What if artificial intelligence could finally explain the why behind molecular behavior, rather than just predicting the what? For decades, AI in science has been a black box, delivering accurate answers without showing its work. This lack of transparency has hindered trust in critical fields like drug discovery and materials science. Now, researchers have introduced a new model that treats structural data not as abstract numbers, but as inspectable evidence that can be traced and understood.

The new system, called SciReasoner, is a multimodal foundation model designed to reason about the physical structures of proteins, small molecules, and inorganic crystals. Instead of relying on vague correlations, it breaks down complex 3D structures into a unified vocabulary of tokens. These tokens act as addressable units of evidence, allowing the model to apply scientific principles like bonding, symmetry, and energetics directly to its predictions. By treating structure as the substrate for reasoning, the model ensures that every prediction is grounded in verifiable physical constraints rather than statistical guesswork.

The results are striking, particularly in areas where traditional methods struggle. In protein analysis, SciReasoner significantly improved the annotation of low-homology and orphan-like proteins, boosting performance metrics from 0.42 to 0.55. In chemistry, it increased the accuracy of single-step retrosynthesis planning from 0.63 to 0.72, while generating detailed traces that show exactly how it arrived at its conclusions. Across 86 different benchmarks, the model achieved state-of-the-art performance on 67 tasks. Perhaps most importantly, double-blind evaluations by experts found its reasoning traces to be preferred or comparable to those of a frontier large language model in 98% of cases.

This research marks a shift from pure prediction to interpretable inference. By making the reasoning process transparent, SciReasoner connects high accuracy with scientific understanding. For researchers, this means AI tools that do not just output answers but provide the logical, evidence-based pathways to get there. As science demands greater rigor and reproducibility, models that can explain their structural logic may become the new standard for discovery.