Pith. sign in

REVIEW 5 major objections 6 minor 67 references

SciForge argues that the bottleneck in AI-driven science is not the model but the missing persistent, auditable research state, and that a local-first workbench with evidence graphs and translate-then-reason input can supply it.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

An open-source research workbench that attaches provenance, audit, and human approval gates to agentic science workflows, with eight biological/chemical use-case demos.

T0 review reviewed 2026-08-01 challenge →

load-bearing objection An honest systems paper whose integrated design is real, but whose evidence trails are thinner than the word 'demonstrated' implies; §5.2's admission about incomplete provenance capture is the load-bearing gap. the 5 major comments →

arxiv 2607.16038 v1 pith:ME5ZDBA2 submitted 2026-07-17 cs.AI

SciForge: An AI-Native, Multimodal Workbench for Scientific Discovery

SciForge Team: Zhangyang Gao , Minghao Fang , Yifei Liu , Hanhui Yang , Xinyu Gu , Shixiang Tang , Siqi Sun , Lei Bai
show 4 more authors
Cheng Tan Mengdi Liu Hao Wu Shuizhou Chen
This is my paper
classification cs.AI
keywords AI research workbenchevidence governancescientific provenancemultimodal scientific inputagentic workflowsgoal-scoped governancereproducibilityscientific discovery
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.

The reading

Scientific AI assistants normally operate session-by-session, losing the connections between a claim, the dataset it came from, the script that produced it, the model run, and the human who approved it. SciForge claims this missing persistent research state is the central bottleneck, and that a local-first workbench can fix it structurally: a thin human-judgment interface over modular agent services, with every task flowing through one evidence-aware control chain. That chain translates scientific files into structured expert observations before reasoning, records provenance as an Evidence DAG and a goal-scoped Project DAG, runs read-only audits, and exposes candidate and certified release gates under human oversight. The paper argues that eight end-to-end use cases—ranging from gene discovery, protein and molecular design, to genome mining and paper reproduction—show this design can carry long-horizon, multi-session research while keeping the trail inspectable. A sympathetic reader should take the central claim as: auditable agent-driven science is achievable by architecture, not by prompting discipline.

Core claim

SciForge's central claim is architectural: agents fail to support real research not because of model capability but because research state—artifacts, actions, decisions, and evidence traces—is not preserved as a coherent, auditable whole. The paper proposes a five-pillar workbench whose graphical interface is reserved for human judgment while search, parsing, model routing, execution, plotting, writing, and presentation generation run as modular agent-accessible services. The load-bearing device is the Evidence DAG, a claim-source-reasoning graph built automatically from agent turns, snapshotted into a goal-scoped Project DAG that merges findings across sessions and carries human review deci

What carries the argument

The machinery that carries the argument is the two-layer evidence graph plus the router. The Evidence DAG is a thread-scoped graph automatically compiled from completed agent turns, linking claims to source assertions, reasoning nodes, and support or contradiction edges, with node-level provenance visible to the researcher. The Project DAG consumes immutable Evidence DAG snapshots across sessions, merges equivalent findings while preserving independent source paths, and provides candidate and certified release gates tied to human decision records. The Scientific Model Router adds translate-then-reason: scientific files are detected by modality and passed through specialized translators that

Load-bearing premise

The whole audit value proposition rests on the Evidence DAG capturing complete, faithful provenance from agent runs; the paper concedes that software, parameters, environment, logs, and seed are not automatically populated, so the graph is only as auditable as whatever the runtime happens to record.

What would settle it

Run a controlled scripted session with known steps—specific tool versions, parameters, and a fixed random seed—then export the Evidence DAG and compare its provenance fields against ground truth; or inject an unsupported claim into an agent session and check whether the read-only audit flags it as ungrounded before any human review.

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

If this is right

  • If the architecture holds, agent-generated scientific claims become inspectable end to end: each conclusion can be traced back to the files, scripts, model calls, parameters, and human approvals that produced it.
  • Multi-day, multi-session research stays continuous: goals, evidence snapshots, and review decisions persist in the Project DAG, so work does not reset when a session ends.
  • Scientific file formats stop being dead letters to language agents; proteins, structures, molecules, and single-cell data enter as structured expert observations rather than raw prompts, with unsupported formats failing closed.
  • The same evidence-aware control chain can restructure routine research tasks—manuscript review and rebuttal, paper reproduction, figure and presentation generation—into auditable revision packages.
  • Honest outcome reporting becomes a governance feature: in the molecular optimization use case the pre-registered primary criterion was not met, and that failure was recorded transparently rather than adjusted.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If provenance capture were made complete and enforced at execution time—software version, parameters, environment, logs, and seed—the audit sidecar would become a reproducibility linter for agent runs; the paper's own limitation note suggests this is the natural next step and a testable extension.
  • The audit claim is only as strong as the capture layer; a stress test with deliberately planted fabricated claims or omitted provenance would show whether the DAG exposes them before a human reviews.
  • The thin-GUI-plus-modular-services division of labour is not science-specific; the same pattern could govern other evidence-heavy, approval-gated domains such as regulated analysis or clinical decision support, though the paper does not say this.
  • The eight use cases vary in maturity; adopting pre-registered success criteria uniformly would let the workbench generate its own evaluation evidence rather than relying on narrative demonstrations.
Share X Bluesky LinkedIn Reddit HN

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

5 major / 6 minor

Summary. The paper describes SciForge, a local-first, AI-native research workbench designed to keep scientific artifacts, agent actions, and human decisions in one auditable state. Its architecture has five layers: thin human-judgment interaction surfaces; six research capability patterns; a core engine with an Agent Runtime, Workflow Engine, scoped memory, and evidence governance; a Scientific Model Router with translate-then-reason handling for protein, structure, molecule, and single-cell modalities; and local-first infrastructure. The claimed differentiators are goal-scoped decision governance, thread-level Evidence DAGs linked to a goal-scoped Project DAG, release gates, and multimodal scientific ingress. The system is 'demonstrated through eight end-to-end user cases,' including a multi-day agentic meiosis research sprint, protein contact prediction, reviewer/rebuttal support, MCFST paper reproduction, cross-scale cell atlas construction, de novo protein design, molecular optimization, and genome-to-BGC discovery. The paper is unusually candid: several use cases explicitly report unmet criteria, post-hoc thresholds, provenance mismatches, and the absence of baselines or third-party validation.

Significance. If the architecture works as claimed, SciForge would be a useful contribution to the emerging category of persistent, auditable AI research environments: it goes beyond single-session chat by linking claims to files, model calls, parameters, and human approvals through PROV-aligned Evidence/Project DAGs, and it makes a concrete open-source release with per-use-case repositories. The paper's main strengths are its modular architecture, fail-closed modality routing, explicit release-gate semantics, and the exceptionally transparent reporting of negative or ambiguous results (e.g., §4.7's unmet primary criterion, §4.4's post-hoc threshold, §4.6's provenance mismatch). These strengths are, however, self-assessed: there are no baselines, ablations, user studies, third-party reproductions, or capture-completeness benchmarks, and the central 'auditable traceability' claim depends on runtime provenance capture that the paper itself concedes is incomplete. The contribution is therefore best read as an architecture proposal with pilot illustrations, not as a validated demonstration of the auditability value proposition.

major comments (5)
  1. [§5.2, §3.4] The central claim of auditable traceability rests on automatic Evidence DAG capture, but §5.2 concedes that the DAG 'depends on runtime capture; software, parameters, environment, log/output, and seed are not automatically populated.' These are the same provenance fields the architecture (§3.4, §1.2) promises to attach to every agent action. Without a capture-completeness benchmark or independent audit measuring what fraction of model calls, tool invocations, file reads, and parameters actually appear in the DAG, the 'automatically constructs' claim is unsupported. I recommend either instrumenting the runtime to report capture coverage on the existing use cases or sharply narrowing the auditability claim in the abstract/contributions.
  2. [§4.4] The guided reproduction case is presented as an 'independently auditable evidence' demonstration, yet the reported evidence is internally inconsistent: the text says 25 independent training runs, verify.py reports n_runs=5, the 0.05 success threshold was applied post-hoc, and the selection rule for the five runs is not pre-specified. The best ARI of 0.7007 differs from the paper's 0.693 by +0.0077 with no stability analysis. This is precisely the kind of irreproducible verification record the workbench is supposed to prevent. The case should be reframed as a feasibility prototype (as the text later says) and the verification contract should be fixed with a pre-registered selection rule and reconciled run counts before the audit trail is presented as a positive demonstration.
  3. [§4.6] The self-audit reveals a provenance mismatch: the agent's final narrative cited ProteinMPNN scores (1.0363 and 1.0421) from unverified sequence samples rather than the scores of the sequences actually submitted to Boltz-2 (1.1039 and 1.1481). This demonstrates the failure mode in which an Evidence DAG can look coherent while pointing to the wrong artifacts. Since this is the flagship protein-design demonstration, the paper needs to show how the DAG and audit sidecar would surface such mismatches automatically — e.g., by linking each claimed score to the exact sequence hash and run record — rather than relying on a separate post-hoc audit.
  4. [§4.7] The abstract lists 'molecular optimization' as a demonstrated flagship scenario, but the primary pre-registered criterion was not met: the best improvement is −1.7 kcal/mol against a −2.0 threshold, and the text notes the erlotinib docking-score variance across replicates was ±2.0 kcal/mol. The observed improvement is therefore within the noise floor of the docking protocol. The honest reporting is commendable, but 'demonstrated' overstates the outcome; this should be described as a pilot/SAR illustration with the quantitative limitation stated in the abstract.
  5. [§5.2, §2.4] The paper's positioning against Claude Science, OmicOS, OmicsClaw, Operon, and research agents is based on a feature-level reading of public materials, and §5.2 explicitly states there are 'no baselines, ablation studies, user studies, or third-party reproductions.' For a systems paper whose contribution is the workbench claim, this leaves the main differentiators — evidence governance, capture completeness, and goal-scoped release semantics — unbenchmarked. I would not require a full user study for acceptance, but an evaluation section comparing at least provenance capture and task completion on a small standardized set (e.g., rerunning one case under two runtimes) would materially strengthen the central claim.
minor comments (6)
  1. [§4.1] The 'informal post-hoc literature mapping' is used to support the 23-gene atlas's plausibility. The paper already notes it was informal and not pre-registered; please mark it clearly as anecdotal or move it to supplementary so it is not read as validation.
  2. [§4.2] The repository URL 'https://github.com/BruthYU/autoresearch base' contains a space and is likely 'autoresearch-base'. Please fix the link.
  3. [§4.7] Figure 13 is referenced before Figure 12 in the text ('scaffold modification strategy is depicted in Fig. 13' precedes 'Figure 12 summarizes...'). Renumber or adjust the in-text references.
  4. [§3.2, §4] The abbreviation 'PI' is used without definition on first use; please define it (e.g., principal investigator) in §3 or §4.
  5. [Title page, Appendix C] The title-page attribution 'Written by SciForge with DeepSeek-v4-pro (text) and gpt-image-2 (figures), Guided and Verified by Humans' appears to conflict with Appendix C, which lists human manuscript contributors. Clarify the drafting/verification workflow or move the model attribution to a footnote.
  6. [Table 1] The dagger marker for the planned 'Team Workspace' capability is ambiguous about which column it annotates. Move the dagger into the SciForge column and state in the caption that the capability is planned, not implemented.

Circularity Check

0 steps flagged

No circular derivation; the workbench's core auditability claim rests on a stated validation gap, not on circular reasoning.

full rationale

SciForge is a systems paper: it does not derive predictions from fitted parameters or mathematical inputs. The central claim is architectural—that an Evidence DAG/Project DAG can make agent workflows auditable—and it is demonstrated through case studies. The load-bearing assumption is that runtime provenance capture is complete and faithful. The paper explicitly concedes in §5.2 that 'the Evidence DAG schema can represent complete provenance but depends on runtime capture; software, parameters, environment, log/output, and seed are not automatically populated,' and that 'no baselines, ablation studies, user studies, or third-party reproductions have been conducted.' These are validation gaps, not circular reductions. The MCFST reproduction (§4.4) discloses that the 0.05 success threshold was applied post-hoc and that verify.py reports n_runs=5, conflicting with 25 total predictions; this is an admitted statistical weakness, not a hidden fit called a prediction. The protein-design audit (§4.6) discloses a provenance mismatch between cited and executed ProteinMPNN scores; this is a self-audit finding that lowers confidence but does not define a result in terms of its own inputs. No self-citation chain or uniqueness theorem is load-bearing; external references are standard tools/translators (RFdiffusion, ProteinMPNN, Boltz-2, antiSMASH, etc.), and the workbench code is open-source. Therefore there is no significant circularity.

Axiom & Free-Parameter Ledger

2 free parameters · 3 axioms · 2 invented entities

The central architecture rests on unverified assumptions about translator fidelity, provenance completeness, and external-tool correctness; these are explicitly acknowledged in §5.2. No free parameters are fitted to data in a predictive sense, but two evaluation thresholds are hand-set/post-hoc and directly affect reported use-case outcomes.

free parameters (2)
  • MCFST success threshold = 0.05
    Applied post-hoc to select 5 of 25 runs in the MCFST reproduction; not pre-specified and contradicts verify.py run count. It directly affects the reported best ARI 0.7007 result in §4.4.
  • BGC rubric weights/thresholds (v0.1) = unknown
    Author-defined rubric for scoring 430 Candidate BGC Cards; paper states weights and thresholds 'have not been prospectively calibrated against expert consensus or experimental outcomes' (§4.8 Limitations).
axioms (3)
  • domain assumption The four domain translators (Esm2Text, Prot2Text, BioT5+, C2S) produce structured expert observations sufficiently reliable for agent reasoning about scientific files.
    The translate-then-reason pipeline in §3.5 routes all supported modalities through these translators; §5.2 states 'no systematic validation against ground-truth annotations has been performed.'
  • domain assumption Automatically captured Evidence Snapshots and the Evidence DAG accurately represent the provenance of agent actions (software, parameters, environment, logs, seed).
    §5.2 evidence completeness: 'depends on runtime capture; software, parameters, environment, log/output, and seed are not automatically populated.'
  • domain assumption External scientific resources (UniProt, Reactome, DepMap, antiSMASH, MIBiG, BiG-SCAPE, PDB, structure-prediction models) are treated as ground truth for use-case conclusions.
    Use cases 4.5–4.8 rely on these resources without independent verification; the paper does not audit their correctness or coverage.
invented entities (2)
  • Evidence DAG (thread-scoped) no independent evidence
    purpose: Automatically constructed claim–source–reasoning graph with provenance and audit metadata for each agent session.
    Core audit construct; implemented in code but no external benchmark validates its completeness or faithfulness. §3.4, Fig. B1.
  • Project DAG (goal-scoped) no independent evidence
    purpose: Cross-session aggregation of evidence snapshots, claims, decisions, and candidate/certified release gates.
    Core governance construct; implemented but no external evaluation. §3.4, Fig. B2.

reviewed 2026-08-01 · how reviews work

0 comments
Cite this review

Pith. "Pith review of SciForge: An AI-Native, Multimodal Workbench for Scientific Discovery." pith.science (2026). https://pith.science/paper/ME5ZDBA2

@misc{pith2026260716038,
  author       = {Pith},
  title        = {Pith review of: SciForge: An AI-Native, Multimodal Workbench for Scientific Discovery},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ME5ZDBA2}},
  note         = {Machine review of arXiv:2607.16038}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Scientific work increasingly spans heterogeneous artifacts -- papers, code, datasets, scientific file formats, model outputs, figures, manuscripts, and team decisions -- yet general-purpose AI assistants rarely preserve these objects as a coherent, auditable research state. We present SciForge, a multimodal research-native AI workbench that reserves the graphical interface for human judgment while search, parsing, model routing, workflow execution, plotting, writing, and presentation generation run as modular agent-accessible services. SciForge is built around five pillars: (i) \emph{goal-scoped scientific decision governance} for \textbf{goal-oriented} research, with review gates and shared review surfaces; (ii) \emph{translate-then-reason} for \textbf{multimodal} input, routing scientific objects through domain translators before the agent reasons; (iii) \emph{evidence governance} for \textbf{auditable} traceability, linking claims to provenance chains and audit findings; (iv) \emph{collaborative team science} for \textbf{collaborative} research, enabling multi-role decision governance, with shared team workspaces planned for future releases; and (v) \emph{real-world application scenarios} for \textbf{practical} impact, demonstrated through eight end-to-end user cases, with flagship demonstrations including multi-day agentic research sprints for gene discovery, AI-guided de novo protein design, molecular optimization, and genome-to-BGC discovery. The system combines a thin interaction layer, contextual research capability patterns, an Agent Runtime and Workflow Engine, an Evidence-DAG audit sidecar and a Scientific Model Router. SciForge currently runs as a desktop application, with mobile supervision support; future releases will deepen team collaboration. The system is open-source and available at https://github.com/AGI4Sci/SciForge

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

67 extracted references · 11 canonical work pages

  1. [1]

    arXiv preprint arXiv:2307.14367

    Abdine H, Chatzianastasis M, Bouyioukos C, et al (2023) Prot2Text: Multi- modal protein’s function generation with GNNs and transformers. arXiv preprint arXiv:2307.14367

    Pith/arXiv arXiv 2023

  2. [2]

    Nature 630:493–500

    Abramson J, Adler J, Dunger J, et al (2024) Accurate Structure Prediction of Biomolecular Interactions with AlphaFold 3. Nature 630:493–500. https://doi.or g/10.1038/s41586-024-07487-w

    doi:10.1038/s41586-024-07487-w 2024

  3. [3]

    Bioinformatics 31(13):2214–2216

    Alhossary A, Handoko SD, Mu Y, et al (2015) Fast, Accurate, and Reliable Molecular Docking with QuickVina 2. Bioinformatics 31(13):2214–2216. https: 36 //doi.org/10.1093/bioinformatics/btv082

    doi:10.1093/bioinformatics/btv082 2015

  4. [4]

    Available at https://allenai.org/bl og/ai2-scholarqa (accessed 4 July 2026)

    Allen Institute for AI (2025) Ai2 ScholarQA. Available at https://allenai.org/bl og/ai2-scholarqa (accessed 4 July 2026)

    2025

  5. [5]

    Available at https: //www.anthropic.com/news/model-context-protocol (accessed 4 July 2026)

    Anthropic (2024) Introducing the Model Context Protocol. Available at https: //www.anthropic.com/news/model-context-protocol (accessed 4 July 2026)

    2024

  6. [6]

    Available at https://www.anthropic.com/news/claude-science-ai-workbench (accessed 4 July 2026)

    Anthropic (2026) Claude Science: An AI Workbench for Scientists. Available at https://www.anthropic.com/news/claude-science-ai-workbench (accessed 4 July 2026)

    2026

  7. [7]

    Nature 650(8103):857–863

    Asai A, et al (2026) Synthesizing Scientific Literature with Retrieval-Augmented Language Models. Nature 650(8103):857–863. https://doi.org/10.1038/s41586-025 -10072-4

    work page doi:10.1038/s41586-025 2026

  8. [8]

    Journal of Medicinal Chemistry 53(7):2719–2740

    Baell JB, Holloway GA (2010) New Substructure Filters for Removal of Pan Assay Interference Compounds (PAINS) from Screening Libraries and for Their Exclusion in Bioassays. Journal of Medicinal Chemistry 53(7):2719–2740. https: //doi.org/10.1021/jm901137j

    doi:10.1021/jm901137j 2010

  9. [9]

    Nature Chemistry 4:90–98

    Bickerton GR, Paolini GV, Besnard J, et al (2012) Quantifying the Chemical Beauty of Drugs. Nature Chemistry 4:90–98. https://doi.org/10.1038/nchem.1243

    doi:10.1038/nchem.1243 2012

  10. [10]

    Nucleic Acids Research 51(W1):W46–W50

    Blin K, Shaw S, Augustijn HE, et al (2023) antiSMASH 7.0: New and Improved Predictions for Detection, Regulation, Chemical Structures and Visualisation. Nucleic Acids Research 51(W1):W46–W50. https://doi.org/10.1093/nar/gkad344

    doi:10.1093/nar/gkad344 2023

  11. [11]

    Nature 624:570–578

    Boiko DA, MacKnight R, Kline B, et al (2023) Autonomous Chemical Research with Large Language Models. Nature 624:570–578. https://doi.org/10.1038/s415 86-023-06792-0

    doi:10.1038/s415 2023

  12. [12]

    Nature Machine Intelligence 6(5):525–535

    Bran AM, Cox S, Schilter O, et al (2024) Augmenting Large Language Models with Chemistry Tools. Nature Machine Intelligence 6(5):525–535. https://doi.org/ 10.1038/s42256-024-00832-8

    doi:10.1038/s42256-024-00832-8 2024

  13. [13]

    Chen Z, Chen S, Ning Y, et al (2025) ScienceAgentBench: Toward Rigorous Assess- ment of Language Agents for Data-Driven Scientific Discovery. In: International Conference on Learning Representations, available at https://proceedings.iclr.cc/ paper files/paper/2025/hash/f12b4df26344f3be803c06b555252efe-Abstract-Con ference.html

    2025

  14. [14]

    Available at https://github.com/delibae/claude-prism (accessed 4 July 2026) 37

    ClaudePrism Contributors (2026) ClaudePrism: Desktop Workspace for Scientific Writing. Available at https://github.com/delibae/claude-prism (accessed 4 July 2026) 37

    2026

  15. [15]

    Available at https://help.consensus .app/en/articles/12641232-research-agent (accessed 4 July 2026)

    Consensus (2026) Consensus Research Agent. Available at https://help.consensus .app/en/articles/12641232-research-agent (accessed 4 July 2026)

    arXiv 2026

  16. [16]

    Science 378(6615):49–56

    Dauparas J, Anishchenko I, Bennett N, et al (2022) Robust Deep Learning- Based Protein Sequence Design Using ProteinMPNN. Science 378(6615):49–56. https://doi.org/10.1126/science.add2187

    doi:10.1126/science.add2187 2022

  17. [17]

    Available at https://elicit.com/ (accessed 4 July 2026)

    Elicit (2026) Elicit Research Agent. Available at https://elicit.com/ (accessed 4 July 2026)

    2026

  18. [18]

    Journal of Cheminformatics 1:8

    Ertl P, Schuffenhauer A (2009) Estimation of Synthetic Accessibility Score of Drug-Like Molecules Based on Molecular Complexity and Fragment Contributions. Journal of Cheminformatics 1:8. https://doi.org/10.1186/1758-2946-1-8

    doi:10.1186/1758-2946-1-8 2009

  19. [19]

    Available at https://www.futurehouse.org/news/launching-futurehouse-platform-ai-agents (accessed 4 July 2026)

    FutureHouse (2025) Launching FutureHouse Platform AI Agents. Available at https://www.futurehouse.org/news/launching-futurehouse-platform-ai-agents (accessed 4 July 2026)

    2025

  20. [20]

    Advanced Materials https://doi.org/10.1002/adma.202413523

    Ghafarollahi A, Buehler MJ (2025) SciAgents: Automating Scientific Discovery through Bioinspired Multi-Agent Intelligent Graph Reasoning. Advanced Materials https://doi.org/10.1002/adma.202413523

    doi:10.1002/adma.202413523 2025

  21. [22]

    Gillespie M, Jassal B, Stephan R, et al (2024) Reactome Pathway Knowledgebase

    2024

  22. [23]

    Nature 655(8122):487–496

    Gottweis J, Weng WH, Daryin A, et al (2026) Accelerating Scientific Discovery with Co-Scientist. Nature 655(8122):487–496. https://doi.org/10.1038/s41586-026 -10644-y

    work page doi:10.1038/s41586-026 2026

  23. [24]

    Science 387(6736):850–858

    Hayes T, Rao R, Akin H, et al (2025) Simulating 500 Million Years of Evolution with a Language Model. Science 387(6736):850–858. https://doi.org/10.1126/scie nce.ads0018

    doi:10.1126/scie 2025

  24. [25]

    Nature Methods 18(11):1342–1351

    Hu J, Li X, Coleman K, et al (2021) SpaGCN: Integrating Gene Expression, Spatial Location and Histology to Identify Spatial Domains and Spatially Variable Genes by Graph Convolutional Network. Nature Methods 18(11):1342–1351. https: //doi.org/10.1038/s41592-021-01255-8

    doi:10.1038/s41592-021-01255-8 2021

  25. [26]

    bioRxiv https://doi.org/10.1101/2025.05.30.656746 38

    Huang K, Zhang S, Wang H, et al (2025) Biomni: A General-Purpose Biomedical AI Agent. bioRxiv https://doi.org/10.1101/2025.05.30.656746 38

    doi:10.1101/2025.05.30.656746 2025

  26. [27]

    W3c member submission, World Wide Web Consortium, available at https://www.w3.org/submissions/201 3/SUBM-prov-json-20130424/

    Huynh TD, Jewell MO, Keshavarz AS, et al (2013) The PROV-JSON Serialization: A JSON Representation for the PROV Data Model. W3c member submission, World Wide Web Consortium, available at https://www.w3.org/submissions/201 3/SUBM-prov-json-20130424/

    2013

  27. [28]

    NEJM AI 2(1)

    Ifargan T, Hafner L, Kern M, et al (2025) Autonomous LLM-Driven Research from Data to Human-Verifiable Research Papers. NEJM AI 2(1). https://doi.org/ 10.1056/AIoa2400555

    doi:10.1056/aioa2400555 2025

  28. [29]

    In: Advances in Neural Information Processing Systems, pp 10088–10116, https: //doi.org/10.52202/079017-0324

    Jansen P, Cˆ ot´ e MA, Khot T, et al (2024) DiscoveryWorld: A Virtual Environ- ment for Developing and Evaluating Automated Scientific Discovery Agents. In: Advances in Neural Information Processing Systems, pp 10088–10116, https: //doi.org/10.52202/079017-0324

    doi:10.52202/079017-0324 2024

  29. [30]

    Nature 596(7873):583–589

    Jumper J, Evans R, Pritzel A, et al (2021) Highly Accurate Protein Structure Prediction with AlphaFold. Nature 596(7873):583–589. https://doi.org/10.1038/ s41586-021-03819-2

    2021

  30. [31]

    Nature Reviews Microbiology 17:167–180

    Keller NP (2019) Fungal Secondary Metabolism: Regulation, Function and Drug Discovery. Nature Reviews Microbiology 17:167–180. https://doi.org/10.1038/s4 1579-018-0121-1

    doi:10.1038/s4 2019

  31. [32]

    Preprint at https://arxiv.org/abs/ 2409.13740

    L´ ala J, O’Donoghue O, Shtedritski A, et al (2024) Language Agents Achieve Superhuman Synthesis of Scientific Knowledge. Preprint at https://arxiv.org/abs/ 2409.13740

    Pith/arXiv arXiv 2024

  32. [33]

    https://ww w.rdkit.org; software archive https://doi.org/10.5281/zenodo.591637

    Landrum G (2016) RDKit: Open-Source Cheminformatics Software. https://ww w.rdkit.org; software archive https://doi.org/10.5281/zenodo.591637

    doi:10.5281/zenodo.591637 2016

  33. [34]

    Science 379(6637):1123–1130

    Lin Z, Akin H, Rao R, et al (2023) Evolutionary-scale prediction of atomic-level protein structure with a language model. Science 379(6637):1123–1130. https: //doi.org/10.1126/science.ade2574

    doi:10.1126/science.ade2574 2023

  34. [35]

    Advanced Drug Delivery Reviews 46(1–3):3–26

    Lipinski CA, Lombardo F, Dominy BW, et al (2001) Experimental and Compu- tational Approaches to Estimate Solubility and Permeability in Drug Discovery and Development Settings. Advanced Drug Delivery Reviews 46(1–3):3–26. https://doi.org/10.1016/S0169-409X(00)00129-0

    doi:10.1016/s0169-409x(00)00129-0 2001

  35. [36]

    Nature Communications 14(1):1155

    Long Y, Ang KS, Li M, et al (2023) Spatially Informed Clustering, Integration, and Deconvolution of Spatial Transcriptomics with GraphST. Nature Communications 14(1):1155. https://doi.org/10.1038/s41467-023-36796-3

    doi:10.1038/s41467-023-36796-3 2023

  36. [37]

    Nature 651:914–919

    Lu C, Lu C, Lange RT, et al (2026) Towards End-to-End Automation of AI Research. Nature 651:914–919. https://doi.org/10.1038/s41586-026-10265-5

    doi:10.1038/s41586-026-10265-5 2026

  37. [38]

    W3C Recommendation, https://www.w3.org/TR/prov-dm/ 39

    Moreau L, Missier P (2013) PROV-DM: The PROV Data Model. W3C Recommendation, https://www.w3.org/TR/prov-dm/ 39

    2013

  38. [39]

    Nature Chemical Biology 16:60–68

    Navarro-Mu˜ noz JC, Selem-Mojica N, Mullowney MW, et al (2020) A Computa- tional Framework to Explore Large-Scale Biosynthetic Diversity. Nature Chemical Biology 16:60–68. https://doi.org/10.1038/s41589-019-0400-9

    work page doi:10.1038/s41589-019-0400-9 2020

  39. [40]

    Available at https://docs.omicos.cn/zh/part 1/01-overview.html (accessed 4 July 2026)

    OmicOS (2026) OmicOS Overview. Available at https://docs.omicos.cn/zh/part 1/01-overview.html (accessed 4 July 2026)

    2026

  40. [41]

    Available at https://docs.omicos.cn/zh/pa rt1/08-starting-cli.html (accessed 4 July 2026)

    OmicOS (2026) Starting OmicOS CLI. Available at https://docs.omicos.cn/zh/pa rt1/08-starting-cli.html (accessed 4 July 2026)

    2026

  41. [42]

    Available at https://docs.omicos.cn/zh/ part1/07-starting-serve.html (accessed 4 July 2026)

    OmicOS (2026) Starting OmicOS Serve. Available at https://docs.omicos.cn/zh/ part1/07-starting-serve.html (accessed 4 July 2026)

    2026

  42. [43]

    Available at https://github.com/TianGzlab/OmicsClaw (accessed 4 July 2026)

    OmicsClaw Contributors (2026) OmicsClaw: Local-First Multi-Omics Research Assistant. Available at https://github.com/TianGzlab/OmicsClaw (accessed 4 July 2026)

    2026

  43. [44]

    Nature Genetics 53:322–331

    Papalexi E, Mimitou EP, Butler A W, et al (2021) Characterizing the Molecular Regulation of Inhibitory Immune Checkpoints with Multimodal Single-Cell Screens. Nature Genetics 53:322–331. https://doi.org/10.1038/s41588-021-00778-2

    doi:10.1038/s41588-021-00778-2 2021

  44. [45]

    Biochemical Journal 448(3):417–423

    Park JH, Liu Y, Lemmon MA, et al (2012) Erlotinib Binds Both Inactive and Active Conformations of the EGFR Tyrosine Kinase Domain. Biochemical Journal 448(3):417–423. https://doi.org/10.1042/BJ20121513

    work page doi:10.1042/bj20121513 2012

  45. [46]

    bioRxiv https://doi.org/10.1101/2025.06.14 .659707

    Passaro S, Corso G, Wohlwend J, et al (2025) Boltz-2: Towards Accurate and Efficient Binding Affinity Prediction. bioRxiv https://doi.org/10.1101/2025.06.14 .659707

    work page doi:10.1101/2025.06.14 2025

  46. [47]

    arXiv preprint arXiv:2402.17810

    Pei Q, Wu L, Gao K, et al (2024) BioT5+: Towards generalized biological understanding with IUPAC integration and multi-task tuning. arXiv preprint arXiv:2402.17810

    Pith/arXiv arXiv 2024

  47. [48]

    Nature Biomedical Engineering 10:245–258

    Qu Y, Huang K, Yin M, et al (2026) CRISPR-GPT for Agentic Automation of Gene-Editing Experiments. Nature Biomedical Engineering 10:245–258. https: //doi.org/10.1038/s41551-025-01463-z

    doi:10.1038/s41551-025-01463-z 2026

  48. [49]

    bioRxiv https://doi.org/10.1101/2025.04.14.64885 0

    Rizvi SA, Levine D, Patel A, et al (2025) Scaling large language models for next- generation single-cell analysis. bioRxiv https://doi.org/10.1101/2025.04.14.64885 0

    work page doi:10.1101/2025.04.14.64885 2025

  49. [50]

    Science 357(6347):168–175

    Rocklin GJ, Chidyausiku TM, Goreshnik I, et al (2017) Global Analysis of Protein Folding Using Massively Parallel Design, Synthesis, and Testing. Science 357(6347):168–175. https://doi.org/10.1126/science.aan0693

    doi:10.1126/science.aan0693 2017

  50. [51]

    Journal of Chemi- cal Information and Modeling 50(5):742–754

    Rogers D, Hahn M (2010) Extended-Connectivity Fingerprints. Journal of Chemi- cal Information and Modeling 50(5):742–754. https://doi.org/10.1021/ci100050 t 40

    work page doi:10.1021/ci100050 2010

  51. [52]

    PLoS Computational Biology 9(10):e1003285

    Sandve GK, Nekrutenko A, Taylor J, et al (2013) Ten Simple Rules for Repro- ducible Computational Research. PLoS Computational Biology 9(10):e1003285. https://doi.org/10.1371/journal.pcbi.1003285

    doi:10.1371/journal.pcbi.1003285 2013

  52. [53]

    In: Findings of the Association for Computational Linguistics: EMNLP 2025

    Schmidgall S, Su Y, Wang Z, et al (2025) Agent Laboratory: Using LLM Agents as Research Assistants. In: Findings of the Association for Computational Linguistics: EMNLP 2025. Association for Computational Linguistics, pp 5977–6043, https: //doi.org/10.18653/v1/2025.findings-emnlp.320

    doi:10.18653/v1/2025.findings-emnlp.320 2025

  53. [54]

    Science 353(6294):78–82

    St ˚ ahl PL, Salm´ en F, Vickovic S, et al (2016) Visualization and Analysis of Gene Expression in Tissue Sections by Spatial Transcriptomics. Science 353(6294):78–82. https://doi.org/10.1126/science.aaf2403

    doi:10.1126/science.aaf2403 2016

  54. [55]

    Nature 646:716–723

    Swanson K, Wu W, Bulaong NL, et al (2025) The Virtual Lab of AI Agents Designs New SARS-CoV-2 Nanobodies. Nature 646:716–723. https://doi.org/10.1 038/s41586-025-09442-9

    2025

  55. [56]

    Available at https://github.com/swaruplab/operon (accessed 4 July 2026)

    Swarup Lab (2026) Operon: AI-Powered IDE for Bioinformatics. Available at https://github.com/swaruplab/operon (accessed 4 July 2026)

    2026

  56. [57]

    Nature 624:86–91

    Szymanski NJ, Rendy B, Fei Y, et al (2023) An Autonomous Laboratory for the Accelerated Synthesis of Inorganic Materials. Nature 624:86–91. https://doi.org/ 10.1038/s41586-023-06734-w

    doi:10.1038/s41586-023-06734-w 2023

  57. [58]

    Cell 170(3):564–576.e16

    Tsherniak A, Vazquez F, Montgomery PG, et al (2017) Defining a Cancer Depen- dency Map. Cell 170(3):564–576.e16. https://doi.org/10.1016/j.cell.2017.06.0 10

    work page doi:10.1016/j.cell.2017.06.0 2017

  58. [59]

    UniProt Consortium (2025) UniProt: The Universal Protein Knowledgebase in

    2025

  59. [60]

    Nature 620:1089–1100

    Watson JL, Juergens D, Bennett NR, et al (2023) De Novo Design of Protein Structure and Function with RFdiffusion. Nature 620:1089–1100. https://doi.org/ 10.1038/s41586-023-06415-8

    doi:10.1038/s41586-023-06415-8 2023

  60. [61]

    Scientific Data 3:160018

    Wilkinson MD, Dumontier M, Aalbersberg IJ, et al (2016) The F AIR Guiding Prin- ciples for Scientific Data Management and Stewardship. Scientific Data 3:160018. https://doi.org/10.1038/sdata.2016.18

    doi:10.1038/sdata.2016.18 2016

  61. [62]

    Genome Medicine 16(1):12

    Xu H, Fu H, Long Y, et al (2024) Unsupervised Spatially Embedded Deep Representation of Spatial Transcriptomics. Genome Medicine 16(1):12. https: //doi.org/10.1186/s13073-024-01283-x

    work page doi:10.1186/s13073-024-01283-x 2024

  62. [63]

    Preprint at https://arxi v.org/abs/2504.08066 41

    Yamada Y, Lange RT, Lu C, et al (2025) The AI Scientist-v2: Workshop-Level Automated Scientific Discovery via Agentic Tree Search. Preprint at https://arxi v.org/abs/2504.08066 41

    Pith/arXiv arXiv 2025

  63. [64]

    arXiv preprint ArXiv:2604.27646v1

    Yang K, et al (2026) Benchmarking virtual cell models for in-the-wild perturbation response. arXiv preprint ArXiv:2604.27646v1

    Pith/arXiv arXiv 2026

  64. [65]

    Nucleic Acids Research 53(D1)

    Zdouc MM, Blin K, Louwen NLL, et al (2025) MIBiG 4.0: Advancing Biosynthetic Gene Cluster Curation through Global Collaboration. Nucleic Acids Research 53(D1). https://doi.org/10.1093/nar/gkae1115

    doi:10.1093/nar/gkae1115 2025

  65. [66]

    Bioinformatics 42(7)

    Zhang Z, Duan H, Gao X (2026) MCFST: Spatial domain identification method based on multi-view graph convolutional network and graph fusion network. Bioinformatics 42(7). https://doi.org/10.1093/bioinformatics/btag469

    work page doi:10.1093/bioinformatics/btag469 2026

  66. [67]

    Nature Biotechnology 39(11):1375–1384

    Zhao E, Stone MR, Ren X, et al (2021) Spatial Transcriptomics at Subspot Resolution with BayesSpace. Nature Biotechnology 39(11):1375–1384. https: //doi.org/10.1038/s41587-021-00935-2 42

    work page doi:10.1038/s41587-021-00935-2 2021

  67. [2025]

    https://doi.org/10.1093/nar/gk ae1010

    Nucleic Acids Research 53(D1):D609–D617. https://doi.org/10.1093/nar/gk ae1010

    doi:10.1093/nar/gk

This paper was first reviewed by deepseek-v4-flash on August 1, 2026.