SciReasoner

Structure-aware scientific foundation model for proteins, molecules, nucleic acids, and materials.

SciReasoner is a multimodal scientific foundation model for native structural reasoning. It turns protein structures, 3D molecules, crystals, sequences, formulas, and text into structure-aware evidence tokens, then reasons over those tokens for scientific QA, prediction, classification, and generation/design tasks.

Highlights

• Broad benchmark leadership: state-of-the-art on 67 / 86 evaluated tasks and best-performing model against generalist LLM baselines on 75 / 86 tasks.
• Specialist-level accuracy: matches or surpasses published domain specialists on 26 / 33 specialist-baseline comparisons.
• Native structural reasoning: uses Foldseek 3Di, ConfSeq, and SLICES structural encodings instead of relying only on text serialization.
• Cross-domain scope: covers proteins, DNA/RNA, small molecules, 3D molecular structures, and inorganic crystals in one model.
• Inspectable traces: double-blind experts rated SciReasoner traces as preferred or comparable to DeepSeek-V4-Pro in 98% of judgments.

News

• 2026-07-08: SciReasoner preprint released on arXiv: arXiv:2607.07708.

Model

SciReasoner is initialized from Qwen3-14B and aligned with domain-native structural vocabularies. Structural tokens are treated as addressable evidence units, allowing the model to connect residues, fragments, conformers, symmetry, coordination environments, and periodic bonding patterns to downstream scientific properties.

Model Scope

Area	Inputs	Task coverage
Chemistry	SMILES, IUPAC names, molecular formulas, text descriptions, ConfSeq 3D molecular tokens	Scientific QA and extraction: chemical entity recognition, chemical-protein interaction extraction, chemical-disease interaction extraction, multiple-choice, true/false, open QA, name conversion, molecular description generation, molecule captioning; property prediction: ESOL, DUD-E, LIPO, physicochemical prediction; classification: BBBP, ClinTox, HIV Prediction, SIDER; generation/design: forward synthesis, forward reaction prediction, reagent prediction, retrosynthesis mol/USPTO-50K/smol, molecule generation, description-guided molecule design.
Materials	chemical formulas/compositions, text descriptions, crystal structures/CIF-derived descriptions, SLICES tokens	Property prediction: MP regression, SNUMAT regression, JARVIS-DFT, JARVIS-QETB, GNoME, hMOF, Cantor HEA, QMOF, OQMD, OMDB; classification: MP classification, SNUMAT classification; generation/design: composition material, bulk modulus material.
Protein	amino-acid sequences, text descriptions, functional annotations, Foldseek 3Di protein-structure tokens	Scientific QA: Function, General function; property prediction: fluorescence, stability, structural similarity, TM-score; classification/annotation: solubility, antibody-antigen interaction, yeast/human PPI, protein function, domain motif, fold type, subcellular localization, EC number, keywords, metal ion binding, GO-BP, GO-CC, GO-MF; generation/design: function-guided protein design, catalytic activity.
DNA/RNA and genomics	nucleotide sequences, text descriptions, gene/sequence metadata	Property prediction: enhancer activity, isoform, mean ribosome loading, programmable RNA switches, CRISPR on target, siRNA efficiency; classification: gSymbol2Tissue, gName2Cancer, gSymbol2Cancer, RNA protein interaction, epigenetic marks, TF-m, TF-h, enhancer-promoter interaction, PD-prom 300 all/notata/tata, CPD-prom core all/notata/tata, non-coding RNA family, modification.

Evaluation

Evaluation scripts and processed benchmark configs will be added with the public code release. The figure and specialist table below report the paper results across generalist and specialist comparisons.

Full Specialist-Baseline Results

Category	Task	Metric	Specialist baseline	SciReasoner
Generation & Design	Retrosynthesis USPTO-50K	Exact Match ↑	0.63 RSGPT	0.72
Prediction	Fluorescence	Spearman ↑	0.70 SaprotHub	0.77
Prediction	Isoform	R2 ↑	0.59 APARENT	0.86
Prediction	TM-score	Spearman ↑	0.83 SaprotHub	0.83
Prediction	ESOL	RMSE ↓	1.11 MolCLR	1.03
Prediction	GNoME	MAD/MAE ↑	15.60 LLM-Prop	21.91
Prediction	QMOF	MAD/MAE ↑	1.96 LLM-Prop	8.61
Prediction	MP regression	MAD/MAE ↑	4.39 LLM-Prop	5.83
Prediction	JARVIS-DFT	MAD/MAE ↑	2.91 LLM-Prop	5.67
Prediction	SNUMAT regression	MAD/MAE ↑	1.51 LLM-Prop	2.26
Prediction	hMOF	MAD/MAE ↑	1.48 LLM-Prop	1.67
Prediction	OQMD	MAD/MAE ↑	6.02 LLM-Prop	7.22
Prediction	OMDB	MAD/MAE ↑	1.51 LLM-Prop	1.50
Prediction	DUD-E	5.0% EF ↑	7.12 ConfSeq	7.70
Prediction	Cantor HEA	MAD/MAE ↑	8.40 LLM-Prop	7.79
Prediction	LIPO	RMSE ↓	0.65 MolCLR	0.80
Classification	BBBP	ACC ↑	0.74 MolCLR	0.84
Classification	ClinTox	ACC ↑	0.93 MolCLR	0.95
Classification	HIV Prediction	ACC ↑	0.81 MolCLR	0.92
Classification	SIDER	ACC ↑	0.68 MolCLR	0.74
Classification	MP classification	AUC ↑	0.72 LLM-Prop	0.73
Classification	TF-m	MCC ↑	0.57 NT	0.64
Classification	PD-prom 300 all	MCC ↑	0.91 NT	0.89
Classification	CPD-prom core all	MCC ↑	0.67 NT	0.68
Classification	RNA protein interaction	MCC ↑	0.74 RPI-Pred	0.81
Classification	Non-coding RNA family	ACC ↑	0.89 RNA-MSM	0.90
Classification	GO-BP	Fmax ↑	0.49 SaprotHub	0.52
Classification	GO-CC	Fmax ↑	0.48 SaprotHub	0.58
Classification	GO-MF	Fmax ↑	0.67 SaprotHub	0.66
Classification	Human PPI	ACC ↑	0.77 ESM2	0.73
Classification	Subcellular localization	ACC ↑	0.84 ESM2	0.88
Classification	Metal ion binding	ACC ↑	0.73 ESM2	0.74
Classification	Solubility	ACC ↑	0.77 DeepLoc	0.72

Reasoning Examples

SciReasoner is designed to expose intermediate structural evidence, not only final predictions.

Protein Function Annotation

Structure-aware prompting lets the model reason over secondary-structure segments, binding-pocket evidence, and motif-level cues before predicting protein function.

Materials Property Prediction

For crystalline materials, SciReasoner can use space-group information, atom-index connectivity, unit-cell boundaries, and bonding patterns to support property estimates.

Citation

If you find SciReasoner useful, please cite:

@misc{tang2026scireasoner,
  title={Accurate, Interdisciplinary and Transparent Structure-property Understanding with Deep Native Structural Reasoning},
  author={Tang, Chen and Wang, Yizhou and Wu, Jianyu and Wang, Lintao and Tang, Shixiang and Li, Pengze and Su, Encheng and Yao, Jun and Xiao, Jiabei and Shi, Yuqi and Li, Jielan and Hao, Hongxia and Gao, Zhangyang and Wu, Fang and Fei, Ben and Yue, Xiangyu and Tan, Pan and Zhong, Bozitao and Zhang, Jinouwen and Wang, Aoran and Lu, Yan and Liu, Jiaheng and Ma, Xinzhu and Hong, Liang and Zheng, Mingyue and Torr, Phil and Zhou, Bowen and Ouyang, Wanli and Bai, Lei},
  year={2026},
  eprint={2607.07708},
  archivePrefix={arXiv},
  primaryClass={cs.CL},
  url={https://arxiv.org/abs/2607.07708}
}

Contact

For updates, visit scireasoner.github.io or follow this repository.