Accurate, Interdisciplinary and Transparent Structure-property Understanding with Deep Native Structural Reasoning

Motivation

• Scientific AI often separates native structural representations from evidence-linked reasoning, leaving a gap between language-based explanations and task-specific predictors that expose intermediate evidence.Mechanistic evidence can span local motifs, non-local contacts, chemical environments, conformational geometry and long-range periodic order, requiring scientific constraints to compare structural hypotheses.Source: §1 · p. 2

Approach

• SciReasoner discretizes coordinates, topologies and periodic connectivities into a unified structure-aware vocabulary whose tokens serve as addressable evidence in reasoning trajectories.The model interleaves explanatory text with specific residues, molecular fragments, conformational cues or crystal descriptors, enabling intermediate claims to be traced to structural evidence.Source: §1 · p. 2

Results

• Fmax rises from 0.42 to 0.55 for Cellular Component prediction in low-homology and orphan-like proteins, while single-step retrosynthesis accuracy rises from 0.63 to 0.72.The traces localize protein attention to contact-defined DNA-binding residues and decompose products into reaction-relevant fragments, strategic bond disconnections and chemically verified precursors; in crystals, they emphasize symmetry-distinct sites, coordination environments and periodic bonding motifs.Source: §1 · p. 2
• Across 86 benchmarks, SciReasoner achieves state-of-the-art performance on 67 tasks, and experts prefer or tie its reasoning traces to a frontier language model in most comparisons.The benchmarks span proteins, DNA, RNA, small molecules, inorganic crystals, scientific question answering, property prediction and generation tasks.Source: §1 · p. 2

2.1 Overview of SciReasoner

• SciReasoner outperforms previous state-of-the-art methods on 67 tasks and matches or exceeds domain specialists on 26 benchmarks with available baselines.Across the overview results, it also improves RNA-function metrics and DUD-E enrichment, while providing natural-language rationales for predictions and open-ended biomedical QA and protein-function description.Source: §2.1 · p. 5

2.2.1 SciReasoner precisely predicts protein GO term and learns biologically meaningful structure

• At ≤30% sequence identity, SciReasoner improves Cellular Component Fmax by +0.21 over BLAST and +0.13 over ESM2, while reaching overall GO Fmax 0.59.Attention aligns with DNA-binding sites (AUROC 0.91 and 4.2× top-20% enrichment for catabolite control protein A), and reinforcement learning improves GO traces, scoring 8.33 overall versus 7.77 for SciReasoner-ColdStart and 6.96 for DeepSeek-V4-Pro.Source: §2.2.1 · pp. 5, 7, Figure 2 · p. 6

2.2.2 SciReasoner formulates chemically viable retrosynthetic routes with high precision

• SciReasoner reaches 0.72 Exact Match on USPTO-50K, exceeding RSGPT at 0.63 and Opus-4.7 at 0.48.Its structural-token traces proceed through analysis, disconnection, verification, and feasibility; SciReasoner includes a correct reactant set in the Top-3 for 5/5 representative products, versus 2/5 for both RSGPT and Opus-4.7.Source: §2.2.2 · pp. 7–9, Figure 3 · p. 8

2.2.3 SciReasoner captures plausible 3D molecular structure for similarity representation

• On 102 DUD-E targets, SciReasoner’s molecular embedding matches ConfSeq’s prior-best AUC of 0.76 and raises best 5.0% EF from 7.12 to 7.70.Its embedding clusters actives by binding-pocket geometry, including chemically distinct ligands with low fingerprint similarity that adopt similar 3D pharmacophores.Source: Figure 3 · p. 8, §2.2.3 · p. 9

2.2.4 SciReasoner learns structure- and property-aware representations for materials tasks

• SciReasoner outperforms CGCNN across all ten materials tasks and LLM-Prop on most numerical properties, with formation-energy parity R2 = 0.895.Its latent space separates C, Si, and SiC and organizes polymorphs and band gaps along property gradients; a TiGaCo2 trace links symmetry and periodic-connectivity tokens to crystallographic features.Source: Figure 4 · p. 10, §2.2.4 · p. 11

2.3 Structural evidence drives native scientific reasoning

• Removing structural inputs consistently weakens performance across proteins, small molecules, and materials, while SciReasoner represents structure in a distinct embedding space from sequence.Case studies show structural tokens redirect pore-size and GO reasoning from misleading composition or sequence cues toward symmetry, coordination, connectivity, secondary structure, and binding-site evidence.Source: §2.3 · pp. 11–12

2.4 Intra-domain grounding and cross-domain consolidation stimulate self-bootstrapped native structural reasoning

• Cross-domain reasoning consolidation raises retrosynthesis pass@1 from 0.49 after intra-domain grounding to 0.72, while improving first-sample performance across four task families.Intra-domain grounding raises GO pass@1 from 0.48 to 0.52, QMOF from 2.71 to 4.76, and DUD-E from 4.65 to 6.31; consolidation further raises GO to 0.59, QMOF from 6.80 to 8.61, and DUD-E from 7.29 to 7.70. The two-stage procedure first grounds evidence within domains using expert-generated traces, then pools those traces to consolidate cross-domain reasoning and reduce pass@10–pass@1 gaps.Source: §2.4 · p. 12

2.5 Human Expert Evaluation

• Experts rated SciReasoner reasoning tie-or-better in 98% of comparisons with DeepSeek-V4-Pro across three scientific tasks.The comparison covered GO protein-function annotation, crystalline-material property prediction, and retrosynthesis; 73% strongly preferred SciReasoner, 21% preferred it, and 4% tied. Tie-or-better rates were 100%, 95%, and 98% by task, respectively.Source: §2.5 · pp. 12, 15, Figure 6 · p. 14

2.4 Intra-domain grounding and cross-domain consolidation stimulate self-bootstrapped native structural reasoning

• Self-bootstrapped reasoning shifted LLM-judged quality upward most on DUD-E (∆mean = +3.10), followed by GO (+1.79), QMOF (+1.61), and retrosynthesis (+0.48).The smaller retrosynthesis shift reflects its already high reasoning-quality distribution; the gains indicate improved coherence and evidence grounding as well as answer correctness.Source: §2.4 · p. 12, Figure 6 · p. 14
• Intra-domain evidence grounding and cross-domain consolidation raised rewards and pass@1 while reducing the pass@10–pass@1 gap, improving first-sample reliability.Task-specific experts were independently optimized to ground structural evidence, then their traces were pooled to train a unified SciReasoner that matched or exceeded expert variants across scientific tasks.Source: §2.4 · p. 12, Figure 6 · p. 14

2.5 Human Expert Evaluation

• SciReasoner averaged 8.7/10 across expert quality axes versus 4.3/10 for DeepSeek-V4-Pro, with higher means on all five axes.Axis means were 8.8 vs 4.5 for evidence grounding, 8.5 vs 4.4 for domain plausibility, 8.9 vs 4.3 for target-regime alignment, 8.7 vs 3.9 for reasoning coherence, and 8.8 vs 4.6 for anti-hallucination; all differences were significant (Wilcoxon signed-rank, P < 0.001).Source: Figure 6 · p. 14, §2.5 · p. 15

3 Discussion

• SciReasoner discretizes coordinates, molecular topologies, and crystallographic lattices into native tokens that act as addressable evidence units in reasoning trajectories.The model integrates these tokens with language instructions, making structural evidence inspectable against the physical object studied.Source: §3 · p. 15
• SciReasoner raised Cellular Component Fmax from 0.42 to 0.55 and single-step retrosynthesis accuracy from 0.63 to 0.72, while materials traces invoked symmetry, coordination, and periodic bonding.These results are highlighted in regimes where superficial similarity is insufficient.Source: §3 · p. 15
• 98% of double-blind expert evaluations preferred or tied SciReasoner over a frontier large language model, suggesting its inspectable intermediate claims aid scientific assessment.Examples include substructures supporting retrosynthetic disconnections and crystal-connectivity tokens supporting stability or electronic-response inferences.Source: §3 · p. 15
• SciReasoner unifies sequence, structure, and natural-language reasoning across scientific modalities; the authors are collecting more human judgments.The discussion frames this as a route beyond prediction toward reasoning over physical structures underlying functions, properties, and mechanisms.Source: §3 · p. 15

4.1.1 Protein Data

• Protein pretraining pairs amino-acid sequences with aligned Foldseek 3Di residue tokens, masks or replaces regions below pLDDT 70, and excludes text-linked records above 30% identity to held-out proteins.Sequences are linked to protein names, literature, and UniProtKB/Swiss-Prot annotations; AlphaFoldDB structures are used when available.Source: §4.1.1 · p. 16

4.1.2 Small-molecule Data

• Small-molecule data spans chemical text, property and assay labels, reaction and spectral resources, and 3D conformers linked to language, with benchmark-overlapping molecule–label and reaction-product examples excluded.Conformers from ChEMBL and BindingDB encode connectivity and internal-coordinate information.Source: §4.1.2 · p. 16

4.1.3 Materials Data

• Materials data links formulas, elemental composition, CIF crystal structures, generated structural descriptions, and property labels; validation and test samples are removed from continued pretraining.The material-sample-level split is 80/10/10.Source: §4.1.3 · pp. 16–17

4.1.4 DNA/RNA Data; 4.1.5 General Text and Instruction Following Data

• RNAcentral sequences and metadata, NCBI DNA fragments with organism and coordinate context, quality-filtered web text, reasoning instructions, and scientific-constraint prompts broaden training beyond the three structural domains.DNA is sampled in 1,000-bp organism-level fragments.Source: §4.1.4 · p. 17, §4.1.5 · p. 17

4.2 Model Architecture

• Foldseek, SLICES, and ConfSeq preserve protein, crystal, and molecular structure as discrete tokens, which a dedicated vocabulary embeds alongside language for autoregressive generation conditioned on both.The architecture uses a Qwen3-14B backbone and avoids continuous structural encoders during the language-model forward pass.Source: §4.2.1 · p. 17, §4.2.2 · p. 17, §4.2.3 · p. 18

4.3.1 Optimization Objective and Multi-Stage Pretraining

• A three-stage next-token curriculum first freezes the backbone to anchor structural tokens, then trains all parameters on paired structure–text data, and finally increases QA data during annealing.Stages 2 and 3 share an optimizer state and warmup-stable-decay schedule.Source: §4.3.1 · p. 18, Stage 1: Warm-up Training · p. 18, Stage 2: Full-parameter Training · p. 18

4.4 Post-training

• Post-training separates heterogeneous tasks into expert groups, replaces teacher-supplied rationales with verification-passing expert-generated traces, and pools those traces for unified coldstart and cross-domain consolidation.Coldstart supports both deliberation-formatted and direct-answer modes, with supervision loss applied only to response tokens.Source: §4.4 · p. 19, §4.4.1 · p. 19, Reasoning-formatted data construction · p. 19

4.4.2 Reinforcement learning

• Using eight independent stochastic rollouts per example, with decoding parameters held constant across seeds, the method estimates empirical solve rates, selects examples closest to 0.5, and expands outward until 2,000 examples are collected.A monotone calibration function keeps reward magnitudes comparable across tasks; prediction uses distance-based rewards, retrieval and extraction use matching-based rewards, and tool-verified rewards use professional scientific software.Source: §4.4.2 · pp. 20–21

Appendix A Detailed experimental results

• SciReasoner is compared with Opus-4.7, GPT-5.5, Kimi-K2.6, and DeepSeek-V4-Pro across benchmark tasks with available model measurements.The comparison tables cover Chemistry, Material Science, and Biology.Source: Appendix A · p. 22
• The result tables are organized by Chemistry, Material Science, and Biology, with rows grouped into Scientific QA, Property Prediction, Property Classification, or Generation and Design.Source: Appendix A · p. 22
• Bold and underlined values mark the best and second-best results, respectively, while ↑ and ↓ indicate metrics where larger and smaller values are better.Source: Appendix A · p. 22

A.1 Task and metric descriptions

• Task descriptions clarify expected model behavior and evaluation metrics using the result tables’ organization.Source: §A.1 · p. 22

Chemistry tasks

• Text and molecular-language tasks assess entity and interaction extraction, scientific question answering, SMILES/name/formula conversion, and molecule description or captioning.The listed metrics are F1 for extraction; ACC and BertScore for question answering; Split Match, Element Match, and Exact Match for name conversion; and ROUGE-L and MENTOR for descriptions and captions.Source: Chemistry tasks · p. 22
• Molecular benchmarks cover continuous-property prediction, virtual screening, molecular classification, and synthesis, reaction, reagent, retrosynthesis, and constrained-molecule generation.They include ESOL, LIPO, and physicochemical prediction (RMSE or MAE); DUD-E early enrichment at 5.0%; BBBP, ClinTox, HIV, and SIDER classification (ACC); and forward and retrosynthetic reaction tasks, reagent prediction, molecule generation, and description-guided design (Exact Match).Source: Chemistry tasks · pp. 22–23

Property Prediction

• Database-level property benchmarks report normalized MAD/MAE scores, where larger values indicate lower error relative to target dispersion.The score follows the domain LLM-Prop approach.Source: Material science tasks · p. 23
• Regression spans ten materials datasets and targets structural, electronic, energetic, porous-material, and alloy properties.The datasets are MP, SNUMAT, JARVIS-DFT, JARVIS-QETB, GNoME, hMOF, Cantor HEA, QMOF, OQMD, and OMDB.Source: Material science tasks · p. 23

Classification

• Classification covers Materials Project and SNUMAT attributes, including direct-gap status, thermodynamic stability, and direct or indirect band-gap status.Source: Material science tasks · p. 23

Material Generation

• Composition-generation tasks use SMACT to assess chemical validity or plausibility, including generation conditioned on a target bulk modulus.Source: Material science tasks · p. 23

Biology tasks.

• Function and general-function tasks generate biological descriptions or annotations and compare them with reference text using ROUGE-L.The inputs are biological context or biological sequences, respectively.Source: Biology tasks · p. 23
• Tasks assess protein and nucleic-acid predictions, including fluorescence, stability, enhancer activity, RNA processing and translation, CRISPR efficacy, and structural similarity.Evaluation metrics include Spearman correlation, HK-PCC, R2, mixed-score, and MAE.Source: Biology tasks · p. 24
• The biology benchmark includes classification and annotation tasks spanning protein properties, molecular interactions, regulatory elements, and functional labels.Examples include solubility, protein–protein interactions, transcription-factor binding, promoter detection, subcellular localization, and GO-BP, GO-CC, and GO-MF prediction.Source: Biology tasks · p. 24
• The tasks also evaluate function-guided protein sequence design and textual description of enzyme-catalyzed reactions.Design is scored by normalized sequence similarity, while catalytic-activity descriptions are compared with reference text using ROUGE-L.Source: Biology tasks · p. 24

• ACC (↑) is the fraction of samples whose predicted label exactly matches the reference label.Source: Metric definitions · p. 24
• AUC (↑) measures positive-over-negative ranking; F1 (↑) is the precision-recall harmonic mean; Fmax (↑) is maximum F1 across thresholds for multi-label annotation; MCC (↑) remains informative with imbalanced classes.Source: Metric definitions · p. 25
• RMSE (↓) is root mean squared error and penalizes larger errors; MAE (↓) is mean absolute error; MAD MAE (↑) compares target dispersion with model error.Source: Metric definitions · p. 25
• Spearman (↑) measures rank correlation; HK-PCC (↑) is Pearson correlation for housekeeping enhancer activity; R2 (↑) measures explained variance; Mixed-score (↑) combines siRNA numeric error and activity-range agreement.Source: Metric definitions · p. 25
• BertScore (↑) measures contextual-embedding semantic similarity; ROUGE-L (↑) measures longest-common-subsequence overlap; MENTOR (↑) compares generated molecule captions with references.Source: Metric definitions · p. 25
• Split Match (↑) measures component-level matching in molecular name conversion; Element Match (↑) checks formula element composition; Exact Match (↑) requires strict string or structured-answer equality.Source: Metric definitions · p. 25
• 5.0% EF (↑) measures enrichment among the top 5.0% of virtual-screening rankings; SMACT (↑) checks material validity; Normalized SW (↑) is maximum normalized Smith–Waterman alignment to reference protein sequences.Source: Metric definitions · p. 25

A.2 Detailed results

• 67 of 86 tasks: SciReasoner leads across the full benchmark suite, spanning structured prediction, generation, classification, and design-oriented tasks.The appendix groups tasks within each discipline as Scientific QA, Property Prediction, Property Classification, or Generation and Design.Source: §A.2 · p. 25
• SciReasoner shows broad gains in Chemistry, strong regression and classification in Material Science, and robust Biology performance across sequence-to-function, property, promoter, interaction, RNA, and protein tasks.The Material Science results have pronounced margins on JARVIS-QETB, GNoME, QMOF, and OQMD.Source: §A.2 · p. 25, Table A2 · p. 27, Table A3 · p. 27
• 26 of 33 comparisons: SciReasoner matches or surpasses specialist baselines.Table A1 summarizes per-task comparisons against specialist baselines.Source: §A.2 · p. 25, Table A1 · p. 26
• 75 of 86 tasks: SciReasoner was best-performing against LLM baselines (22 of 28 Chemistry, 13 of 14 Material Science, and 40 of 44 Biology tasks).Tables A2–A4 present the discipline-level task comparisons; the few non-leading results cluster in molecule captioning, selected chemistry QA and classification, bulk modulus, and some protein benchmarks.Source: §A.2 · p. 25, Table A2 · p. 27, Table A3 · p. 27

Appendix B Human Evaluation Form

• Three representative task items show an input prompt, anonymized Model A and B traces and outputs, and a read-only ground-truth fact sheet for evaluators.Evaluators score each model on five trace-quality axes and answer overall comparison questions.Source: Appendix B · p. 29

B.1 General Evaluation Instructions

• Evaluators prioritize evidence grounding, domain plausibility, target alignment, reasoning coherence, and hallucination risk over final-answer closeness alone.The five axes are scored from 1 to 10 or marked N.A. when there is no checkable claim.Source: §B.1 · p. 29
• Hallucination checks penalize unsupported fabricated or over-specific claims, including invented biological features, materials topology, or chemical entities and mechanisms.Real entities applied incorrectly are routed to domain-plausibility or target-alignment questions rather than counted as Critical fabrication.Source: §B.1 · pp. 29–30
• Grounding checks whether cited entities, tokens, positions, or input facts appear in the task input, with task-specific evidence spanning crystal representations, protein sequence and 3Di, and product chemistry.Materials checks include formula, space-group tokens, atom and edge lists, periodic offsets, and coordination counts; protein checks include sequence length, residue positions, and motifs; chemistry checks include product SMILES, functional groups, atom maps, and connectivity.Source: §B.1 · pp. 29–30
• Coherence is judged by whether evidence supports the committed conclusion, from decoded structure to property call, sequence and structural evidence to GO hypothesis, or product parsing to reactants.The form flags unresolved contradiction, circularity, padding, and unused observations as coherence defects.Source: §B.1 · pp. 29–30
• Task-specific plausibility and alignment assess crystal chemistry and property regimes, biological identity and GO region, and retrosynthetic route feasibility, reaction family, and formed bond.The form also specifies boundary handling for property regimes, GO-space alignment from correct to wrong-super-class, and borderline neighboring reaction families sharing the formed bond.Source: §B.1 · p. 30
• Four overall fields ask how each model compares with expert expectation, which model is better, and how confident the evaluator is on a 1–10 scale with a free-text note.These are recorded as Q6 through Q9.Source: §B.1 · p. 30

B.2 Blank Scoring Sheet Used for Each Sample

• The blank sheet records Q1–Q5, model scores, evidence and claim verdicts, plus an overall question rated from significantly falls short to significantly exceeds.It includes a notes field.Source: §B.2 · p. 30
• The sheet separately rates each model against expert expectations, asks for a direct A-versus-B comparison, and records evaluator confidence from 1 to 10 with a note.The direct-comparison choices range from A much better to B much better, including a tie.Source: §B.2 · pp. 30–31

B.3 Materials: Ag2HgI4, shear modulus

• For Ag2HgI4, Model A predicts 5.62 GPa against a 5.77 GPa ground truth, while Model B predicts 8.00 GPa; both are described as in the correct soft regime.The sheet's example notes that Model A's value is close to ground truth.Source: §B.3 · p. 31, Example claim prompts shown to the… · p. 32

Input prompt.

• The model prompt asks for a concise JSON prediction from Ag2HgI4's chemical formula and encoded crystal-structure information.The target property is shear modulus, gv.Source: Input prompt · p. 31

Example claim prompts shown to the evaluator.

• The evaluator-facing trace parses the atom indices into Ag, Hg, and I and lists twelve metal–iodine edges.It cites the atom list, selected edge tokens, and tetrahedral metal–iodine coordination.Source: Example claim prompts shown to the… · p. 31
• The trace links tetrahedral coordination and heavy, polarizable iodide chemistry to a compliant lattice and the target mechanical property.It treats the material as a soft iodide solid and invokes tetrahedral coordination.Source: Example claim prompts shown to the… · pp. 31–32
• Evaluators check whether the final estimate follows from cited evidence and whether named phases, database identifiers, or literature-like statements are fabricated.The example says no concrete fabricated entity is apparent when the trace stays within decoded structure and standard halide chemistry.Source: Example claim prompts shown to the… · p. 32

B.4 Gene Ontology: 1bd8 A-P55273, biological process

• F1 = 0.967 for Model A versus 0.209 for Model B; precision was 0.993 versus 0.314 and recall was 0.943 versus 0.156, respectively.Model A predicted 139 BP terms and Model B 70; the ground truth contained 145 terms.Source: §B.4 · p. 32

Input prompt.

• The input asks for biological-process terms only and supplies the protein sequence and a structure representation.Source: §B.4 · p. 32

Model outputs shown in the questionnaire.

• The displayed 70-term output is dominated by transcription and gene-expression terms, whereas the 145-term ground truth centers on cell-cycle, CDK, cell-death, DNA-damage, and stress-response processes.A separate displayed 139-term list includes regulation of cell cycle, apoptosis, DNA-damage response, and kinase activity.Source: Model outputs shown in the questionnaire · p. 32

Example claim prompts shown to the evaluator.

• Evaluator examples contrast a trace’s sequence and 3Di evidence with its cell-cycle and apoptosis predictions, while flagging bZIP, WRKY, and leucine-zipper claims for scrutiny.The examples ask evaluators to assess whether committed terms follow from grounded evidence and whether named biological entities are misassigned or fabricated.Source: Example claim prompts shown to the… · p. 33

B.5 Retrosynthesis: USPTO-50K sample 4, other

• Model A matches the gold reactants, whereas Model B proposes a related but non-gold reactant set.The example is an uncategorized USPTO-50K retrosynthesis reaction.Source: §B.5 · p. 33
• The trace parses a trifluoroacetamide, aryl sulfone, and cyclopropyl group, then disconnects the C1–N7 amide bond to propose TFAA and a primary amine.Model A’s proposed disconnection matches the gold amide-forming reaction family and formed bond.Source: Example claim prompts shown to the… · p. 34
• Model B makes the same C–N disconnection but proposes trifluoroacetic acid rather than the gold route’s more activated TFAA.The passage notes that both trifluoroacetic acid and the amine are real, and evaluators check for unsupported reaction claims or invented reagents.Source: Example claim prompts shown to the… · p. 34