Pith. sign in

REVIEW 4 major objections 8 minor 92 references

Reviewed by Pith at T0; open to challenge.

T0 means a machine referee read the full paper against a public rubric. The mark states how deep the mechanical check went, never who wrote it. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · glm-5.2

One model reasons across proteins, molecules, and crystals by citing structural evidence

2026-07-09 01:29 UTC pith:N5QYTGJ7

load-bearing objection Ambitious integration of structure-aware tokens across three scientific domains with strong protein/chemistry results, but materials benchmarks have a plausible leakage problem and reproducibility is currently zero. the 4 major comments →

arxiv 2607.07708 v1 pith:N5QYTGJ7 submitted 2026-07-08 cs.CL cs.AIcs.CEcs.LG

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
show 21 more authors
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
This is my paper
classification cs.CL cs.AIcs.CEcs.LG
keywords structural reasoningscientific foundation modelstructure-aware vocabularyautoregressive reasoningmultimodalprotein function predictionretrosynthesiscrystal properties
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

SciReasoner is a foundation model that converts 3D coordinates, molecular topologies, and crystal lattices into a unified token vocabulary, then treats those tokens as addressable evidence within a single autoregressive reasoning trajectory. The central claim is that when structural information is represented as discrete tokens that preserve domain-native semantics—rather than as text strings or black-box inputs—a single model can reason across proteins, small molecules, and inorganic crystals while producing inspectable chains of thought that cite specific residues, molecular fragments, or crystallographic features. The paper argues this is the first model to unify sequence, structure, and language reasoning across all four scientific modalities (proteins, DNA/RNA, small molecules, crystals) within one autoregressive pass, achieving state-of-the-art on 67 of 86 benchmarks. The key mechanism is a two-stage post-training procedure: first, domain-specific experts learn to use structural tokens as reasoning evidence via reinforcement learning; then, expert-generated reasoning traces are pooled and consolidated into a single unified model, avoiding the destructive interference that would arise from jointly training all domains at once.

Core claim

The paper's central object is the structure-aware token—a discrete representation of local geometry, bonding, symmetry, or conformation that preserves scientific semantics and can be cited as evidence within a reasoning chain. The core discovery is that by discretizing structural information into such tokens (using Foldseek for proteins, ConfSeq for molecules, SLICES for crystals) and integrating them into a language model's vocabulary, structural features become inspectable reasoning substrates rather than opaque input descriptors. The authors demonstrate this through three probe results: (1) protein function prediction improves most in low-homology regimes where sequence similarity is un帮助

What carries the argument

Structure-aware tokens + two-stage post-training (intra-domain expert grounding via RL, then cross-domain consolidation)

Load-bearing premise

The cross-domain consolidation step assumes that reasoning patterns learned independently by domain experts (proteins, molecules, materials) are compatible when merged into a single model, but the paper provides no formal guarantee against destructive interference during this merge.

What would settle it

If structural tokens are ablated and performance does not drop, the tokens are not load-bearing for reasoning.

Watch this falsifier — get emailed when new claim-graph text bears on it.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 8 minor

Summary. The manuscript introduces SciReasoner, a multimodal foundation model for scientific reasoning across proteins, small molecules, and inorganic crystals. The core architectural idea is a structure-aware vocabulary that discretizes 3D coordinates, molecular topologies, and crystallographic lattices into domain-native tokens (via Foldseek, ConfSeq, and SLICES encoders), integrated into a Qwen3-14B backbone through a dedicated embedding layer. A multi-stage pretraining curriculum (warm-up, full-parameter, annealing) is followed by a two-stage post-training framework (intra-domain structural evidence grounding via RL, then cross-domain reasoning consolidation). The paper evaluates on 86 benchmarks spanning biology, chemistry, and materials science, reporting SOTA on 67 tasks. Key qualitative claims include: (1) GO prediction gains are largest in low-homology regimes, (2) retrosynthesis traces are auditable and chemically grounded, (3) materials representations separate polymorphs and band-gap regimes, and (4) double-blind expert evaluation (N=1776 judgments) prefers SciReasoner over DeepSeek-V4-Pro in 98% of comparisons.

Significance. The paper addresses a genuine gap: unifying structure-native representation with inspectable reasoning traces across multiple scientific domains in a single autoregressive model. The structure-aware vocabulary design—treating structural tokens as addressable evidence units rather than auxiliary descriptors—is a clean and well-motivated architectural contribution. The homology-stratified GO analysis (Fig. 2A) and the structure-ablation experiments (Fig. 5) provide falsifiable evidence that performance gains are not driven by sequence-similarity shortcuts. The retrosynthesis case studies (Fig. 3B-C) with atom-by-auditable SMILES fragments in the reasoning trace are a concrete strength. The double-blind human evaluation protocol (Appendix B) with a structured rubric and ground-truth fact sheets is more rigorous than typical LLM-judge-only evaluations. The self-bootstrapped post-training framework that avoids pooling heterogeneous CoT supervision in a single pass is a reasonable methodological contribution to the RL-for-reasoning literature.

major comments (4)
  1. §4.1.3 and Table A3: The materials data split uses '80/10/10 random split at the material-sample level' with only exact-sample removal from pretraining. Unlike proteins (>30% sequence identity filtering, §4.1.1) or molecules (canonicalized identifier exclusion, §4.1.2), the materials split does not control for structural or compositional similarity. This is a load-bearing concern because JARVIS-QETB is listed as both a pretraining source (§4.1.3) and a benchmark target (Table A3), and the reported MAD/MAE = 108.98 vs 0.73 for the next-best model is a ~149x improvement—far beyond typical ML gains. Similarly, GNoME shows 21.91 vs 1.94 (~11x). These magnitudes could reflect the model exploiting isostructural or compositionally similar compounds in train and test rather than learning genuine structure-property relationships. The MAD/MAE metric itself can amplify this: if the target variance,
  2. §2.5, Fig. 6E: The double-blind human evaluation compares SciReasoner only against DeepSeek-V4-Pro, a general-purpose LLM, rather than against domain specialists or structure-aware baselines (e.g., SaProt for proteins, RSGPT for chemistry). The 98% preference rate (8.7/10 vs 4.3/10) is so one-sided that it raises the question of whether the comparison is informative about SciReasoner's scientific reasoning quality or primarily about the weakness of general-purpose LLMs on structure-intensive tasks. The per-axis scores for DeepSeek-V4-Pro (e.g., 3.9/10 on reasoning coherence) suggest a baseline that may not be competitive enough to establish that SciReasoner's traces meet a high scientific standard. Including at least one domain-specialist baseline in the human evaluation would substantially strengthen the claim that the reasoning traces are genuinely useful to experts, not merely better-
  3. §4.4, Fig. 6B-C: The cross-domain reasoning consolidation step pools expert-generated traces and performs a single RL pass, but the paper does not isolate whether consolidation preserves or degrades individual expert capabilities relative to the experts themselves. Fig. 6B shows reward dynamics and Fig. 6C shows pass@1/pass@10, but neither directly compares the unified model against the domain experts on their respective domains. If the unified model underperforms the experts on some domains, this would indicate destructive interference during consolidation—a risk the paper acknowledges (§4.4: 'joint training induces destructive interference') but does not empirically rule out. A table comparing per-domain performance of the unified model vs. each domain expert would address this.
  4. Tables A2-A4: No error bars, confidence intervals, or significance tests are reported for any of the 86 benchmark results. Given that several margins are narrow (e.g., GO-MF 0.66 vs 0.67 for SaProt in Table A1, Non-coding RNA family 0.90 vs 0.89 for RNA-MSM), it is unclear whether these differences are statistically meaningful. The absence of variance estimates is particularly consequential for the headline claim of 'SOTA on 67/86 tasks,' as some of these 67 may not be statistically distinguishable from the second-best model.
minor comments (8)
  1. §2.2.1, Fig. 2D-E: The reasoning-quality evaluation relies on GPT-5.5 as an LLM judge. The paper does not report inter-rater reliability between GPT-5.5 and the human evaluators, making it difficult to assess whether the LLM judge scores generalize to expert assessments.
  2. Table A3: The Cantor HEA result (MAD/MAE = 7.79 for SciReasoner vs 8.40 for LLM-Prop) is the only materials task where SciReasoner underperforms the specialist baseline. This is not discussed in the text.
  3. Fig. 1C: The tokenizer comparison shows compression ratio but does not report vocabulary size for the structure-aware vocabulary, which is relevant for assessing the embedding parameter count (W_v in Eq. 1).
  4. §4.1.3: The footnote numbering for materials databases (footnotes 7-11) appears to use a different numbering scheme than the main text footnotes.
  5. Fig. 4A: The y-axis label and metric for the 10 materials sub-tasks are not clearly specified in the figure caption; the reader must cross-reference Table A3 to determine whether each bar is MAE or AUC.
  6. Table A1: Several entries show SciReasoner underperforming the specialist (LIPO RMSE 0.80 vs 0.65, Human PPI ACC 0.73 vs 0.77, Solubility ACC 0.72 vs 0.77, GO-MF 0.66 vs 0.67). These are not discussed in the main text, which states SciReasoner 'matches or exceeds' specialists.
  7. §2.2.3: The DUD-E evaluation extracts embeddings from 'the 10 tokens generated immediately after the prompt.' The sensitivity of the AUC/EF results to this arbitrary choice of 10 tokens is not discussed.
  8. The paper does not mention whether code, data, or model weights will be released, which limits independent reproducibility of the 86 benchmark results.

Simulated Author's Rebuttal

4 responses · 0 unresolved

We thank the referee for a careful and constructive report. The referee correctly identifies the core contributions of SciReasoner and raises four substantive concerns: (1) the materials data split lacks structural/compositional similarity filtering, and the MAD/MAE gains on JARVIS-QETB and GNoME are suspiciously large; (2) the human evaluation compares only against a general-purpose LLM rather than domain specialists; (3) the cross-domain consolidation step does not empirically rule out destructive interference relative to per-domain experts; and (4) no error bars or significance tests are reported across the 86 benchmarks. We agree that all four points identify genuine gaps in the current manuscript and will address each in revision. Below we respond point by point.

read point-by-point responses

  1. Referee: §4.1.3 and Table A3: The materials data split uses 80/10/10 random split at the material-sample level with only exact-sample removal from pretraining, unlike proteins (>30% sequence identity filtering) or molecules (canonicalized identifier exclusion). JARVIS-QETB is listed as both a pretraining source and benchmark target, and the reported MAD/MAE = 108.98 vs 0.73 (~149x) and GNoME 21.91 vs 1.94 (~11x) could reflect isostructural or compositionally similar compounds leaking between train and test.

    Authors: The referee is correct on all counts. The materials data split in the current manuscript uses only exact-sample removal, which is inconsistent with the stricter leakage controls applied for proteins (>30% sequence identity filtering) and molecules (canonicalized identifier exclusion). This is a genuine weakness in our experimental design. Furthermore, the MAD/MAE magnitudes on JARVIS-QETB (108.98 vs 0.73) and GNoME (21.91 vs 1.94) are far beyond what is typically observed in materials ML and could plausibly be inflated by structural or compositional similarity between training and test compounds. We acknowledge that without similarity-based filtering, we cannot rule out this explanation. In the revision, we will: (1) implement structural similarity filtering for the materials split, using composition-based (e.g., SMACT-style elemental overlap) and structure-based (e.g., pymatgen StructureMatcher with default distance/angle tolerances) deduplication across train/test boundaries, applied consistently to all materials benchmarks including JARVIS-QETB and GNoME; (2) re-run all materials benchmarks with the corrected split and report updated MAD/MAE values; (3) add an explicit discussion of why the original magnitudes were suspicious and how the corrected numbers compare; and (4) clarify in §4.1.3 that JARVIS-QETB and GNoME entries in the pretraining corpus had their test-split material samples removed but that this was insufficient without similarity filtering. If the corrected numbers remain substantially above baselines, we will provide additional analysis (e.g., stratification by compositional novelty) to support the claim. If they do not, we will report the corrected numbers honestly and adjust our claims accordingly. revision: yes

  2. Referee: §2.5, Fig. 6E: The double-blind human evaluation compares SciReasoner only against DeepSeek-V4-Pro, a general-purpose LLM, rather than against domain specialists or structure-aware baselines. The 98% preference rate is so one-sided that it raises the question of whether the comparison is informative about SciReasoner's scientific reasoning quality or primarily about the weakness of general-purpose LLMs on structure-intensive tasks.

    Authors: The referee raises a valid concern. The 98% preference rate against a single general-purpose LLM baseline does not by itself establish that SciReasoner's reasoning traces meet a high scientific standard—it could partly reflect the difficulty general-purpose LLMs face on structure-intensive tasks. We agree that including at least one domain-specialist baseline would substantially strengthen the evaluation. In the revision, we will expand the human evaluation to include a structure-aware domain specialist (e.g., SaProt for protein GO prediction, and a chemistry-specialist model for retrosynthesis) as an additional comparison arm. We will retain the DeepSeek-V4-Pro comparison as a general-purpose reference but will frame the results as 'SciReasoner vs. general-purpose LLM' rather than as an absolute quality claim. We will also add discussion acknowledging that the one-sided margin against DeepSeek-V4-Pro is partly attributable to the absence of structural tokenization in general-purpose LLMs, and that the specialist comparison is the more informative test of whether the reasoning traces are genuinely useful to experts. We note that recruiting qualified domain experts for double-blind evaluation is logistically constrained, and we may not be able to add all three domain specialists in time for the next revision, but we commit to including at least one. revision: partial

  3. Referee: §4.4, Fig. 6B-C: The cross-domain reasoning consolidation step pools expert-generated traces and performs a single RL pass, but the paper does not isolate whether consolidation preserves or degrades individual expert capabilities relative to the experts themselves. A table comparing per-domain performance of the unified model vs. each domain expert would address this.

    Authors: This is a fair and important point. The current manuscript shows reward dynamics (Fig. 6B) and pass@1/pass@10 (Fig. 6C) but does not directly compare the unified model against the per-domain experts on their respective domains. Without this comparison, we cannot empirically demonstrate that consolidation preserves expert capabilities rather than degrading them. We acknowledge this gap. In the revision, we will add a table comparing the unified SciReasoner against each domain-structure expert (protein, molecule, material) on the representative tasks used in Fig. 6C (GO prediction, DUD-E, QMOF, Retrosynthesis USPTO-50K). If the unified model underperforms any expert on its domain, we will report this transparently and discuss it as evidence of partial destructive interference, consistent with the caveat we already note in §4.4. If the unified model matches or exceeds all experts, this will strengthen the consolidation claim. Either way, the comparison will be included. revision: yes

  4. Referee: Tables A2-A4: No error bars, confidence intervals, or significance tests are reported for any of the 86 benchmark results. Several margins are narrow (e.g., GO-MF 0.66 vs 0.67 for SaProt, Non-coding RNA family 0.90 vs 0.89 for RNA-MSM), and it is unclear whether these differences are statistically meaningful, particularly for the headline claim of 'SOTA on 67/86 tasks.'

    Authors: The referee is correct. The absence of variance estimates is a significant omission, especially given that several of the 67 claimed SOTA results have narrow margins where statistical distinguishability is uncertain. We will address this in the revision by: (1) reporting confidence intervals or standard errors for all benchmark results where multiple evaluation runs or bootstrap resampling is feasible; (2) applying paired significance tests (e.g., bootstrap or McNemar's test for classification, bootstrap CI for regression metrics) for the narrow-margin comparisons the referee identifies, including GO-MF (0.66 vs 0.67) and Non-coding RNA family (0.90 vs 0.89); and (3) revising the headline claim to distinguish between tasks where SciReasoner is statistically significantly better than the second-best model and tasks where the difference is within noise. We will report the revised count as 'SOTA on N of 86 tasks with statistical significance (at p<0.05)' alongside the original count, and will list the borderline cases explicitly. We acknowledge that for some benchmarks, only a single evaluation run was performed due to computational cost, and in those cases we will use bootstrap resampling over the test set to estimate confidence intervals rather than requiring full re-runs. revision: yes

Circularity Check

0 steps flagged

No significant circularity: benchmark results are evaluated against external datasets with stated leakage controls, and the central derivation is self-contained.

full rationale

The paper's central claim—SOTA on 67 of 86 benchmarks—is evaluated against external datasets (DeepFRI-GO, USPTO-50K, DUD-E, Materials Project, JARVIS-DFT, etc.) with explicitly stated leakage controls (§4.1.1: >30% sequence identity filtering for proteins; §4.1.2: canonicalized molecular identifier exclusion; §4.1.3: removal of test material samples from pretraining). The model architecture (§4.2) is initialized from Qwen3-14B and trained via a multi-stage pipeline using standard next-token prediction (Eq. 2) and DAPO reinforcement learning (Eqs. 11-13). The post-training framework (§4.4) uses self-bootstrapped reasoning traces, but this is a training methodology, not a circular definition: the RL reward is computed from task-specific correctness metrics (exact match, F1, MAE, etc.) that are externally defined, not from the model's own outputs fed back as ground truth. The reasoning-quality evaluation uses GPT-5.5 as judge (§2.2.1, Fig. 2D-E), which creates a dependency on another LLM's outputs, but this is a methodological choice for evaluating reasoning traces, not a circular derivation of the model's core predictions. The structure-aware vocabulary (§4.2.1-4.2.2) uses domain-specific encoders (Foldseek, SLICES, ConfSeq) that are externally developed tools, not defined in terms of SciReasoner's own outputs. No equation or definition in the paper reduces to its own inputs by construction. The self-bootstrapped post-training (§4.4) generates on-policy traces from expert models and pools them for unified training, but the final model's performance is measured against held-out external benchmarks, not against the expert-generated traces themselves. The one minor concern is the LLM-as-judge evaluation creating a partial dependency on GPT-5.5, but this affects only the reasoning-quality assessment (a secondary claim), not the primary benchmark results or the architectural derivation. No self-citation chain is load-bearing for the central mathematical or architectural claims. The derivation is self-contained against external benchmarks, so the circularity score is 1, not higher. The materials data split concern (random split without structural-similarity filtering, JARVIS-QETB appearing in both pretraining and benchmark) is a correctness risk, not a circularity issue, as the benchmark targets are externally defined DFT properties, not quantities defined in terms of the model's own predictions. The extraordinary magnitudes on some materials tasks (JARVIS-QETB, G

Axiom & Free-Parameter Ledger

6 free parameters · 6 axioms · 2 invented entities

The axiom ledger captures the key dependencies and assumptions. The free parameters are mostly standard RL/training hyperparameters, but several critical values (vocabulary size, DAPO clip bounds, reward softening function) are not reported, limiting reproducibility. The domain assumptions about tokenizer fidelity are inherited from prior work but are load-bearing for this model's claims. The ad-hoc assumptions about LLM-as-judge and cross-domain compatibility are the most fragile.

free parameters (6)
  • Structure-aware vocabulary embedding matrix W_v = learned (size not specified)
    New embedding table for structural tokens, trained during Stage 1-3. Vocabulary size |V_v| is not reported.
  • RL training temperature T = tuned per sub-task, starting from 0.9
    Adaptive temperature for stochastic rollouts, tuned once per sub-task before collection (§4.4.2).
  • RL subset size K = 2000
    Number of examples selected per sub-task for RL training (§4.4.2, Eq. 10). Chosen by the authors.
  • Solve-rate filter bounds = 0.125 and 0.875
    Empirical solve-rate thresholds for mid-difficulty filtering of RL training data (§4.4.2, Eq. 8).
  • DAPO clip parameters = not specified
    Asymmetric clipping bounds epsilon_low and epsilon_high are mentioned (§4.4.2, Eq. 12) but values are not reported.
  • Reward softening function g(·) = not specified
    Monotone calibration function for converting task metrics to [0,1] rewards (§4.4.2, Eq. 13). Form not specified.
axioms (6)
  • domain assumption Foldseek 3Di tokens faithfully represent protein structural environments sufficient for functional inference
    The entire protein structure representation relies on 3Di tokens from Foldseek (§4.1.1, §4.2.1). If 3Di loses information critical for function, the model's structural reasoning is compromised.
  • domain assumption SLICES representation preserves crystallographic information sufficient for property prediction
    Crystal structures are encoded via SLICES (§4.2.1). The paper assumes this discrete representation captures enough of the periodic connectivity and symmetry for materials tasks.
  • domain assumption ConfSeq tokens preserve 3D molecular geometry sufficient for similarity and property tasks
    3D molecule structures are encoded via ConfSeq (§4.2.1). The DUD-E virtual screening results depend on this representation capturing binding-pocket geometry.
  • ad hoc to paper GPT-5.5 is a reliable judge of scientific reasoning quality
    The reasoning-quality evaluation in Fig. 2D-E uses GPT-5.5 to score traces on five axes. This assumes the LLM judge aligns with expert human judgment of reasoning quality.
  • ad hoc to paper Domain experts' reasoning patterns are compatible for cross-domain consolidation
    The cross-domain consolidation step (§4.4) pools traces from independently trained domain experts. This assumes no destructive interference, which is not formally proven.
  • domain assumption AlphaFold-predicted structures with pLDDT<70 can be safely masked without losing critical information
    §4.1.1 states structurally unreliable regions (pLDDT<70) are masked or replaced. This assumes the remaining structure is sufficient for functional inference.
invented entities (2)
  • Structure-aware vocabulary (unified structural token set) independent evidence
    purpose: Discrete cross-modal projection layer mapping structural tokens into LLM embedding space
    The vocabulary is evaluated on 86 external benchmarks with leakage controls. Its utility is tested via structure-ablation experiments (Fig. 5) showing performance drops when structural inputs are removed.
  • Self-bootstrapped native structural reasoning (post-training framework) independent evidence
    purpose: Two-stage RL procedure: intra-domain grounding then cross-domain consolidation
    The framework is evaluated via pass@1/pass@10 improvements (Fig. 6C) and reasoning-quality distributions (Fig. 6D) across four task families. The ablation in Fig. 5 provides additional evidence.

pith-pipeline@v1.1.0-glm · 47231 in / 2987 out tokens · 484256 ms · 2026-07-09T01:29:28.288367+00:00 · methodology

0 comments
read the original abstract

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.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

92 extracted references · 92 canonical work pages · 5 internal anchors

  1. [1]

    C., Gogotsi, Y

    Thakur, A., Kim, J., Wyatt, B. C., Gogotsi, Y. & Anasori, B. Composition–structure–property relationships in mxenes.Nature Reviews Materials1–16 (2026)

    work page 2026

  2. [2]

    C., Burden, F

    Le, T., Epa, V. C., Burden, F. R. & Winkler, D. A. Quantitative structure–property relationship modeling of diverse materials properties.Chemical reviews112, 2889–2919 (2012)

    work page 2012

  3. [3]

    A complete representation of structure–property relationships in crystals.Nature materials 7, 455–458 (2008)

    Van de Walle, A. A complete representation of structure–property relationships in crystals.Nature materials 7, 455–458 (2008)

    work page 2008

  4. [4]

    E.Proteins: structures and molecular properties(Macmillan, 1993)

    Creighton, T. E.Proteins: structures and molecular properties(Macmillan, 1993)

    work page 1993

  5. [5]

    Tai, T.-Y.et al.Leveraging isothermal titration calorimetry to explore structure–property relationships of pro- tein immobilization in metal–organic frameworks.Angewandte Chemie International Edition61, e202209110 (2022)

    work page 2022

  6. [6]

    & Kolodny, R

    Osadchy, M. & Kolodny, R. Maps of protein structure space reveal a fundamental relationship between protein structure and function.Proceedings of the National Academy of Sciences108, 12301–12306 (2011)

    work page 2011

  7. [7]

    Anfinsen, C. B. Principles that govern the folding of protein chains.Science181, 223–230 (1973)

    work page 1973

  8. [8]

    Abramson, J.et al.Accurate structure prediction of biomolecular interactions with alphafold 3.Nature630, 493–500 (2024)

    work page 2024

  9. [9]

    & Grossman, J

    Xie, T. & Grossman, J. C. Crystal graph convolutional neural networks for an accurate and interpretable prediction of material properties.Physical review letters120, 145301 (2018)

    work page 2018

  10. [10]

    Merchant, A.et al.Scaling deep learning for materials discovery.Nature624, 80–85 (2023)

    work page 2023

  11. [11]

    Strieth-Kalthoff, F.et al.Artificial intelligence for retrosynthetic planning needs both data and expert knowledge.Journal of the American Chemical Society146, 11005–11017 (2024)

    work page 2024

  12. [12]

    H., Preuss, M

    Segler, M. H., Preuss, M. & Waller, M. P. Planning chemical syntheses with deep neural networks and symbolic ai.Nature555, 604–610 (2018)

    work page 2018

  13. [13]

    & Lane, L

    Duek, P., Mary, C., Zahn-Zabal, M., Bairoch, A. & Lane, L. Functionathon: a manual data mining workflow to generate functional hypotheses for uncharacterized human proteins and its application by undergraduate students.Database2021, baab046 (2021)

    work page 2021

  14. [14]

    Zhao, D., Tu, S. & Xu, L. Efficient retrosynthetic planning with mcts exploration enhanced a* search. Communications Chemistry7, 52 (2024)

    work page 2024

  15. [15]

    Achiam, J.et al.Gpt-4 technical report.arXiv preprint arXiv:2303.08774(2023)

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  16. [16]

    Mirza, A.et al.Are large language models superhuman chemists?Nature Chemistry(2025)

    work page 2025

  17. [17]

    Narayanan, S.et al.Training a scientific reasoning model for chemistry.Advances in Neural Information Processing Systems38, 157671–157710 (2026)

    work page 2026

  18. [18]

    A., MacKnight, R., Kline, B

    Boiko, D. A., MacKnight, R., Kline, B. & Gomes, G. Autonomous chemical research with large language models.Nature624, 570–578 (2023)

    work page 2023

  19. [19]

    E.et al.A multi-agent system for automating scientific discovery.Nature1–3 (2026)

    Ghareeb, A. E.et al.A multi-agent system for automating scientific discovery.Nature1–3 (2026)

    work page 2026

  20. [20]

    Ayg¨ un, E.et al.An ai system to help scientists write expert-level empirical software.Nature1–3 (2026)

    work page 2026

  21. [21]

    Gottweis, J.et al.Accelerating scientific discovery with co-scientist.Nature1–3 (2026)

    work page 2026

  22. [22]

    Wang, H.et al.Scientific discovery in the age of artificial intelligence.Nature620, 47–60 (2023)

    work page 2023

  23. [23]

    Yang, K.et al.Analyzing learned molecular representations for property prediction.Journal of chemical information and modeling59, 3370–3388 (2019)

    work page 2019

  24. [24]

    System card: Claude opus 4.7

    Anthropic. System card: Claude opus 4.7. https://anthropic.com/claude-opus-4-7-system-card (2026)

    work page 2026

  25. [25]

    Introducing GPT-5.5

    OpenAI. Introducing GPT-5.5. https://openai.com/index/introducing-gpt-5-5/ (2026). Accessed: 2026-06-05. 39

    work page 2026

  26. [26]

    Deepseek-v4: Towards highly efficient million-token context intelligence (2026)

    DeepSeek-AI. Deepseek-v4: Towards highly efficient million-token context intelligence (2026)

    work page 2026

  27. [27]

    Kimi k2.6 tech blog: Advancing open-source coding

    Kimi. Kimi k2.6 tech blog: Advancing open-source coding. https://www.kimi.com/blog/kimi-k2-6 (2026). Accessed: 2026-07-07

    work page 2026

  28. [28]

    Su, J.et al.Democratizing protein language model training, sharing and collaboration.Nature Biotechnology 1–7 (2025)

    work page 2025

  29. [29]

    URL https://www.science.org/doi/abs/10.1126/science.ade2574

    Lin, Z.et al.Evolutionary-scale prediction of atomic-level protein structure with a language model.Science 379, 1123–1130 (2023). URL https://www.science.org/doi/abs/10.1126/science.ade2574

    work page doi:10.1126/science.ade2574 2023

  30. [30]

    F., Gish, W., Miller, W., Myers, E

    Altschul, S. F., Gish, W., Miller, W., Myers, E. W. & Lipman, D. J. Basic local alignment search tool.Journal of molecular biology215, 403–410 (1990)

    work page 1990

  31. [31]

    Deng, Y.et al.Rsgpt: a generative transformer model for retrosynthesis planning pre-trained on ten billion datapoints.Nature communications16, 7012 (2025)

    work page 2025

  32. [32]

    M., Carchia, M., Irwin, J

    Mysinger, M. M., Carchia, M., Irwin, J. J. & Shoichet, B. K. Directory of useful decoys, enhanced (dud-e): better ligands and decoys for better benchmarking.Journal of medicinal chemistry55, 6582–6594 (2012)

    work page 2012

  33. [33]

    Ashburner, M.et al.Gene ontology: tool for the unification of biology.Nature genetics25, 25–29 (2000)

    work page 2000

  34. [34]

    The gene ontology resource: enriching a gold mine.Nucleic acids research49, D325–D334 (2021)

    work page 2021

  35. [35]

    Consortium, T. U. Uniprot: the universal protein knowledgebase in 2023.Nucleic acids research51, D523– D531 (2023)

    work page 2023

  36. [36]

    Automated protein function prediction—the genomic challenge.Briefings in bioinformatics7, 225–242 (2006)

    Friedberg, I. Automated protein function prediction—the genomic challenge.Briefings in bioinformatics7, 225–242 (2006)

    work page 2006

  37. [37]

    Radivojac, P.et al.A large-scale evaluation of computational protein function prediction.Nature methods 10, 221–227 (2013)

    work page 2013

  38. [38]

    P., Hern´ andez-Plaza, A., Letunic, I., Bork, P

    Cantalapiedra, C. P., Hern´ andez-Plaza, A., Letunic, I., Bork, P. & Huerta-Cepas, J. eggnog-mapper v2: functional annotation, orthology assignments, and domain prediction at the metagenomic scale.Molecular biology and evolution38, 5825–5829 (2021)

    work page 2021

  39. [39]

    Gligorijevi´ c, V.et al.Structure-based protein function prediction using graph convolutional networks.Nature communications12, 3168 (2021)

    work page 2021

  40. [40]

    Van Kempen, M.et al.Fast and accurate protein structure search with foldseek.Nature biotechnology42, 243–246 (2024)

    work page 2024

  41. [41]

    M., Sayle, R

    Schneider, N., Lowe, D. M., Sayle, R. A., Tarselli, M. A. & Landrum, G. A. Big data from pharmaceutical patents: a computational analysis of medicinal chemists’ bread and butter.Journal of medicinal chemistry 59, 4385–4402 (2016)

    work page 2016

  42. [42]

    W., Rogers, L., Green, W

    Coley, C. W., Rogers, L., Green, W. H. & Jensen, K. F. Computer-assisted retrosynthesis based on molecular similarity.ACS central science3, 1237–1245 (2017)

    work page 2017

  43. [43]

    Segler, M. H. & Waller, M. P. Neural-symbolic machine learning for retrosynthesis and reaction prediction. Chemistry–A European Journal23, 5966–5971 (2017)

    work page 2017

  44. [44]

    & Song, L

    Dai, H., Li, C., Coley, C., Dai, B. & Song, L. Retrosynthesis prediction with conditional graph logic network. Advances in Neural Information Processing Systems32(2019)

    work page 2019

  45. [45]

    & Jung, Y

    Chen, S. & Jung, Y. Deep retrosynthetic reaction prediction using local reactivity and global attention.JACS Au1, 1612–1620 (2021)

    work page 2021

  46. [46]

    & Huang, J

    Yan, C., Zhao, P., Lu, C., Yu, Y. & Huang, J. Retrocomposer: composing templates for template-based retrosynthesis prediction.Biomolecules12, 1325 (2022)

    work page 2022

  47. [47]

    & Tang, J

    Shi, C., Xu, M., Guo, H., Zhang, M. & Tang, J. A graph to graphs framework for retrosynthesis prediction 8818–8827 (2020). 40

    work page 2020

  48. [48]

    Yan, C.et al.Retroxpert: Decompose retrosynthesis prediction like a chemist.Advances in Neural Information Processing Systems33, 11248–11258 (2020)

    work page 2020

  49. [49]

    Wang, X.et al.Retroprime: A diverse, plausible and transformer-based method for single-step retrosynthesis predictions.Chemical Engineering Journal420, 129845 (2021)

    work page 2021

  50. [50]

    R., Fuchs, J

    Chen, Z., Ayinde, O. R., Fuchs, J. R., Sun, H. & Ning, X. G 2 retro as a two-step graph generative models for retrosynthesis prediction.Communications Chemistry6, 102 (2023)

    work page 2023

  51. [51]

    Gao, Z., Tan, C., Wu, L. & Li, S. Z. Semiretro: Semi-template framework boosts deep retrosynthesis prediction. arXiv preprint arXiv:2202.08205(2022)

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  52. [52]

    & Chen, C

    Zhong, W., Yang, Z. & Chen, C. Y.-C. Retrosynthesis prediction using an end-to-end graph generative architecture for molecular graph editing.Nature Communications14, 3009 (2023)

    work page 2023

  53. [53]

    & Yang, Y

    Zheng, S., Rao, J., Zhang, Z., Xu, J. & Yang, Y. Predicting retrosynthetic reactions using self-corrected transformer neural networks.Journal of chemical information and modeling60, 47–55 (2019)

    work page 2019

  54. [54]

    Sacha, M.et al.Molecule edit graph attention network: modeling chemical reactions as sequences of graph edits.Journal of Chemical Information and Modeling61, 3273–3284 (2021)

    work page 2021

  55. [55]

    & Coley, C

    Tu, Z. & Coley, C. W. Permutation invariant graph-to-sequence model for template-free retrosynthesis and reaction prediction.Journal of chemical information and modeling62, 3503–3513 (2022)

    work page 2022

  56. [56]

    Zhong, Z.et al.Root-aligned smiles: a tight representation for chemical reaction prediction.Chemical Science 13, 9023–9034 (2022)

    work page 2022

  57. [57]

    Yao, L.et al.Node-aligned graph-to-graph: elevating template-free deep learning approaches in single-step retrosynthesis.JACS Au4, 992–1003 (2024)

    work page 2024

  58. [58]

    Xiong, J.et al.Bridging three-dimensional molecular structures and artificial intelligence with a conformation description language.Nature Machine Intelligence1–15 (2026)

    work page 2026

  59. [59]

    C., Steinhauer, V

    Hemmer, M. C., Steinhauer, V. & Gasteiger, J. Deriving the 3d structure of organic molecules from their infrared spectra.Vibrational spectroscopy19, 151–164 (1999)

    work page 1999

  60. [60]

    & Lesyk, R

    Devinyak, O., Havrylyuk, D. & Lesyk, R. 3d-morse descriptors explained.Journal of Molecular Graphics and Modelling54, 194–203 (2014)

    work page 2014

  61. [61]

    T., Kaiser, D

    Klein, C. T., Kaiser, D. & Ecker, G. Topological distance based 3d descriptors for use in qsar and diversity analysis.Journal of chemical information and computer sciences44, 200–209 (2004)

    work page 2004

  62. [62]

    D.et al.A simple representation of three-dimensional molecular structure.Journal of medicinal chemistry60, 7393–7409 (2017)

    Axen, S. D.et al.A simple representation of three-dimensional molecular structure.Journal of medicinal chemistry60, 7393–7409 (2017)

    work page 2017

  63. [63]

    J., Puranen, J

    Vainio, M. J., Puranen, J. S. & Johnson, M. S. Shaep: molecular overlay based on shape and electrostatic potential (2009)

    work page 2009

  64. [64]

    M., Dixon, S

    Sastry, G. M., Dixon, S. L. & Sherman, W. Rapid shape-based ligand alignment and virtual screening method based on atom/feature-pair similarities and volume overlap scoring.Journal of chemical information and modeling51, 2455–2466 (2011)

    work page 2011

  65. [65]

    Liu, X., Jiang, H. & Li, H. Shafts: a hybrid approach for 3d molecular similarity calculation. 1. method and assessment of virtual screening.Journal of chemical information and modeling51, 2372–2385 (2011)

    work page 2011

  66. [66]

    & Zhang, Y

    Hu, J., Liu, Z., Yu, D.-J. & Zhang, Y. Ls-align: an atom-level, flexible ligand structural alignment algorithm for high-throughput virtual screening.Bioinformatics34, 2209–2218 (2018)

    work page 2018

  67. [67]

    Niyongabo Rubungo, A., Arnold, C., Rand, B. P. & Dieng, A. B. Llm-prop: predicting the properties of crystalline materials using large language models.npj Computational Materials11, 186 (2025)

    work page 2025

  68. [68]

    Dana, J. M.et al.Sifts: updated structure integration with function, taxonomy and sequences resource allows 40-fold increase in coverage of structure-based annotations for proteins.Nucleic Acids Research47, D482–D489 (2019)

    work page 2019

  69. [69]

    M.et al.The protein data bank.Nucleic Acids Research28, 235–242 (2000)

    Berman, H. M.et al.The protein data bank.Nucleic Acids Research28, 235–242 (2000). 41

    work page 2000

  70. [70]

    https://pubmed.ncbi.nlm.nih.gov/ (2026)

    Pubmed. https://pubmed.ncbi.nlm.nih.gov/ (2026). Accessed: 2026-06-21

    work page 2026

  71. [71]

    https://www.ncbi.nlm.nih.gov/pmc/ (2026)

    Pubmed central. https://www.ncbi.nlm.nih.gov/pmc/ (2026). Accessed: 2026-06-21

    work page 2026

  72. [72]

    Varadi, M.et al.Alphafold protein structure database: massively expanding the structural coverage of protein-sequence space with high-accuracy models.Nucleic Acids Research50, D439–D444 (2022)

    work page 2022

  73. [73]

    Su, J.et al.Saprot: Protein language modeling with structure-aware vocabulary (2024)

    work page 2024

  74. [74]

    & S¨ oding, J

    Steinegger, M. & S¨ oding, J. Mmseqs2 enables sensitive protein sequence searching for the analysis of massive data sets.Nature biotechnology35, 1026–1028 (2017)

    work page 2017

  75. [75]

    Nature Communications16, 2017 (2025)

    Kshirsagar, M.et al.Rapid and accurate prediction of protein homo-oligomer symmetry using seq2symm. Nature Communications16, 2017 (2025)

    work page 2017

  76. [76]

    Wu, Z.et al.Moleculenet: a benchmark for molecular machine learning.Chemical science9, 513–530 (2018)

    work page 2018

  77. [77]

    Huang, K.et al.Therapeutics data commons: Machine learning datasets and tasks for drug discovery and development.Advances in neural information processing systems(2021)

    work page 2021

  78. [78]

    Frontiers in Pharmacology(2020)

    Polykovskiy, D.et al.Molecular Sets (MOSES): A Benchmarking Platform for Molecular Generation Models. Frontiers in Pharmacology(2020)

    work page 2020

  79. [79]

    NVIDIA Nemotron Nano 2: An Accurate and Efficient Hybrid Mamba-Transformer Reasoning Model

    NVIDIAet al.Nvidia nemotron nano 2: An accurate and efficient hybrid mamba-transformer reasoning model (2025). URL https://arxiv.org/abs/2508.14444. arXiv:2508.14444

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  80. [80]

    Olmo 3

    Olmo, T.et al.Olmo 3 (2025). URL https://arxiv.org/abs/2512.13961. arXiv:2512.13961

    work page internal anchor Pith review Pith/arXiv arXiv 2025

Showing first 80 references.