SciReasoner

Scientific machine learning has largely grown as a patchwork of modality-specific tools: one model reads protein sequences, another predicts small-molecule reactions, a third handles crystal structures, and each speaks its own representation. SciReasoner, developed by the open-sciencelab team at Shanghai AI Laboratory, is a scientific reasoning foundation model that dissolves these boundaries by aligning natural language with heterogeneous scientific representations inside a single autoregressive model.

Its defining idea is native structural reasoning. Rather than flattening a molecule or protein into an unstructured string, SciReasoner discretizes atomic coordinates, molecular topologies, and periodic connectivities into a unified structure-aware vocabulary. Structural tokens become addressable evidence units that the model can attend to and cite while it reasons, letting it move between a protein's sequence, its three-dimensional form, a chemical reaction, and the surrounding natural-language question.

The model family spans domains directly relevant to the life sciences — protein sequence and structure, DNA and RNA, and small-molecule chemistry — alongside inorganic materials. This lets a single backbone cover tasks that have historically each required a bespoke specialist model. A companion structure-property reasoning study extends the same system with deeper native structural reasoning across proteins, molecules, and crystals.

#Key Features

• Unified structure-aware vocabulary: Coordinates, topologies, and periodic connectivities are tokenized into one shared vocabulary, so protein, molecule, and crystal structure share a common representational substrate with text.
• Cross-modal reasoning: A single checkpoint maps between molecular sequences, structures, and natural language, supporting both understanding and generation without per-task architectures.
• Homology-controlled protein annotation: On Gene Ontology prediction stratified by sequence homology, it improves Cellular Component F-max for low-homology and orphan-like proteins from 0.42 to 0.55.
• Interpretable chemistry traces: For single-step retrosynthesis it raises accuracy from 0.63 to 0.72 while emitting fragment-level disconnection and precursor-verification reasoning traces.
• Open release: Evaluation code is Apache-2.0 licensed and the 8B checkpoint is downloadable, alongside the instruction-tuning and evaluation datasets.

#Technical Details

SciReasoner is an autoregressive transformer built on the Qwen architecture and released in 1.7B- and 8B-parameter variants. It is pretrained on a 206-billion-token corpus that mixes scientific text, pure biological and chemical sequences, and paired sequence-text data, then aligned through supervised fine-tuning on roughly 40 million instructions spanning many scientific capability families. Evaluation covers 86 benchmarks, on which the model reaches state-of-the-art performance on 67 tasks, including forward reaction prediction, reagent selection, retrosynthesis, text-guided molecule design, and protein function annotation. Scaling from the 1.7B to the 8B checkpoint strengthens both conditional and unconditional generation across molecules, proteins, RNAs, and materials.

#Applications

SciReasoner serves computational biologists and chemists who would otherwise stitch together several single-purpose models. Protein scientists can use it for function annotation and structure-conditioned reasoning, medicinal chemists for retrosynthetic planning and molecule design with human-readable justification, and materials researchers for property understanding of inorganic crystals. Because it accepts and produces natural language, it fits into interactive research assistants and automated pipelines where a scientist poses a question and expects both an answer and a traceable rationale.

#Impact

By showing that structure, sequence, and text can share one reasoning substrate, SciReasoner advances the case for genuinely multimodal scientific foundation models rather than isolated per-domain systems. The open release of the 8B checkpoint, evaluation harness, and training datasets under permissive terms lowers the barrier to reproducing and extending cross-disciplinary results. Its principal trade-off is scope: the same generality that spans proteins, molecules, and crystals means it is not a specialist for any single benchmark, and the crystal and materials capabilities fall outside biological use. As a preprint-stage system, its broader claims await independent replication.