Published 2026-07-08 · Shanghai AI Laboratory · Chen Tang, Yizhou Wang, Jianyu Wu et al.

Accurate, Interdisciplinary and Transparent Structure-property Understanding with Deep Native Structural Reasoning

Abstract (verbatim)

Structure-property relationships are foundational to biology, chemistry and materials science, where function, reactivity and physical response emerge from spatial, chemical and periodic organization. Mechanistically explaining these relationships requires interpreting structural evidence through scientific principles and physical constraints, from stereochemistry and bonding to symmetry, energetics and periodic order. However, applying artificial intelligence to this process presents a joint challenge of representation and reasoning: models must preserve domain-native structural information while showing how specific evidence supports predictions under these constraints. Here we introduce SciReasoner, a multimodal scientific foundation model for native structural reasoning across proteins, small molecules and inorganic crystals. SciReasoner discretizes coordinates, topologies and periodic connectivities into a unified structure-aware vocabulary, treating structural tokens as addressable evidence units during reasoning. In homology-controlled Gene Ontology prediction, SciReasoner improves Cellular Component annotation for low-homology and orphan-like proteins, increasing F_{max} from 0.42 to 0.55. In chemistry, it raises single-step retrosynthesis accuracy from 0.63 to 0.72 while generating fragment-level disconnection and precursor-verification traces. In materials science, its representations separate elemental and compound phases and resolve high- and low-band-gap regimes. Across 86 benchmarks, SciReasoner achieves state-of-the-art performance on 67 tasks. Double-blind expert evaluation rates its reasoning traces as preferred or at least comparable to those of a frontier large language model in 98% of cases. By making structure an inspectable substrate for reasoning under scientific constraints, SciReasoner connects accurate prediction with interpretable scientific inference.

Background

Background Analysis

Technical Context: In biology, chemistry, and materials science, understanding "structure-property relationships" is fundamental—e.g., how a protein’s 3D structure determines its function, how a molecule’s chemical bonds influence its reactivity, or how a material’s crystal arrangement dictates its conductivity. Scientists need to extract patterns from vast structural data (e.g., protein sequences, molecular conformations, or crystal lattices), but manual analysis is slow and error-prone. AI offers a solution, but it must address two key challenges: accurately predicting properties (e.g., protein function or material band gap) while explaining why the prediction was made (i.e., ensuring reasoning aligns with scientific principles).

Previous Limitations: Existing AI models face two critical bottlenecks. First, insufficient representation: Traditional approaches (e.g., deep learning or graph neural networks) often reduce structural information to numerical features, losing domain-specific scientific constraints (e.g., stereochemical rules in chemistry or periodic symmetry in crystals). Second, opaque reasoning: Even when models predict outcomes (e.g., protein function classification), their decision-making is often a "black box," leaving scientists unable to verify or trust the logic. For example, low-homology protein function prediction suffers from accuracy issues due to a lack of similar structural references, while materials science struggles to distinguish between elemental and compound phases in band gap predictions.

Proposed Solution: The paper introduces SciReasoner, which addresses these issues through three innovations. First, unified structural representation: It converts coordinates, topologies, and periodic connectivities of proteins, small molecules, and crystals into a "structure-aware vocabulary," allowing the model to parse structural evidence (e.g., chemical bonds in molecules or atomic arrangements in crystals) as traceable "evidence units." Second, constraint-aware reasoning: Scientific principles (e.g., chemical bonding rules or energy optimization) are explicitly integrated into reasoning, ensuring predictions adhere to physical constraints. For instance, in drug synthesis planning, the model not only outputs the target molecule but also generates step-by-step chemical disconnection justifications. Third, enhanced interpretability: Detailed reasoning traces (e.g., labeling which structural features support a prediction) allow scientists to validate the model’s logic against domain knowledge.

Key Differentiation: Unlike prior work, SciReasoner’s core innovation is treating structure as an interpretable foundation for reasoning. Previous models either prioritized prediction accuracy (ignoring physical constraints) or relied on manually designed rules (lacking generalization). SciReasoner achieves end-to-end modeling of "structure-property relationships" while preserving domain-native information (e.g., molecular stereochemistry or crystal periodicity) and providing transparent reasoning. This approach outperforms existing methods across tasks (e.g., protein function annotation, retrosynthesis planning, and materials phase analysis) and earns high expert approval for its logical consistency.