrinsic error is driving the diffusion error, then the claim of near-QM accuracy for dynamics is misleading. The authors openly list this as unresolved (they say 'much work remains'), but the abstract and introduction pitch the model as recovering experimental densities and diffusion, which overstates what is actually shown. The reactivity study is also not entirely blind—only the aldol cycle was searched de novo; benzoin and cobalamin were seeded from published structures, which the authors honestly note.\n\nThis paper belongs in a good journal. It pushes the field forward on static properties and chemical coverage. But readers who care about dynamics should treat the transport numbers as a known limitation, not a solved problem. I'd recommend acceptance with a request to substantially tone down the dynamic claims, or to add explicit NQE simulations or model refinements.","headline":"A landmark ML force field for biomolecules, with impressive static accuracy and broad coverage, but the dynamic validation is weaker than claimed—the 47% liquid diffusion error attributed to nuclear quantum effects without testing is a real gap.","tokens_in":67985,"tokens_out":1546,"duration_ms":21575},"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek/deepseek-v4-flash","headline":"SO3LR Tauri is a pre-trained machine-learned force field that simulates (bio)molecules across 85 elements and reactive chemistry at near-DFT accuracy without system-specific fitting, scaling to tens of millions of atoms on a single GPU.","keywords":["machine-learned force field","biomolecular simulation","SO3krates","long-range interactions","transferability","periodic table","molecular dynamics","equivariant neural network"],"falsifier":"Simulate a set of 10 small proteins and 10 RNA loops for 1 microsecond each with SO3LR-T-M and compare the resulting folded-state populations or RNA secondary structure against experiment or explicit-solvent QM/MM references; significant deviations would undermine the claim of general biomolecular applicability. Alternatively, measure liquid self-diffusion coefficients for 20 organic solvents under the same protocol: a mean relative error exceeding 60% would indicate a systematic failure to capture dynamical transport properties.","tokens_in":68521,"feed_emoji":"🧬","tokens_out":4378,"duration_ms":57374,"temperature":0.7,"pith_summary":"The paper presents SO3LR Tauri, a family of machine-learned force fields that aim to be a single, unified model for simulating all major classes of biomolecules—proteins, nucleic acids, lipids, carbohydrates—as well as reactive chemistry, at near-quantum-mechanical accuracy. The models combine a fast equivariant message-passing neural network (SO3krates) with explicit physics-based terms for short-range repulsion (pre-tabulated NLH potentials), electrostatics (particle-mesh Ewald), and dispersion (environment-dependent C6/C8/C10 via quantum Drude oscillators). Trained on 148 million structures spanning 85 elements from seven datasets at two DFT levels (PBE0+MBD-NL and ωB97M-V), the model reproduces static benchmarks (interaction energies, conformational energetics, dipoles), locates transition states for organic and enzymatic reactions, and recovers condensed-phase properties (densities, ion hydration, peptide pKa, protein/DNA/membrane structure) from unmodified molecular dynamics. The work is presented as a step toward a universal force field, with openly discussed limitations including liquid diffusion errors (47% MRE), water dynamics, and the need for nuclear quantum effects.","feed_headline":"Single ML force field covers 85 elements, no fitting needed","feed_subtitle":"SO3LR Tauri simulates proteins, DNA, membranes, and reactions at near-DFT accuracy, scaling to tens of millions of atoms on one GPU.","key_machinery":"The key machinery is the SO3LR Tauri architecture: a SO3krates equivariant message-passing neural network that predicts per-atom energies, partial charges, polarizability ratios, and C6 ratios, which parametrize three explicit physics-based terms. The short-range repulsion uses pre-tabulated, pair-specific screened Coulomb potentials (NLH) from dedicated DFT calculations, with no trainable parameters. Electrostatics uses damped pairwise Coulomb interactions with particle-mesh Ewald summation for periodic systems. Dispersion is modelled with environment-dependent atomic polarizabilities and C6 coefficients (from two learned ratios per atom), pairwise Casimir–Polder mixing, and higher-order C8/C10 via quantum Drude oscillator relations, with the leading C6 term also evaluated by PME. The network is trained on 148M structures from seven datasets at two DFT levels (PBE0+MBD-NL and ωB97M-V) using a multi-head scheme: separate heads per level of theory in early stages, then a single production head fine-tuned on both levels.","core_discovery":"The core claim is that SO3LR Tauri can serve as a single, pre-trained machine-learned force field that is generally applicable across the periodic table (elements 1–85) and across biomolecular classes, achieving near-DFT accuracy for static properties and recovering key condensed-phase observables from unmodified simulations, without any system-specific parametrization. The discovery is that a combination of a fast equivariant backbone (SO3krates) with separate physics-based modules for short-range repulsion (fixed pair-specific NLH potentials), electrostatics (PME), and environment-dependent dispersion (C6/C8/C10 with learned per-atom ratios) can be trained jointly on a diverse 148M-structure dataset to yield a model that transfers across chemical spaces and scales to tens of millions of atoms on a single GPU.","pith_inferences":["The multi-head training scheme that combines data from two DFT levels could become a standard strategy for training foundation force fields on heterogeneous quantum-chemistry datasets, because it allows the backbone to learn a representation that captures physics common to both functionals.","The paper's explicit discussion of limitations—particularly liquid diffusion, water structuring, reaction barrier accuracy, and nuclear quantum effects—sets a constructive precedent and should guide benchmark design for the next generation of universal force fields.","The use of pre-tabulated NLH repulsion (rather than learning short-range repulsion from data) may be a generally applicable technique to improve stability in out-of-distribution configurations encountered in high-temperature, high-pressure, or generative-AI-proposed geometries."],"forward_implications":["If correct, a single pre-trained model can replace separate classical force fields for proteins, lipids, nucleic acids, carbohydrates, and water, eliminating cross-parameter inconsistencies and system-specific fitting.","The model's ability to locate transition states and predict reaction barriers (MAE 2.9 kcal/mol for an aldol cycle) suggests it could serve as a fast surrogate for DFT in enzyme mechanism studies and drug discovery.","Scaling to 84 million atoms on one GPU enables all-atom simulations of large biomolecular assemblies (e.g., organelles or viral capsids) with near-DFT accuracy, though the authors note nanosecond sampling remains a bottleneck.","The reported 47% mean relative error in liquid self-diffusion indicates that static benchmark accuracy does not guarantee correct dynamics, highlighting a key area for improvement in future general-purpose MLFFs."],"supporting_citations":[{"why":"Provides the SO3krates equivariant message-passing backbone that is the core of the neural network.","marker":"[18]"},{"why":"Describes the predecessor SO3LR model, establishing the baseline architecture and performance that SO3LR Tauri extends.","marker":"[19]"},{"why":"Supplies the QCML dataset (33.5M structures at PBE0+MBD-NL) with per-atom dispersion labels, a major component of the training corpus.","marker":"[20]"},{"why":"Supplies the QCell dataset (526k structures at PBE0+MBD(-NL)) covering diverse biomolecular fragments, used for training and testing.","marker":"[21]"},{"why":"Supplies the OMol25 dataset (100M structures at ωB97M-V) covering proteins, nucleic acids, electrolytes, and metal-organic systems, used for training and testing.","marker":"[22]"},{"why":"Supplies the OPoly26 dataset (6.6M structures at ωB97M-V) covering polymers and lipids, used for training and testing.","marker":"[23]"},{"why":"Supplies the pre-tabulated Nordlund–Lehtola–Hobler (NLH) pair-specific short-range repulsion potentials that replace the learned ZBL term.","marker":"[31]"},{"why":"Provides the JAX-PME implementation for particle-mesh Ewald summation of long-range electrostatics and dispersion.","marker":"[33]"},{"why":"Establishes the quantum Drude oscillator (QDO) relations that govern the higher-order dispersion coefficients C8 and C10.","marker":"[37]"}],"fun_headline_variants":["One MLFF for 85 elements, no refitting needed","SO3LR Tauri: universal force field for biomolecules","Single pre-trained model covers 85-element periodic table","DFT accuracy across proteins, DNA, and reactions, no tuning"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The mixed quantum-mechanical reference data (PBE0+MBD-NL for most fragments and ωB97M-V for OMol25 and OPoly26) are together sufficient to represent 'quantum-mechanical accuracy' for all simulated systems, and the multi-head training scheme does not introduce systematic errors that cancel in static benchmarks but amplify in dynamics such as liquid self-diffusion (47% mean relative error).","fun_headline_variants_meta":{"raw":{"variants":["One MLFF for 85 elements, no refitting needed","SO3LR Tauri: universal force field for biomolecules","Single pre-trained model covers 85-element periodic table","DFT accuracy across proteins, DNA, and reactions, no tuning"]},"model":"deepseek/deepseek-v4-flash","effort":"low","cost_usd":9e-05,"raw_usage":{"cost":8.976408e-05,"is_byok":false,"cost_details":{"upstream_inference_cost":8.976408e-05,"upstream_inference_prompt_cost":4.988004e-05,"upstream_inference_completions_cost":3.988404e-05},"total_tokens":1397,"prompt_tokens":998,"completion_tokens":399,"prompt_tokens_details":{"audio_tokens":0,"video_tokens":0,"cached_tokens":0,"cache_write_tokens":0},"completion_tokens_details":{"audio_tokens":0,"image_tokens":0,"reasoning_tokens":328}},"tokens_in":998,"tokens_out":399,"cost_source":"openrouter_usage","duration_ms":5529,"image_pages":[],"temperature":1.0,"billed_attempts":1,"reasoning_tokens":328,"transport_provider":"openrouter","cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-11-05T13:28:07.545785+00:00","model_set":{"reader":"deepseek/deepseek-v4-flash"},"pipeline_version":"pith-pipeline@v1.4.1-daily-deepseek-or-medium","input_kind":null,"input_chars":null,"input_total_chars":null,"passes_missing":null,"published_standing":null,"published_confidence":null,"objection_audit":null,"run_record":null,"basis":{"kind":"not_recorded","source":null,"chars_given":null,"chars_total":null},"receipt":null,"falsifier":"Simulate a set of 10 small proteins and 10 RNA loops for 1 microsecond each with SO3LR-T-M and compare the resulting folded-state populations or RNA secondary structure against experiment or explicit-solvent QM/MM references; significant deviations would undermine the claim of general biomolecular applicability. Alternatively, measure liquid self-diffusion coefficients for 20 organic solvents under the same protocol: a mean relative error exceeding 60% would indicate a systematic failure to capture dynamical transport properties.","supporting_citations":[{"cited_title":"Multiscale Neural Network Potential with Anisotropic Message Passing for the Fast and Accurate Simulation of Protein Dynamics and Enzymatic Reactions","cited_arxiv_id":null,"evidence_quote":"Provides the SO3krates equivariant message-passing backbone that is the core of the neural network."},{"cited_title":"Analyzing Atomic Interactions in Molecules as Learned by Neural Networks","cited_arxiv_id":null,"evidence_quote":"Describes the predecessor SO3LR model, establishing the baseline architecture and performance that SO3LR Tauri extends."},{"cited_title":"P.; Simm, G.; Ortner, C.; Cs´ anyi, G","cited_arxiv_id":null,"evidence_quote":"Supplies the QCML dataset (33.5M structures at PBE0+MBD-NL) with per-atom dispersion labels, a major component of the training corpus."},{"cited_title":"Molecular Dynamics Simulation for All","cited_arxiv_id":null,"evidence_quote":"Supplies the QCell dataset (526k structures at PBE0+MBD(-NL)) covering diverse biomolecular fragments, used for training and testing."},{"cited_title":"Machine Learning Force Fields","cited_arxiv_id":null,"evidence_quote":"Supplies the OMol25 dataset (100M structures at ωB97M-V) covering proteins, nucleic acids, electrolytes, and metal-organic systems, used for training and testing."},{"cited_title":"SpookyNet: Learning force fields with electronic degrees of freedom and nonlocal effects","cited_arxiv_id":null,"evidence_quote":"Supplies the OPoly26 dataset (6.6M structures at ωB97M-V) covering polymers and lipids, used for training and testing."},{"cited_title":"and Woolf, Lawrence A","cited_arxiv_id":null,"evidence_quote":"Supplies the pre-tabulated Nordlund–Lehtola–Hobler (NLH) pair-specific short-range repulsion potentials that replace the learned ZBL term."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the JAX-PME implementation for particle-mesh Ewald summation of long-range electrostatics and dispersion."},{"cited_title":"NIST Chemistry WebBook, NIST Standard Reference Database 69","cited_arxiv_id":null,"evidence_quote":"Establishes the quantum Drude oscillator (QDO) relations that govern the higher-order dispersion coefficients C8 and C10."}],"lane_accuracy":{"status":"not yet measured","lane":{"model":"deepseek/deepseek-v4-flash","input_kind":null},"url":"https://pith.science/accuracy"},"review_version":1}