SciReasoner: A Unified Structural Language for Proteins, Molecules, and Crystals

Forum topic · 小凯 · 2026-07-10

Summary

SciReasoner, developed by researchers from SenseTime, Shanghai AI Lab, and Oxford, is a foundation model that translates diverse scientific structures—protein 3D folds, molecular SMILES/bond topologies, and crystal CIF descriptions—into a unified vocabulary of structural tokens. Instead of treating structures as opaque embeddings from graph neural networks, the model performs reasoning by explicitly citing structural tokens as evidence, making predictions auditable. Evaluated across 86 benchmarks spanning protein function, molecular property, reaction, and crystal property prediction, SciReasoner achieves state-of-the-art results on 67 of them. Highlights include improved F_max (0.42 to 0.55) on low-homology protein function prediction, single-step retrosynthesis accuracy rising from 0.63 to 0.72 with generated explanatory pathways, and embedding spaces that separate high- and low-bandgap materials. In a double-blind expert evaluation, 98% of judgments rated SciReasoner's reasoning traces as comparable to or better than a frontier LLM, underscoring both accuracy and interpretability gains.

Overview

Original paper: Accurate, Interdisciplinary and Transparent Structure-property Understanding with Deep Native Structural Reasoning arXiv: 2607.07708 Authors: Chen Tang, Yizhou Wang, Jianyu Wu, et al. (SenseTime, Shanghai AI Lab, Oxford, etc.) Categories: cs.CL, cs.AI, cs.CE, cs.LG

The Tower of Babel in Science

Biology studies proteins, chemistry studies molecules, and materials science studies crystals—three fields, three languages, three sets of intuitions. Yet their underlying structure is the same: atoms arranged in space according to specific rules, producing specific functions.

The problem is that these structural facts are described in incompatible formats: PDB files for proteins, SMILES strings for molecules, CIF files for crystals. An AI trained on proteins struggles to understand a chemical molecule because it learned a different representation.

SciReasoner aims to tear down this Babel—teaching AI the lingua franca of science.

Structure as Language

The core insight: convert all scientific structures into a unified vocabulary.

Three Structures, One Grammar

1. Proteins: amino acid sequence → 3D coordinates → topological connections 2. Small molecules: atom types and coordinates → bond topology → functional group patterns 3. Inorganic crystals: unit cell parameters → periodic connections → element arrangement

SciReasoner discretizes all of these into structure tokens—just as GPT turns text into tokens, SciReasoner turns structures into processable units encoding spatial positions, topological connections, and chemical elements.

Why This Is Hard

Traditional approaches feed structures into a graph neural network (GNN) to get a vector representation, then make predictions. Effective—but the reasoning is uninterpretable: you know a protein is predicted to be an enzyme, but not which structural features support that conclusion.

SciReasoner's solution: make structure tokens addressable evidence units. During inference, the model must cite specific structural tokens step by step to support its conclusions—like a scientist showing X-ray diffraction data and analysis, not just a verdict.

Cross-Domain Validation

Proteins: Finding a Home for "Orphans"

For Gene Ontology prediction, the hardest subtask is low-homology proteins, whose sequences differ greatly from known proteins. SciReasoner raises F_max from 0.42 to 0.55, inferring function from 3D structure alone—much as biologists do.

Chemistry: Transparent Retrosynthesis

SciReasoner lifts single-step retrosynthesis accuracy from 0.63 to 0.72, and crucially generates explanation pathways: which bond was cut and why, and why each precursor is valid. This is auditable chemical reasoning, not black-box prediction.

Materials Science: Separating Good and Bad Band Gaps

SciReasoner's embedding space cleanly separates elements, compound phases, and high- vs. low-bandgap materials—showing the model "sees" the microstructures that determine electronic properties rather than memorizing training examples.

Results at a Glance

Evaluated on 86 benchmarks across protein function, molecular property, crystal property, and chemical reaction prediction
State-of-the-art on 67 of them
Double-blind expert evaluation: in 98% of cases, experts (unaware of model identity) rated SciReasoner's reasoning traces as comparable to or better than a frontier LLM

Why Not Just Prompt a Frontier LLM?

Scientific reasoning requires a sense of structure. GPT reads text—it "knows" benzene is a hexagon but cannot "see" it; it has no intuitive grasp of bond angles, ring strain, or π-electron distributions. For an LLM, this knowledge is descriptive, not experiential.

SciReasoner encodes structures directly as tokens, so the model reasons in structure space—it doesn't read books about proteins, it directly "looks at" 3D shapes, computing over coordinates and topological connections.

One-Sentence Takeaway

SciReasoner doesn't make AI read more science books—it teaches AI science's universal grammar: structure itself. When chemical bonds and protein folding speak the same language, cross-disciplinary breakthroughs can happen directly at the structural level, without human domain translation.

---

*Auto-collected on 2026-07-10.*