SciReasoner - A multimodal foundational model for scientific reasoning developed by Shanghai AI Lab and others. - AiBoss

What is SciReasoner?

SciReasoner is a next-generation multimodal scientific reasoning model developed by the Shanghai Artificial Intelligence Laboratory in collaboration with teams from the Chinese University of Hong Kong, the Institute of Materia Medica of the Chinese Academy of Sciences, Stanford University, and Oxford University. The model is the first to propose a native structural reasoning paradigm, transforming protein 3D structures, molecular topology, and crystal lattices into addressable units of reasoning evidence, enabling AI to organize its analysis around real structural evidence, much like scientists do. Based on continuous pre-training on Qwen3, the model achieves state-of-the-art (SOTA) results in 67 out of 86 interdisciplinary benchmarks, covering protein functional annotation, retrosynthetic planning, drug screening, and material property prediction, thus making a leap from black-box prediction to an interpretable scientific framework.

SciReasoner's main functions

Protein Functional AnnotationPredicting Gene Ontology terms based on 3D structure (Foldseek 3Di representation) is particularly effective for low-homology and orphan proteins.
Inverse Synthesis ProgrammingThe goal is to analyze the product skeleton and functional groups, identify key bond breaking sites, verify the reactant structure, and determine the feasibility of the reaction.
Molecular virtual screeningDrug molecule screening is based on the similarity of three-dimensional drug clusters, eliminating the need for additional docking scoring or task fine-tuning.
Material property prediction: Analyze crystal space group, coordination environment and periodicity to predict physicochemical properties such as formation energy and band gap.
Science Q&A and GenerationSupports open-ended scientific Q&A and molecular design across proteins, DNA/RNA, small molecules, and crystals.

SciReasoner's technical principles

Unified Structure-Aware VocabularyDifferent scientific objects are discretized into structure-aware tokens: ConfSeq for small molecules, Foldseek 3Di for proteins, and SLICES for crystals, avoiding the destruction of chemical semantics by traditional word segmentation.
Three-stage continuous pre-trainingBased on Qwen3, we perform Warm-up training, full-parameter multimodal training, and annealing training to establish a stable structure-language interface.
Post-training inference frameworkIt employs bootstrapping native structural reasoning, ground-training protein/molecule/material experts through domain-specific structural evidence, and then unifying expert capabilities through cross-domain reasoning integration, using reinforcement learning to connect structural vocabulary semantics and thought chain strategies.
Censorable reasoning trajectoryModel output <think.../think> The reasoning chain of the package, with its central claim traceable to a specific structural token, realizes the transformation from black-box prediction to an explainable scientific chain.

How to use SciReasoner

Obtain resourcesAccess the GitHub repository, clone the code, and download the model weights.
Installation EnvironmentConfigure Python dependencies and runtime environment according to the README.
Prepare dataOrganize structural inputs such as protein PDB/3Di, molecular SMILES/ConfSeq, or crystal CIF/SLICES according to the task type.
Loading ModelInitialize the SciReasoner model and connect it to the structure-aware vocabulary.
Input command: Combine structured data with natural language questions and feed them into the model.
Obtaining reasoningModel output band <think.../think> The censorable inference chain of the tags and the final prediction results.
Verification resultsThe structure tokens in the reasoning trajectory are mapped back to the three-dimensional structure for expert review.

SciReasoner's core advantages

Original structural reasoningThe structural token is not an auxiliary input, but the core evidence unit for reasoning. The reasoning process can revolve around spatial relationships, coordination environment, symmetry, etc.
Significant advantages in low homogeneity scenariosIn protein function prediction with sequence similarity below 30%, Cellular Component Fmax improved from 0.42 to 0.55, far exceeding BLAST and ESM2.
Interdisciplinary Unified CapabilitiesThe single model covers three major fields: biology, chemistry, and materials science, achieving state-of-the-art (SOTA) results in 67 out of 86 benchmarks, breaking the barriers of traditional dedicated models.
Explainability LeadingIn double-blind expert evaluation, 98% of the cases considered its reasoning trajectory to be superior to or comparable to cutting-edge large language models, and the reasoning chain could be directly mapped to the three-dimensional structural region.
High accuracy in chemical reasoningThe USPTO-50K retrosynthesis achieved a Top-1 accuracy of 72.0%, significantly exceeding the previous best method, RSGPT, at 63.4%.

SciReasoner's project address

Project official website：https://scireasoner.github.io/
GitHub repository：https://github.com/SpectrAI-Initiative/SciReasoner
arXiv technical paper：https://arxiv.org/pdf/2607.07708

Comparison of SciReasoner's similar products

Dimension	SciReasoner	ESM2 (Meta)
position	Interdisciplinary Scientific Reasoning Foundation Model	Protein Language Model
Input mode	Protein structure/sequence, molecule, crystal	Protein sequence only
Reasoning Paradigm	Original structure reasoning, trajectory can be examined.	Sequence embedding prediction, black box output
Low homology protein	CC Fmax 0.55 (≤30% similarity)	CC Fmax 0.42
Explainability	Inference chains can be mapped to structured tokens.	Attention heatmap, without natural language reasoning
cross-domain	Unified model of biology, chemistry and materials	Protein field only

Application scenarios of SciReasoner

orphan protein function discoveryFor unknown proteins lacking homologous sequences, their molecular function and subcellular localization can be inferred through their three-dimensional folding structure.
Drug retrosynthetic designIt assists chemists in planning synthetic routes from target molecules to commercially available precursors, and in identifying bond-breaking sites and reaction conditions.
New material screeningBased on the prediction of material stability and electronic properties using crystal structure, we can accelerate the research and development of functional materials such as battery materials and catalysts.
AI-assisted science educationIt utilizes its interpretable reasoning trajectory to help students understand the scientific thinking process of "structure determines properties".
Cross-modal science discoveryIt integrates multimodal tasks such as protein-small molecule interactions and RNA function prediction to support systems biology research.