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

The Core Claim

Deep learning models can predict molecular and material properties with both high accuracy and human-readable scientific reasoning—if you treat 3D structures as discrete tokens that preserve their spatial and chemical meaning. A new model called SciReasoner demonstrates this across proteins, drug molecules, and inorganic crystals, achieving state-of-the-art results on 67 of 86 benchmarks while generating explanations that experts prefer over outputs from frontier language models in 98% of cases.

How It Works

The key insight is representation. Most AI systems either compress structures into dense vectors (losing interpretability) or feed raw coordinates to attention mechanisms (obscuring the chemical logic). SciReasoner takes a third path: it discretizes the 3D world into a unified vocabulary of structural tokens.

For proteins, this means encoding coordinates, bonding patterns, and stereochemistry into discrete units. For small molecules, similar tokenization captures bonding topology and reactivity hotspots. For inorganic crystals, the method encodes periodic connectivity—how atoms repeat in lattice space. These tokens become addressable evidence units. When the model makes a prediction, it points to specific tokens as support, making reasoning inspectable.

The model then reasons over these tokens while respecting scientific constraints: bonding rules, stereochemical principles, symmetry, and periodicity. This is not arbitrary attention over words—it's structured reasoning grounded in domain knowledge.

Why It Matters

Proteins: Gene Ontology prediction is hard for rare or poorly-characterized proteins (low-homology and "orphan" sequences). SciReasoner improved Cellular Component annotation F-score from 0.42 to 0.55 on these difficult cases—a 31% relative gain. This matters because function annotation drives drug discovery and systems biology.

Chemistry: Retrosynthesis—predicting how to build a molecule step-by-step—improved from 63% to 72% accuracy. More importantly, the model generates fragment-level disconnection traces and verifies that precursors exist, meaning chemists can trust and act on predictions.

Materials: The learned representations separate elemental from compound phases and correctly resolve band-gap regimes (which determine whether a material conducts electricity or light). This is the kind of physical reasoning typically done by hand.

Interpretability at scale: Double-blind expert evaluation rated the model's reasoning traces as preferred or equivalent to a frontier LLM in 98% of cases. This is rare. Most high-performing models are black boxes; SciReasoner couples accuracy with explainability.

The broader value is methodological: by treating domain structure as a first-class object in the token vocabulary, the authors show that you can build foundation models that are simultaneously predictive and scientifically transparent. This bridges the long-standing gap between deep learning's empirical power and science's need for mechanistic understanding—essential for adoption in regulated domains like drug development and materials engineering.