English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

SciReasoner: A Unified Structural Language for Chemistry, Proteins, and Materials

Forum topic · 小凯 · 2026-07-10

Summary

A Chinese tech forum post discusses SciReasoner, a foundation model described in the paper "Accurate, Interdisciplinary and Transparent Structure-property Understanding with Deep Native Structural Reasoning" (arXiv:2607.07708) by researchers from SenseTime, Shanghai AI Lab, and Oxford. SciReasoner addresses scientific fragmentation by converting proteins, small molecules, and inorganic crystals into a common vocabulary of structure tokens—encoding spatial positions, topological connections, and chemical elements. Unlike graph neural networks that produce opaque embeddings, the model makes structure tokens addressable evidence units, citing specific structural features step by step during reasoning. Reported results: state-of-the-art on 67 of 86 benchmarks across protein function, molecular property, crystal property, and reaction prediction tasks. F_max for low-homology protein function prediction rises from 0.42 to 0.55, and single-step retrosynthesis accuracy improves from 0.63 to 0.72 with generated explanation paths. In a double-blind expert evaluation, experts judged SciReasoner's reasoning traces comparable to or better than a frontier LLM in 98% of cases, suggesting the model achieves auditable, structure-grounded scientific reasoning rather than text-only inference.

> 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 of Science

Biology studies proteins, chemistry studies molecules, materials science studies crystals. Three fields, three languages, three sets of intuitions.

But their underlying structure is the same: atoms arranged in space according to specific rules, producing specific functions. Proteins fold into 3D shapes to catalyze reactions; chemical bonds in a molecule determine its reactivity; the periodic arrangement of a crystal determines its conductivity.

The problem: these structural descriptions appear in different formats in the scientific literature—protein databases use PDB files, chemical molecules use SMILES strings, crystals use CIF files. An AI trained on proteins struggles to "understand" a chemical molecule because it learned a different representation.

SciReasoner's goal is to tear down this Tower of Babel—teaching AI the "lingua franca" of science.

---

Structure as Language

SciReasoner's core insight is simple: convert all scientific structures into one unified "vocabulary."

Three structures, one grammar

1. Proteins: amino acid sequence → 3D coordinates → topological connections (which atoms are bonded) 2. Small molecules: atom types and coordinates → bond topology → functional group patterns 3. Inorganic crystals: unit cell parameters → periodic connectivity → elemental arrangement

SciReasoner discretizes all of these into structure tokens—just as GPT turns text into tokens, SciReasoner turns structures into processable units. But these tokens are not "words"; they encode spatial positions, topological connections, and chemical elements.

Why is this hard?

The traditional approach: feed the structure to a graph neural network (GNN), get a vector embedding, then make predictions. That works, but at a cost: the reasoning is uninterpretable. You know a protein was predicted to be an "enzyme," but not why—which structural features support that conclusion?

SciReasoner's solution is to make structure tokens addressable evidence units. During inference, the model doesn't just output "enzyme" or "not enzyme"—it must cite specific structure tokens step by step to support its conclusion. Like a scientist who not only states a conclusion in a paper but shows the X-ray diffraction data and the analysis behind it.

---

Cross-Domain Validation: From Proteins to Crystals

Proteins: finding a home for "orphans"

In Gene Ontology prediction, one especially hard subtask is predicting functions of low-homology proteins—proteins whose sequences differ greatly from known ones, leaving traditional methods without "relatives" to infer from.

SciReasoner improves F_max from 0.42 to 0.55—a large boost in functional prediction using 3D structure alone, with almost no sequence clues. It learned to infer "function" from "shape," much as biologists do.

Chemistry: transparent retrosynthesis pathways

Retrosynthesis is a core problem in organic chemistry: given a target molecule, how do you break it down step by step into purchasable precursors?

SciReasoner raises single-step retrosynthesis accuracy from 0.63 to 0.72, and crucially it generates explanation paths. The model shows: "I cut this bond because the electron density here supports this cleavage; I validated this precursor because its structure satisfies the following conditions..."

This is not black-box prediction—it is auditable chemical reasoning.

Materials science: separating "good" and "bad" band gaps

In crystal characterization, SciReasoner's embedding space clearly separates different elements and compound phases, and distinguishes high-band-gap from low-band-gap materials. The model learned to "see" the microscopic structures that determine electronic properties—not just memorize which materials appeared in the training set.

---

86 Benchmarks, 67 SOTA Results

The paper evaluates on 86 benchmarks covering protein function prediction, molecular property prediction, crystal property prediction, and chemical reaction prediction. SciReasoner achieves state-of-the-art (SOTA) on 67 of them.

But that is not the most impressive part. The paper also ran a double-blind expert evaluation: human experts (not knowing which outputs came from SciReasoner vs. a GPT-4-class frontier LLM) judged the quality of reasoning traces. Result: in 98% of cases, experts rated SciReasoner's reasoning as comparable to or better than the frontier LLM.

What does that mean? The AI wins not only on prediction accuracy but also on the credibility of its reasoning.

---

Why Is This Better Than "Big LLM + Science Prompts"?

One might ask: why not just use GPT-5 for these tasks—give it scientific literature and let it reason?

The answer: scientific reasoning requires structural sense. GPT reads text; it "knows" a benzene ring is a hexagon, but it cannot "see" the hexagon. It cannot intuitively grasp bond angles, ring strain, or the distribution of π-electron clouds. Such knowledge is "descriptive" for it, not "experiential."

By encoding structures directly as tokens, SciReasoner lets the model reason in structure space. It doesn't "read" books about proteins—it directly "sees" their 3D shapes. This is not a metaphor—the model genuinely operates on 3D coordinates and topological connections.

---

One-Sentence Takeaway

SciReasoner doesn't make AI "read more science books"—it teaches AI the universal grammar of science: structure itself. When chemical bonds and protein folding are translated into the same language, interdisciplinary breakthroughs no longer depend on human domain translation; they can happen directly at the structural level.

---

*Auto-collected on 2026-07-10*

Tags

#scireasoner#arxiv#scientific-ai#foundation-models#multimodal#protein-folding#retrosynthesis#materials-science

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/178346290