SciReasoner – A Multimodal Foundation Model for Scientific Reasoning Developed by the Shanghai AI Lab and Others

1. What is SciReasoner

SciReasoner is a new-generation multimodal foundation model for scientific reasoning, jointly developed by the Shanghai Artificial Intelligence Laboratory with teams from the Chinese University of Hong Kong, the Shanghai Institute of Materia Medica of the Chinese Academy of Sciences, Stanford University, the University of Oxford, and others. This model first introduces a native structural reasoning paradigm, converting protein 3D structures, molecular topology, and crystal lattices into addressable reasoning evidence units, enabling AI to organize its analytical process around real structural evidence, just like a scientist. Based on Qwen3, the model undergoes continuous pre-training and achieves state-of-the-art (SOTA) performance on 67 out of 86 interdisciplinary benchmark tasks. These tasks include protein function annotation, retrosynthetic planning, drug screening, and material property prediction, marking a leap from black-box prediction to interpretable scientific reasoning pipelines.

Image source: Official article
Image source: official article

Technical Positioning and Domain: SciReasoner is a multimodal foundation model for scientific reasoning, spanning three core scientific domains: biology, chemistry, and materials science. Its unique positioning lies in elevating structural information from auxiliary input to the central evidence unit for reasoning, enabling the model to perform logical reasoning based on 3D spatial relationships, coordination environments, and symmetry, rather than relying solely on sequence pattern matching.

Development Background: This model was led by the Shanghai AI Lab in collaboration with multiple top universities and research institutions both domestically and internationally. The motivation for its development stems from the widespread "black-box prediction" issue in current AI systems for scientific tasks—although models can provide accurate results, they often lack an interpretable reasoning process, especially in challenging scenarios such as low homology and orphan proteins. The team aims to address this by introducing a native structural reasoning paradigm, making AI decision-making processes traceable and verifiable.

Core Value: SciReasoner resolves the fundamental flaw of traditional scientific AI models, which are unable to perform logical reasoning based on real 3D structures. Its innovative value is reflected in three key aspects: first, it unifies heterogeneous structures such as proteins, molecules, and crystals into a structure-aware vocabulary, enabling cross-modal reasoning; second, it significantly improves prediction accuracy in low-sequence-homology scenarios; and third, it outputs verifiable reasoning chains, making the scientific discovery process transparent.

Technical Features: The model employs a unified structure-aware vocabulary (ConfSeq, Foldseek 3Di, SLICES), establishing a stable structure-language interface through a three-stage continuous pre-training process. It leverages a bootstrapped native structural reasoning framework and reinforcement learning to connect the semantics of the structure vocabulary with chain-of-thought strategies. The final reasoning trajectory is wrapped in <think...> tags, with intermediate claims traceable back to specific structural Tokens.

2. Key Features

• Protein Function Annotation: Based on the Foldseek 3Di representation derived from the 3D structure of proteins, it predicts Gene Ontology terms. It performs exceptionally well in scenarios with low sequence homology (similarity below 30%) and orphan proteins, achieving a Cellular Component Fmax of 0.55, significantly surpassing traditional methods with a score of 0.42, and outperforming baseline models such as BLAST and ESM2.

• Retrosynthetic Planning: Analyzes the scaffold and functional groups of the product molecule, identifies key bond-breaking sites, and validates the structural feasibility of the reactants. On the USPTO-50K benchmark, it achieves a Top-1 accuracy of 72.0%, an improvement of nearly 9 percentage points over the previous state-of-the-art method, RSGPT, which had 63.4%. It provides chemists with interpretable synthetic route design options.

• Molecular Virtual Screening: Screens drug molecules based on 3D pharmacophore similarity, without requiring additional docking scoring or task-specific fine-tuning. The model directly utilizes molecular conformation information from the structure-aware vocabulary to quickly identify active molecules with similar spatial pharmacophore characteristics in the candidate library, significantly reducing screening costs.

• Material Property Prediction: Analyzes crystal space groups, coordination environments, and periodic connectivity to predict physical and chemical properties such as formation energy and band gap. By converting crystal structures into a sequence of tokens using the SLICES encoding, the model can understand the causal impact of atomic arrangements on macroscopic properties.

• Scientific Q&A and Generation: Supports open-ended scientific questions and molecular design across proteins, DNA/RNA, small molecules, and crystals. Users can ask questions in natural language, such as "Design an inhibitor with a similar electrostatic distribution to target A," and the model generates plausible molecular structures along with their reasoning based on structural knowledge.

• Auditable Reasoning Trajectory: The model outputs a complete reasoning chain wrapped in <think...> tags, with each intermediate claim traceable back to specific structural tokens. In a double-blind expert evaluation, 98% of cases found its reasoning trajectory to be superior or comparable to state-of-the-art large language models, providing a traceable evidence base for scientific validation.

3. How to Use

1. Get the code and weights: Access the SciReasoner GitHub repository (link to be updated after official release).

2. Set up the runtime environment: It is recommended to use the Linux operating system (Ubuntu 22.04+), Python 3.10+, and CUDA 12.1 or higher. Install dependencies using pip install -r requirements.txt. Core dependencies include PyTorch 2.x, Transformers, Foldseek, RDKit, and other scientific computing libraries. It is advised to use conda to create an isolated virtual environment to avoid package conflicts.

3. Prepare structural input: Prepare the corresponding structural data based on the task type. For protein tasks, provide a PDB file or precomputed Foldseek 3Di sequence; for small molecule tasks, provide a SMILES or ConfSeq representation; for crystal tasks, provide a CIF file or SLICES encoding. The project includes conversion scripts that can automatically transform standard structural files into structure-aware Tokens.

4. Perform inference: Load the model via the command line or Python API. Command line example: python run_inference.py --model_path ./SciReasoner --input protein.pdb --question "Predict the molecular function of this protein". The model will output the reasoning chain with <think...> tags and the final prediction result. Inference time depends on the complexity of the input structure, with a single inference taking approximately 5–30 seconds (NVIDIA A100).

5. Result validation and visualization: Map the structural Tokens in the inference trajectory back to 3D coordinates and use visualization tools such as PyMOL or ChimeraX for expert review. The project provides the map_tokens_to_structure.py script to help users map key Tokens in the reasoning chain to actual 3D regions, facilitating the validation of the reasoning logic.

Notes: The first inference requires loading the model weights, and it is recommended to use a GPU with at least 24GB of memory (e.g., NVIDIA A10/A100). For large-scale virtual screening tasks, batch processing mode can be enabled to improve throughput. The model is currently optimized for English scientific terminology, and Chinese scientific questions require term mapping.

4. Pros and Cons Analysis

5. Comparative Analysis with Similar Tools

Selection Recommendations: For research scenarios requiring interdisciplinary scientific reasoning (such as protein-molecule interactions in drug discovery or multi-property prediction in material screening), SciReasoner is currently the only option that supports native structure reasoning and unified cross-domain capabilities, making it particularly suitable for academic research with high interpretability requirements. If the task focuses on protein sequence function prediction and interpretability is not a priority, ESM2 remains a mature specialized model with advantages in deployment convenience and reasoning speed. If only small molecule retrosynthesis planning is needed, RSGPT performs comparably on specific benchmarks and has a more lightweight model. For pure material property prediction, graph neural networks like CGCNN are more efficient, but they lack cross-modal reasoning capabilities.

6. Editor's Summary

SciReasoner has achieved a significant breakthrough in the field of scientific AI in terms of technological innovation. Its core innovation—the native structural reasoning paradigm—unifies protein 3D structures, molecular conformations, and crystal lattices into a structure-aware vocabulary that can be reasoned about, enabling AI models for the first time to perform logical deductions based on real spatial evidence, just like human scientists, rather than relying solely on statistical patterns. This design fundamentally changes the "black box" nature of scientific AI, making the reasoning process traceable and verifiable, marking a milestone in building trust in AI's role in scientific discovery. From a technical implementation perspective, the three-stage continuous pre-training strategy and the bootstrapped reasoning framework effectively integrate structural semantics with the chain-of-thought capabilities of language models, achieving state-of-the-art (SOTA) results on 67 out of 86 interdisciplinary benchmarks, demonstrating the effectiveness of its technical approach.

In terms of practical value, SciReasoner directly addresses several long-standing pain points in scientific AI: low accuracy in predicting the functions of proteins with low homology, poor explainability in retrosynthetic planning, and the fragmentation of multi-disciplinary tools. Its performance on tasks such as orphan protein discovery, retrosynthetic planning for drug molecules, and material property prediction has already reached or exceeded the level of specialized models. Moreover, a single model can cover three major scientific domains, significantly reducing the tool-learning costs for researchers. For pharmaceutical companies, material research institutions, and university laboratories, this model provides a one-stop infrastructure for scientific reasoning.

The target user base is clearly defined: it primarily serves researchers in fields such as computational biology, cheminformatics, and materials genomics, as well as science educators who require explainable AI assistance. For AI developers, its open-source architecture and structure-aware vocabulary design also provide a solid foundation for further customization and expansion. In terms of future development potential, as structural biology data (such as AlphaFold-predicted structures) continues to grow, and with the integration of more scientific domains (such as nanomaterials and complex biological systems), SciReasoner's cross-modal reasoning capabilities are expected to expand further. However, it should be noted that the model's current reliance on high-quality structural data and the potential for optimizing inference speed will be key challenges to overcome for its transition into large-scale industrial applications.

7. Application Scenarios

• Orphan Protein Function Discovery: For unknown proteins lacking homologous sequence information, researchers can input the three-dimensional structure predicted by AlphaFold into SciReasoner. The model infers molecular functions and subcellular localization based on folding patterns and domain organization. This capability holds significant value in scenarios such as metagenomic protein annotation and identification of pathogen virulence factors, drastically reducing the candidate screening cycle before experimental validation.

• Reverse Synthetic Design of Drug Molecules: Medicinal chemists input the target molecule's SMILES and three-dimensional conformation into the model. SciReasoner analyzes the scaffold and functional group characteristics, identifies key bond-breaking sites, and recommends multiple synthetic routes from commercially available precursors to the target molecule. Each route includes a feasibility assessment of the reaction and reasoning basis, helping chemists quickly evaluate synthesis difficulty and optimize their strategies.

• Novel Material Screening: Materials scientists input candidate crystal structures (in CIF files) into the model. SciReasoner predicts key parameters such as formation energy, band gap, and mechanical properties, and explains which coordination environments and symmetry features lead to specific properties. In fields such as battery materials, catalysts, and optoelectronic materials, this capability can accelerate the screening of stable and high-performance materials from millions of candidate structures.

• AI-Assisted Science Education: Using the model's interpretable reasoning trajectory, teachers can demonstrate the core scientific principle of "structure determines properties" in specific cases. For example, showing how the model infers catalytic functions by analyzing the three-dimensional arrangement of protein active sites helps students understand abstract spatial structure concepts and develop scientific thinking.

• Cross-modal Scientific Discovery: Researchers can pose interdisciplinary questions, such as "Design a small molecule that can bind to target protein P and has good oral bioavailability." SciReasoner simultaneously invokes protein structure reasoning and molecular generation capabilities, outputting candidate molecules along with inferred binding patterns. This provides a unified reasoning platform for interdisciplinary fields such as systems biology and chemical biology.

8. FAQ

Q: What is the parameter count of the SciReasoner model?
A: The official has not disclosed the exact parameter count, but the model is based on the Qwen3 series and has undergone continuous pre-training. Based on its functional scale and performance, it is estimated that the base model is at the 7B level (e.g., Qwen3-7B), and with the addition of the structure-aware vocabulary and reasoning module, the total parameter count is approximately 7-10B. Please refer to the official technical paper for the exact parameter count.

Q: Is the model fully open source? Is a license required for commercial use?
A: The code has been released in the GitHub repository under an open-source license (please refer to the LICENSE file in the repository for the specific license). The model weights can be downloaded from the project's official website or HuggingFace. Commercial use is currently not restricted, but it is recommended to confirm the license terms before commercial deployment. The training data comes from publicly available scientific databases and does not involve privacy issues.

Q: How fast is the inference speed? What kind of hardware is required?
A: A single inference (e.g., predicting the function of a single protein) takes approximately 5-15 seconds on an NVIDIA A100 (40GB) and 10-30 seconds on an RTX 4090 (24GB). It is recommended to use a GPU with at least 24GB of memory. If the GPU memory is insufficient, inference can still be performed using a CPU, but the speed will significantly decrease. In batch processing mode, throughput can be improved, making it suitable for tasks like virtual screening.

Q: Does the model support scientific questions in Chinese?
A: The model is based on Qwen3 pre-training, which has both Chinese and English capabilities. However, the training data for SciReasoner is primarily composed of English scientific literature and databases, and the structure-aware vocabulary is also based on English terminology. While the model can understand common Chinese scientific terms, it is recommended to ask questions in English for the best results. The accuracy of mapping for Chinese scientific terminology requires further optimization.

Q: Can the model handle proteins without known 3D structures?
A: Yes, but with reduced performance. The core strength of the model lies in native structure reasoning. If only an amino acid sequence is provided, the model will degrade to a sequence pattern matching mode and will not be able to leverage its structure-based reasoning capabilities. For proteins with unknown structures, it is recommended to first use tools like AlphaFold to predict the structure, and then input it into SciReasoner for further reasoning.

Q: How can I fine-tune the model on my own data?
A: The GitHub repository provides training scripts and example configurations. Users need to prepare data in the structure-aware Token format and follow the three-stage training process for fine-tuning. Due to the model's large size, it is recommended to use multi-A100 GPUs or higher-end configurations for fine-tuning. Detailed tutorials are currently being developed, and users can refer to the experimental setup in the technical paper for now.

Q: What are the unique advantages of SciReasoner compared to other scientific AI models?
A: The core advantage lies in native structure reasoning and cross-disciplinary unification. Traditional models (e.g., ESM2, CGCNN) typically handle only a single modality (sequence or graph), and their reasoning processes are not transparent. SciReasoner unifies proteins, molecules, and crystal structures into a single Token-based reasoning framework and outputs verifiable reasoning chains. This provides significant advantages in low-homology scenarios and tasks requiring interpretability.

9. Project Links

• Project Website: https://scireasoner.github.io/
• GitHub Repository: https://github.com/SpectrAI-Initiative/SciReasoner
• arXiv Technical Paper: https://arxiv.org/pdf/2607.07708