Chen Tang, Yizhou Wang, Jianyu Wu, Lintao Wang, Shixiang Tang, Pengze Li, Encheng Su, Jun Yao, Jiabei Xiao, Yuqi Shi, Jielan Li, Hongxia Hao, Zhangyang Gao, Fang Wu, Ben Fei, Xiangyu Yue, Pan Tan, Bozitao Zhong, Jinouwen Zhang, Aoran Wang, Yan Lu, Jiaheng Liu, Xinzhu Ma, Liang Hong, Mingyue Zheng, Phil Torr, Bowen Zhou, Wanli Ouyang, Lei Bai
5 min
SciReasoner addresses the gap between high-performance predictive models and interpretable scientific reasoning. Traditional AI models for science often treat structures as auxiliary text or black-box inputs, which limits their ability to explain why a specific structure leads to a particular property. SciReasoner introduces a native structural reasoning paradigm by discretizing 3D coordinates, topologies, and periodic connectivities into a unified, structure-aware vocabulary. These tokens act as addressable evidence units within an autoregressive model, allowing the system to interleave natural language explanations with explicit references to structural motifs, such as binding sites or lattice symmetries.
SciReasoner demonstrates significant improvements in scientific inference across biology, chemistry, and materials science. In protein function prediction, it shows robust performance even for low-homology or orphan proteins, where traditional sequence-based methods often fail. In retrosynthesis, the model not only improves prediction accuracy but also provides verifiable traces of fragment-level disconnections. For materials science, it effectively captures the relationship between crystal symmetry, local coordination, and electronic properties like band gaps. Across 86 benchmarks, the model achieves state-of-the-art results on 67 tasks, with expert evaluations confirming that its reasoning traces are highly useful for scientific inspection.
This work represents a shift toward scientific AI systems that function as collaborative partners rather than just predictive engines. By grounding reasoning in domain-native structural evidence, SciReasoner bridges the divide between accurate property prediction and mechanistic explanation. This capability is essential for accelerating discovery in fields where understanding the underlying physical and chemical constraints is as important as the final result, such as drug design and the development of novel functional materials.
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.
Alex: Does that reasoning hold up when the model encounters something genuinely new—a molecule it's never seen before?
Sam: The study suggests it does, and that's one of the more notable findings. In protein function prediction, the model's largest performance gains came on proteins that were very different from anything in its training data. That pattern suggests it isn't just matching against familiar examples—it's applying structural logic, reasoning from physical principles rather than from memory.
Alex: So it's less like a student who memorized the answers, and more like one who actually understood the method.
Sam: That's the distinction the researchers are drawing. And it extends beyond biology. The model was tested across a wide range of scientific domains—proteins, small molecules, and inorganic crystals—and the same approach held up throughout.
Alex: If the whole system depends on those encoders translating physical structure into symbols, what happens when an encoder misses something? If you're looking at an unusual new material that doesn't fit the existing rules, does the reasoning fall apart?
Sam: That's the real limitation, and the paper doesn't sidestep it. The model can only reason about what its encoders can represent. If a particular atomic arrangement is too unusual to be captured in the structural vocabulary, the model is effectively blind to it—it can't reason about what it hasn't been given. So it's highly effective within the chemical space it was trained to represent, but it isn't a universal observer.
Alex: It's a tool for structured reasoning within known frameworks, not a system that discovers entirely new physical laws from scratch.
Sam: That's a fair characterisation. And within those boundaries, the practical value is real. Instead of running hundreds of experiments to screen candidates, a scientist gets a structured, verifiable trace of why a molecule is predicted to behave a certain way. That shifts the work from tedious trial-and-error toward something more deliberate.
Alex: So the contribution isn't that it replaces scientific intuition—it's that it makes the AI's suggestions legible enough to actually be useful. You can interrogate the reasoning, not just accept the output.
Sam: That's well put. By grounding its logic in the physical structure of matter, SciReasoner offers a clearer bridge between raw data and scientific understanding. It has real limitations, but as a step toward more inspectable, accountable AI in the lab, it's a meaningful one.
Alex: Thanks for walking me through it, Sam. And thanks to everyone listening—this has been ResearchPod.