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

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

SciReasoner integrates 3D structural data as addressable tokens into autoregressive reasoning for biology, chemistry, and materials science.

How can a unified multimodal foundation model be trained to perform explicit scientific reasoning across protein, molecular, and materials domains by natively processing structural data?

Scientific AI models often treat proteins, molecules, and crystals as text strings or black-box inputs, forcing them to rely on superficial linguistic associations rather than physical evidence. This makes it impossible for experts to verify the mechanistic basis of a model's prediction. SciReasoner solves this by discretizing 3D coordinates, topologies, and periodic connectivities into a unified, structure-aware vocabulary. These structural tokens act as addressable evidence units that the model can cite and interleave with natural language during its reasoning process. Across 86 benchmarks, SciReasoner achieves state-of-the-art performance on 67 tasks, including significant gains in low-homology protein annotation and retrosynthesis accuracy.

Paper Primer

The core move is the creation of a structure-aware vocabulary that preserves the physical integrity of scientific objects. SciReasoner is a unified reasoning engine: it maps diverse scientific inputs into a shared latent space where structural tokens function like citations in a chain-of-thought, allowing the model to ground its claims in specific residues, fragments, or lattice sites.

SciReasoner improves functional annotation for low-homology proteins where evolutionary transfer fails.

In the Cellular Component aspect of Gene Ontology prediction, the model improved Fmax from 0.42 to 0.55 for proteins with ≤30% sequence identity. +0.13 absolute Fmax improvement over the ESM2 baseline.

The model generates chemically valid retrosynthetic routes by decomposing targets into reaction-relevant fragments.

On the USPTO-50K benchmark, SciReasoner achieved 0.72 single-step retrosynthesis accuracy. +0.09 improvement over the previous best template-free method, RSGPT.

Why is this approach superior to simply feeding structural descriptors into a standard large language model?

Standard models compress structural organization into text strings, which destroys scientific semantics. SciReasoner’s native structural tokens preserve 3D geometry and periodic connectivity, allowing the model to perform reasoning grounded in physical evidence rather than linguistic patterns.

Does the model rely on memorizing known structures to achieve these results?

The authors explicitly tested for this by stratifying performance by sequence identity and performing structure-ablation experiments. The model maintains high performance in low-homology regimes and shifts its reasoning focus to binding pockets and coordination environments when structural tokens are provided, indicating genuine structure-grounded inference.

SciReasoner demonstrates that scientific foundation models can move beyond pattern-matching by treating structural data as a primary, inspectable substrate for reasoning, enabling more reliable and interpretable scientific discovery.

Scientific Reasoning and Structure

We expose the need for native structural reasoning to bridge the gap between structure and function.

Structure–property relationships are foundational to biology, chemistry, and materials science, where function, reactivity, and physical response emerge from spatial, chemical, and periodic organization. Yet AI models face a joint challenge: they must preserve domain‑native structural information while providing explicit, evidence‑linked reasoning, otherwise explanations rely on linguistic associations rather than physical evidence.

Structure‑property understanding is the ability to infer how the spatial and chemical organization of a system gives rise to its function, reactivity, or physical behavior.

Explicit structural reasoning is essential for scientific discovery because it grounds predictions in the physical evidence that truly determines function.

The SciReasoner Framework

Broad evaluation shows SciReasoner delivers specialist‑level accuracy across proteins, molecules, and crystals.

We evaluate SciReasoner on a comprehensive suite of 93 scientific benchmarks covering proteins, nucleic acids, small molecules, and inorganic crystals. The experiments test both closed‑set prediction and open‑ended language generation, probing raw accuracy and the quality of structure‑grounded reasoning traces.

SciReasoner is a single decoder‑only transformer that ingests sequence, 3‑D structure, and free‑text inputs, converts each into a shared “structure‑aware” token stream, and generates a unified reasoning trajectory.

How does SciReasoner differ from a conventional multimodal transformer that simply concatenates raw sequences and graphs?

Conventional models treat structure as an auxiliary feature and tokenize it with generic sub‑word vocabularies, which break chemical bonds and crystal symmetries. SciReasoner replaces those generic tokens with a purpose‑built structure‑aware vocabulary, so the model’s attention operates on physically meaningful units rather than arbitrary byte strings.

**Figure 1.** Overview of the SciReasoner framework. (A) SciReasoner architecture processing multi-modal scientific inputs. (B) Structural tokenization methods for molecules, proteins, and crystals. (C) Comparison of tokenizer compression ratios between Qwen and SciReasoner. (D) Training pipeline consisting of continued pretraining and post-training stages. (E) Performance evaluation across various scientific domains and tasks using radar and bar charts.

SciReasoner matches or exceeds specialist baselines on 67 of the 93 evaluated benchmarks.

Comprehensive testing across protein function, DNA/RNA annotation, small‑molecule retrosynthesis, and crystal property prediction.

Protein Function Prediction

SciReasoner’s GO prediction outperforms baselines and grounds its decisions in protein structure.

SciReasoner attains an overall Fmax of 0.59 on the Gene Ontology benchmark, surpassing all baselines.

BLAST 0.55, Foldseek 0.54, ESM2 0.53, SaProt 0.52, DeepSeek‑V4‑Pro 0.35, GPT‑5.5 0.31.

**Figure 2.** SciReasoner predicts protein Gene Ontology terms through structure-grounded reasoning. (A) Performance on the CAFA-3 Gene Ontology benchmark for Molecular Function, Biological Process and Cellular Component annotations, stratified by maximum BLAST sequence identity to the training set. SciReasoner shows the largest gains in low-homology regimes, particularly for Cellular Component prediction. (B) Attention analysis for DNA-binding Gene Ontology prediction. Residues receiving high attention are enriched at contact-defined DNA-binding sites and localize to protein–DNA interfaces in representative complexes. (C) Reinforcement-learning trajectories for the three ontology aspects. Rewards increase after an initial exploration phase and converge to stable plateaus, indicating consistent optimization across molecular-function, biological-process and cellular-component annotations. (D) LLM-as-judge evaluation of reasoning traces on held-out Biological Process proteins. Reinforcement learning improves input fidelity, biological plausibility, reasoning coherence and functional-hypothesis quality relative to the coldstart checkpoint and DeepSeek-V4-Pro. (E) Reasoning-quality scores stratified by BLAST similarity for Biological Process prediction. Mean GPT-5.5 judge scores are reported on five rubric axes across seven similarity bins relative to the training set (n = 20 proteins per bin; error bars, s.e.m.). SciReasoner-RL achieves the highest scores across axes and remains stable across similarity bins, suggesting that its reasoning generalizes beyond memorized homologues. (F) Structure-token visualization of the reasoning trace for E. coli $\beta$-ketoacyl-ACP synthase I (FabB; 2bz4_A, UniProt P0A953). Load-bearing claims and cited 3Di-token fragments are highlighted in the trace and mapped onto PyMOL-rendered structures. The VVVV fragment maps to canonical $\alpha$-helices, whereas DDLPPDDP maps to an extended-strand/turn region, supporting the inferred $\alpha$/$\beta$ fold and the resulting fatty-acid biosynthesis annotation.

Retrosynthetic Route Planning

Retrosynthesis accuracy and interpretability are quantified on USPTO‑50K.

SciReasoner attains 0.72 Exact Match on USPTO‑50K, surpassing the prior best template‑free model RSGPT (0.63) by +0.09 points.

Exact Match is measured after sampling 16 stochastic completions per query ($T=0.6$, $top\text{-}p=0.95$) and ranking by sample frequency; a prediction is correct only when its canonical SMILES exactly matches the ground‑truth reactant set.

SciReasoner’s chain‑of‑thought weaves natural‑language reasoning steps with structural tokens (SMILES fragments), enabling the model to emit intermediate chemical sub‑structures that can be inspected atom‑by‑atom.

The reasoning trace follows four explicit stages—Analysis, Disconnection, Verification, and Feasibility—so that each strategic bond cut (e.g., the ester C–O bond) is justified with a fragment‑level claim before the final reactants are assembled.

On five additional USPTO‑50K products SciReasoner places the gold reactant pair within its Top‑3 predictions for all 5 cases, whereas RSGPT and Opus‑4.7 succeed on only 2 of 5.

**Figure 3.** SciReasoner formulates chemically viable retrosynthetic routes (A–C) and captures plausible 3D molecular structure for similarity representation (D, E). (A) Retrosynthesis USPTO-50K performance. Across template-based, semi-template-based and template-free methods, SciReasoner reaches 0.72, exceeding the prior best RSGPT by +0.09 points, while Opus-4.7 five-shot scores only 0.48. (B) Chain-of-thought trace for a representative product. SciReasoner proceeds in four interpretable stages—Analysis, Disconnection, Verification and Feasibility, and recovers the two ground-truth reactants by severing the strategic ester C–O bond. Every intermediate claim is itself a SMILES fragment, making the trace auditable atom-by-atom. (C) Cross-model Top-3 reactant predictions on Retrosynthesis USPTO-50K. Five products with their ground-truth reactants and the Top-3 predictions of SciReasoner, RSGPT and Opus-4.7. Molecules in pink match the gold canonical SMILES. SciReasoner recovers a correct candidate for 5/5 products versus 2/5 for RSGPT and 2/5 for Opus-4.7, generalising across mechanistically distinct reaction families, rather than fitting a single template. (D) Per-target distributions on 3D molecule structure (DUD-E). AUC and 5.0% EF box plots over 102 DUD-E targets for nine published baselines, an Opus-4.7 reference, and SciReasoner. SciReasoner matches the previously reported best AUC and establishes a new best 5.0% EF. (E) 3D ligand-target geometry for three DUD-E targets. UMAP of the per-molecule SciReasoner embedding, coloured points are the 400 closest actives for each target, labelled by DUD-E target identifiers—ADA17 (blue), CAH2 (orange), and PGH2 (green); open stars mark the reference queries. Side panels show each query and three representative actives docked into the receptor, then redrawn as 2D structures: chemically distinct scaffolds that nevertheless converge to the same 3D pharmacophore.

Molecular Similarity and Screening

SciReasoner boosts early enrichment while matching AUC, revealing geometry‑aware similarity.

Virtual screening seeks compounds that share a binding‑pocket geometry despite divergent 2D scaffolds. We test whether SciReasoner’s native structural tokens already encode such geometry‑aware similarity.

SciReasoner improves the 5 % enrichment factor from 7.12 to 7.70 while matching the prior best AUC of 0.76, without docking or task‑specific fine‑tuning.

Evaluated on the DUD‑E benchmark across multiple targets; EF measured at the top 5 % of ranked molecules.

UMAP projections of the embeddings reveal distinct clusters for three unrelated targets (ADAM17, CA II, PGHS‑2). Within each cluster, actives are tightly grouped while decoys remain dispersed, indicating that the representation captures family‑specific pocket geometry.

**Figure 4.** Evaluation of material predictive performance and visualization of the learned structure-property latent space. (A) Performance comparison across 10 material property prediction sub-tasks, both regression (MAE as the metric) and classifications (AUC as the metric) tasks are reported. (B) UMAP of the high-dimensional representations extracted by the model on three materials (C, Si and SiC). The latent space inherently clusters materials by chemical space and spontaneously distinguishes distinct structural polymorphs. (C) Parity plots comparing model-predicted formation energies per atom and bandgap against ground-truth values. The strong correlations (R² = 0.895 and 0.785) demonstrate the model's ability to capture physically meaningful materials-property information. (D) Local UMAP embeddings for the Si and SiC subsets, color-coded by their corresponding band gaps. The continuous transition and distinct grouping of high- and low-bandgap regions indicate that the learned representations effectively capture underlying structure-property relationships. (E) Structure-grounded reasoning trace for a representative stability prediction on TiGaCo2. The model's reasoning explicitly identifies the SLICES symmetry substring, which decodes to the Fm3m space group and cites edge substrings. These substrings are mapped back to the corresponding CIF structure, where they visualize the global cubic symmetry framework and local periodic connectivities.

Structural Ablation Studies

Removing structural inputs hurts performance, confirming they are essential for native reasoning.

Providing structural tokens improves GO molecular‑function prediction $F1$ by 0.60.

Figure 6 F shows $F1$ = 1.00 with structure versus 0.40 without.

**Fig. 5:** Structural information empowers native scientific reasoning. (A) Structure-ablation benchmark. Performance with and without structural inputs across materials, proteins and small molecules. Structural cues consistently improve prediction beyond sequence- or text-only inputs. (B) Modality-separated representations. PCA of Qwen and SciReasoner embeddings for paired protein sequence and structure inputs. SciReasoner more clearly separates the two modalities, indicating stronger structure-aware encoding. (C) Structure-grounded materials reasoning. Chain-of-thought comparison for QMOF pore-limiting-diameter prediction. Without structure, the model relies on compositional priors and overestimates pore size; with structural tokens, it uses symmetry and bonding connectivity to recover a value closer to the ground truth. (D) Structure-grounded GO reasoning. Chain-of-thought comparison for biological-process GO prediction. Sequence-only reasoning predicts DNA-related processes, whereas structure-aware reasoning integrates secondary-structure and binding-site evidence to recover protein-modification annotations. (E) Materials prediction with structure. Absolute-error distributions for bandgap, largest-cavity-diameter and pore-limiting-diameter prediction. Structural input reduces errors across all three properties. (F) Structure-guided GO attention. Residue-level attention maps for GO molecular-function prediction. Structure-aware reasoning focuses on the functional binding site, whereas sequence-only reasoning attends to non-binding regions, consistent with its lower accuracy.

Structure‑ablation experiments show that removing structural inputs degrades performance on all three domains, with the largest drops observed on protein function tasks.

Embedding analysis reveals that SciReasoner separates sequence and structure tokens into distinct clusters, unlike the mixed Qwen space.

Structural Reasoning Consolidation

Intra-domain grounding equips each domain with native structural tokens and RL‑trained experts.

Sequence‑only language models struggle to reason about proteins, molecules, or crystals because they never see the underlying structure as a first‑class object. The result is brittle inference on tasks that demand precise geometric or functional cues. Intra-domain grounding solves this by giving each domain its own structural token and a learned expert that teaches the model how to use it.

The model treats a domain’s structural evidence as a native token and learns, via a domain‑specific expert, how to turn that token into a correct reasoning trace.

How does intra‑domain grounding differ from simply fine‑tuning a language model on domain‑specific data?

Fine‑tuning only adjusts parameters to fit the data distribution; it does not give the model an explicit structural token or a policy that guarantees the token is used correctly. Intra‑domain grounding introduces a dedicated token and an RL‑driven expert that teaches the model *how* to incorporate that token into its reasoning steps, ensuring consistent, structure‑aware behavior.

Step 1: Model receives the token sequence “[text] $s$”. The policy predicts “read $s$” and receives reward +1.

Step 2: Model predicts the label “enzyme”. The expert also outputs “enzyme”, so reward +1.

Total reward = 2, indicating a perfect grounding of the structural token.

This tiny example shows that the expert’s trace forces the model to attend to the structural token before emitting the answer, guaranteeing that the token is not ignored.

**Fig. 6:** Self-bootstrapped native structural reasoning unifies scientific reasoning. (A) Intra-domain structural evidence grounding and cross-domain reasoning consolidation. Sketch reasoning traces coldstart task-specific experts, which are optimized independently by reinforcement learning (RL) to ground structural evidence within each domain. Expert-generated traces are then pooled to coldstart and reinforce a unified SciReasoner across all tasks to consolidate cross-domain reasoning patterns. (B) Reward dynamics. RL rewards rise during training, with the unified SciReasoner matching or exceeding expert variants on structure-intensive and other scientific tasks. (C) Improved accuracy and sampling efficiency. Pass@1 and pass@10 are compared across four training stages for GO term prediction, QMOF prediction, 3D molecule structure (DUD-E) and Retrosynthesis USPTO-50K. Lighter hatched bars indicate pass@1 and solid bars pass@10; $\Delta = \text{pass@10} - \text{pass@1}$. The post-training procedure consistently increases pass@1 and reduces $\Delta$, indicating more reliable first-sample predictions. (D) Improved reasoning quality across tasks. Kernel density estimates show LLM-judged reasoning scores for four model variants on GO term prediction, QMOF prediction, 3D molecular similarity (DUD-E), and Retrosynthesis USPTO-50K. Triangles denote means and red brackets indicate the worst-to-best mean difference. Self-bootstrapped native structural reasoning shifts score distributions towards higher quality, and unified SciReasoner generally outperforms domain-structure experts. (E) Double-blinded human-expert evaluation. Domain experts compared SciReasoner and DeepSeek-V4-Pro.

Model Architecture

SciReasoner integrates structural data via discrete tokenization and a multi-stage curriculum to align modalities.

Generalist language models struggle with scientific data because standard text tokenizers fragment physical structures into arbitrary, meaningless sub-words. SciReasoner resolves this by treating structural data as discrete, holistic tokens that preserve physical semantics before they ever reach the language model.

Instead of forcing a 3D molecule or protein into a text-based tokenizer, we use domain-specific encoders to map raw structures into a sequence of discrete, holistic tokens that represent physical motifs and symmetries.

SciReasoner treats structural tokens identically to linguistic tokens, projecting both into a shared semantic space so the model can reason about physical evidence and textual instructions simultaneously.

Stage 1 (Warm-up): Freeze the transformer backbone; train only the structural vocabulary and prediction head to map structural tokens into the linguistic space.

Stage 2 (Full-parameter): Unfreeze all parameters; train on diverse structure-text pairs using a shared Warmup-Stable-Decay (WSD) scheduler.

Stage 3 (Annealing): Continue full-parameter training with an increased proportion of QA-style data to refine reasoning capabilities.

Post-training Curriculum

Post-training refines SciReasoner via a two‑stage expert specialization and unified reasoning consolidation.

The pretrained checkpoint $M_0$ behaves as a plain next‑token generator, not an instruction‑following reasoner. Pooling all chain‑of‑thought (CoT) supervision into a single pass creates destructive interference because scientific tasks exhibit divergent reasoning patterns.

First we train domain‑specialist experts that learn how structural tokens act as evidence, then we merge those experts into a single model that can reason across domains without the interference that joint training caused.

Expert P receives $t_1^{P}, t_2^{P}$ and learns that $t_1^{P}$ signals a binding site while $t_2^{P}$ signals a secondary structure motif.

Expert M receives $t_1^{M}, t_2^{M}$ and learns that $t_1^{M}$ encodes a functional group and $t_2^{M}$ encodes a ring size.

During regeneration, Expert P produces a CoT trace “binding site → motif → answer”, while Expert M produces “functional group → ring size → answer”.

The pooled trace set now contains two domain‑specific rationales that the later unified pass can attend to without mixing the protein and molecule evidence.

Specializing per domain preserves the semantics of structural tokens; when the traces are later merged, the model can still tell which token belongs to which scientific context.

How does this two‑stage scheme differ from a naïve multi‑task fine‑tuning that simply mixes all tasks together?

Naïve mixing lets gradients from unrelated domains collide, erasing domain‑specific token‑evidence mappings. The two‑stage scheme isolates those mappings in per‑group experts first, then recombines only the cleaned‑up rational traces, so each domain’s structural knowledge remains intact.

Initialize each per‑group expert from the shared checkpoint $M_0$.

Feed a compact reasoning‑formatted corpus where every response follows a deliberation template demarcated by explicit anchors.

Train the expert on its domain’s structural‑evidence supervision, learning how structural tokens serve as evidence.

After convergence, let each expert generate rational CoT traces for its own tasks.

Collect all regenerated traces together with non‑thinking data (pure input‑output pairs without reasoning).

Run a single unified pass that combines the cold‑start objective with reinforcement learning, producing the final model.

Reinforcement Learning

Describe how stochastic rollouts and difficulty‑centred sampling generate informative RL data.

Reinforcement learning for scientific tasks suffers from two extremes: examples that are trivially solved give no learning signal, while hopeless cases produce only negative gradients. The method therefore needs a curated pool of “in‑the‑middle” examples that can actually benefit from policy improvement.

Instead of a single deterministic answer, we ask the model to produce several independent completions per prompt, letting randomness expose the range of behaviours the current policy can generate.

Rollout 1 returns “100 °C” → $f_1=1$.

Rollout 2 returns “water boils at 100 °C” → $f_2=1$.

Rollout 3 returns “It boils at 212 °F” → $f_3=1$.

Rollout 4 returns “It’s hot” → $f_4=0$.

Rollout 5 returns “100 °C under standard pressure” → $f_5=1$.

Rollout 6 returns “boiling point is 100 °C” → $f_6=1$.

Rollout 7 returns “water’s boiling point is 100 °C” → $f_7=1$.

Rollout 8 returns “water boils at 100 °C” → $f_8=1$.

Compute $\hat p(x) = (1+1+1+0+1+1+1+1)/8 = 0.875$.

The single failure drags $\hat p(x)$ below 1, flagging the example as partially solved and thus useful for RL learning.

How does stochastic rollout differ from ordinary beam search sampling?

Beam search deterministically keeps the top‑$k$ most likely sequences, discarding randomness; stochastic rollout deliberately injects randomness (via temperature and seed) to expose both successes and failures of the current policy, which is essential for estimating a meaningful solve rate.

We keep the examples whose empirical solve rate sits closest to a 50 % success chance, because those are the cases where the policy’s decisions actually matter – the model can still succeed or fail depending on its actions.

Why not simply sample uniformly from $D_{\text{train}}^{k}$?

Uniform sampling would waste many examples that are already solved or hopeless, yielding near‑zero advantage estimates and noisy gradients. Difficulty‑centred sampling concentrates the RL signal on the region where the policy’s improvement potential is highest.

With the 2 000‑example pool $T_k$ in hand, we train the policy using DAPO, a PPO‑style algorithm that operates on groups of candidate completions.

We treat each prompt as a group of $G$ candidate completions, compute a standardized advantage for each, and update the policy with a clipped importance‑ratio objective that preserves exploration.

Read the original paper

Open the simplified reader on Paperglide

Browse all simplified papers