pith. sign in

arxiv: 2605.18144 · v1 · pith:KVTSRXL4new · submitted 2026-05-18 · 💻 cs.AI

Evidence-Grounded Frontier Mapping and Agentic Hypothesis Generation in Nanomedicine

Christiaan G.A. Viviers , Koen de Bruin , Mirre M. Trines , Ayla M. Hokke , Roy van der Meel , Avi Schroeder , Twan Lammers , Willem J.M. Mulder
show 1 more author
Fons van der Sommen
This is my paper

Pith reviewed 2026-05-20 10:34 UTC · model grok-4.3

classification 💻 cs.AI
keywords nanomedicineliterature mappinghypothesis generationlarge language modelsagentic workflowsresearch frontierscitation analysisdiscovery support
0
0 comments X

The pith

pArticleMap maps low-density bridge regions in nanomedicine literature and generates citation-grounded hypotheses via agentic language models.

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

The paper introduces pArticleMap to navigate the fragmented nanomedicine literature spanning delivery chemistry, immunology, imaging, and translational science. It identifies low-density article-level bridge regions and cluster interfaces using embeddings and similarity graphs, then applies structured evidence retrieval and an audited agentic LLM workflow to generate and score research hypotheses. Retrospective evaluation on four bundles yields a pooled gold recovery rate of 10.8 percent, recall at ten of 15.9 percent, and a future-neighborhood rate of 61.0 percent, showing the system frequently reaches the right forward-looking area even without exact paper matches. Human agreement with internal scores is modest, positioning the tool as an evidence-based support rather than a replacement for expert judgment.

Core claim

Rather than forecasting future concept co-occurrence, pArticleMap targets low-density article-level bridge regions and cluster interfaces, then generates and scores citation-grounded hypotheses with large language models in an agentic setup, obtaining a pooled gold recovery rate of 10.8%, recall@10 of 15.9%, and future-neighborhood rate of 61.0% across retrospective bundles.

What carries the argument

The pArticleMap pipeline of article embeddings, similarity-graph frontier extraction, evidence-pack retrieval, and audited agentic LLM hypothesis generation and scoring.

If this is right

  • The system reaches the correct future research neighborhood in 61 percent of cases even when exact paper recovery is lower.
  • Internal scoring functions as a useful but imperfect support signal that does not replace human expert assessment.
  • Task-retained hypotheses can be produced under the benchmark protocol across multiple nanomedicine sub-areas.
  • The approach emphasizes conservative, citation-grounded ideation over direct prediction of concept co-occurrences.

Where Pith is reading between the lines

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

  • If the retrospective benchmark generalizes, the same bridge-region mapping could surface overlooked connections between subfields such as biomaterials and disease-specific applications.
  • Researchers in other domains with large heterogeneous literatures might adapt the frontier-extraction step to prioritize interface areas for new experiments.
  • Adding loops that incorporate fresh experimental outcomes back into the evidence packs could tighten the scoring over successive cycles.
  • The method could help allocate limited lab resources toward interfaces that already show partial evidence in neighboring clusters.

Load-bearing premise

That the retrospective realization benchmark of generating later literature under a historical cutoff serves as a valid proxy for producing useful forward-looking hypotheses in real research settings.

What would settle it

A prospective trial in which pArticleMap generates hypotheses from current literature cutoffs, experts pursue a subset in the lab, and the rate at which those specific ideas lead to new publications or validations is compared against control sets of expert-only or random ideas.

Figures

Figures reproduced from arXiv: 2605.18144 by Avi Schroeder, Ayla M. Hokke, Christiaan G.A. Viviers, Fons van der Sommen, Koen de Bruin, Mirre M. Trines, Roy van der Meel, Twan Lammers, Willem J.M. Mulder.

Figure 1
Figure 1. Figure 1: Similarity-weighted semantic overlays reveal how major nanotechnology themes occupy distinct but partially overlapping regions of the full-corpus embedding space. fragmented across disease areas, delivery routes, material classes, and experimental models. As a result, potentially important conceptual connections often remain buried in separate subfields even when the underlying evidence already exists in p… view at source ↗
Figure 2
Figure 2. Figure 2: pArticleMap maps article-level literature structure, surfaces sparse frontier regions, and uses evidence-constrained LLM workflows to generate and evaluate grounded research directions. that links corpus construction, representation learning, target identi￾fication, evidence retrieval, structured reasoning, and evaluation. In parallel, AI in nanomedicine has so far been used primarily for tasks such as nan… view at source ↗
Figure 3
Figure 3. Figure 3: Implemented agent workflow. After audit, the pipeline may optionally invoke patch retrieve before continuing to ideate, score, blueprint, and publish. A detailed schematic of the agent orchestration is available in Figure B. During ‘explain‘, the language model receives the serialized ev￾idence pack and produces a structured contrastive account of the target. For gap targets it is asked to explain what lie… view at source ↗
Figure 6
Figure 6. Figure 6: Winner-minus-nonwinner metric lift for pArticleMap. Review-packet winner selection materially improves exact realization metrics, but does not reduce historical confounding. review-packet winner increased gold recovery from 3.6% to 10.8%, recall@10 from 5.4% to 15.9%, and MRR from 0.034 to 0.083 [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 5
Figure 5. Figure 5: Winner-level realization metrics by domain for pArticleMap. Biosensing and payload integration act as positive cases, whereas antimicrobials and vaccines expose confounding and exact-disambiguation difficulty. 5.2. Domain heterogeneity across cue-conditioned tasks Realization behavior varied sharply by domain, as shown in Fig￾ures 4 and 5. Biosensing gave the strongest exact-recovery signal, with 27.1% gol… view at source ↗
Figure 7
Figure 7. Figure 7: Bar graph with pooled human-versus-agent calibration summary for the pArticleMap ideas. Error bars represent disagreement on the average score across the reviewers. 5.5. Interpretation of pArticleMap Taken together, these results position pArticleMap as a conservative research-assistance system rather than an autonomous discovery en￾gine. The strongest empirical claim is not that the system predicts the fu… view at source ↗
Figure 9
Figure 9. Figure 9: Semantic-thresholded subcorpus and operational communities for the vaccines example. Top: papers retained after semantic filtering and thresholding. Bottom: Leiden communities computed on the similarity graph, providing the operational clustering propagated into the stored snapshot. A.2. Similarity graph to operational literature communities Once the retained slice has been converted into a graph, pArticle… view at source ↗
Figure 8
Figure 8. Figure 8: Corpus overview and semantic slicing. Left: global UMAP view of the full nanomedicine corpus, used as an exploratory map rather than as an analysis space. Right: overlay of semantic filters showing how domain-specific queries occupy coherent but partially overlapping regions of the broader literature landscape. After an analyst defines one semantic direction through text￾conditioned embedding similarity, p… view at source ↗
Figure 10
Figure 10. Figure 10: Paper-level gap scores over the same slice, highlighting sparse frontier papers at the interface of denser literature regions. touch define the cluster-pair bridge targets later used by the retrieval and agentic-generation stack. Taken together, Figures 8–11 show the full corpus-creation and analysis flow used by pArticleMap. The process begins with a broad PubMed-derived literature collection, narrows to… view at source ↗
Figure 11
Figure 11. Figure 11: Gap-region extraction and target surfacing. Left: retained connected components among the high-gap-score papers after thresholding and minimum-size filtering. Right: the same gap regions overlaid on Leiden communities, showing how sparse regions induce both direct gap targets and cluster-pair bridge targets for downstream evidence-pack construction [PITH_FULL_IMAGE:figures/full_fig_p015_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Implemented LangGraph orchestration path for pArticleMap. The system builds a target-conditioned evidence pack, explains the frontier, audits the explanation, optionally patches retrieval, generates multiple grounded hypotheses, scores them, and emits a preclinical blueprint only for the top-scored idea. pers and combines cue-hit papers, cluster exemplars, and boundary papers. The top retrieved items incl… view at source ↗
Figure 13
Figure 13. Figure 13: Qwen3 Embedding model architecture = 4.0, plausibility = 4.0, feasibility = 3.5, evaluability = 4.0, and likely impact = 3.5, giving an overall human mean of 3.75∕5. Audit note. The audit reported a supported-claim fraction of approx￾imately 0.92, but still set needs_patch = True and the explain stage recorded insufficient_evidence = True. It explicitly flagged miss￾ing direct evidence for mRNA-LNP vaccin… view at source ↗
read the original abstract

Nanomedicine research spans delivery chemistry, immunology, imaging, biomaterials, and disease-specific translational science, yet its conceptual design space remains fragmented across a large and heterogeneous literature. To date, artificial intelligence in nanomedicine has focused primarily on property prediction and formulation optimization, with much less attention to evidence-grounded discovery support at the level of research direction selection. We introduce pArticleMap, a literature-mapping and research-hypothesis-generation system that combines article embeddings, similarity-graph analysis, sparse frontier extraction, structured evidence-pack retrieval, and an audited large-language-model (LLM) workflow for grounded ideation. Rather than forecasting future concept co-occurrence, pArticleMap targets low-density article-level bridge regions and cluster interfaces, then generates and scores citation-grounded hypotheses with large language models in an agentic setup. We evaluate the system with a retrospective realization benchmark (generate later literature under a historical cutoff) and a blinded human reader assessment layer across cue-conditioned nanomedicine tasks. Across 4 selected retrospective bundles, pArticleMap generated ideas and selected task-retained hypotheses (winner ideas) under the benchmark protocol. For task-level retained hypotheses, a pooled gold recovery rate of 10.8% was obtained, with a recall@10 of 15.9% and a future-neighborhood rate of 61.0%, indicating that the system often reached the correct forward-looking neighborhood (paper ideas) even without exact paper-level recovery. Human-agent agreement is modest overall, indicating that internal scoring is useful as a support signal but does not replace expert judgment. These results position pArticleMap as a conservative, evidence-grounded research assistant for nanomedicine.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper introduces pArticleMap, a literature-mapping and hypothesis-generation system for nanomedicine that uses article embeddings, similarity-graph analysis, sparse frontier extraction to target low-density bridge regions and cluster interfaces, followed by structured evidence-pack retrieval and an audited agentic LLM workflow to generate and score citation-grounded hypotheses. Evaluation uses a retrospective realization benchmark (generating hypotheses from pre-cutoff literature and scoring against later papers) across 4 bundles, yielding a pooled gold recovery rate of 10.8%, recall@10 of 15.9%, and future-neighborhood rate of 61.0%, plus a blinded human reader assessment layer showing modest human-agent agreement.

Significance. If the retrospective metrics hold as a proxy for forward utility, the work offers a conservative, evidence-grounded assistant for research direction selection in a fragmented domain, extending AI in nanomedicine beyond property prediction. The concrete retrospective metrics and emphasis on citation-grounded, audited LLM use are strengths that position the system as a support tool rather than an autonomous discoverer.

major comments (2)
  1. [Evaluation section / retrospective benchmark protocol] The retrospective realization benchmark (described in the evaluation section and abstract) is load-bearing for the central claim that pArticleMap produces useful forward-looking hypotheses. Alignment with later published papers (gold recovery 10.8%, future-neighborhood 61.0%) assumes published outcomes reliably signal intrinsic promise, yet published literature is filtered by funding, feasibility, and trends; this risks the system rediscovering latent embedding-space patterns rather than generating novel bridges. A concrete test against expert-proposed directions or controlled forward experiments is needed to validate the proxy.
  2. [Human reader assessment layer] Modest human-agent agreement (noted in the abstract and human assessment layer) indicates that the internal scoring driving the benchmark numbers diverges from expert judgment. This weakens the claim that the system provides reliable support signals; further breakdown of disagreement cases by task type or bundle would clarify whether the 61.0% neighborhood rate reflects genuine utility or benchmark artifacts.
minor comments (2)
  1. [Abstract and evaluation setup] Specify the exact selection criteria and characteristics of the 4 retrospective bundles to allow reproducibility and assessment of generalizability.
  2. [Sparse frontier extraction description] Clarify quantitative thresholds used for low-density bridge regions and cluster interfaces in the similarity-graph analysis.

Simulated Author's Rebuttal

2 responses · 1 unresolved

We thank the referee for their constructive comments. We address the major concerns regarding the retrospective benchmark and the human assessment layer below, making revisions to improve clarity and transparency where possible.

read point-by-point responses

  1. Referee: [Evaluation section / retrospective benchmark protocol] The retrospective realization benchmark (described in the evaluation section and abstract) is load-bearing for the central claim that pArticleMap produces useful forward-looking hypotheses. Alignment with later published papers (gold recovery 10.8%, future-neighborhood 61.0%) assumes published outcomes reliably signal intrinsic promise, yet published literature is filtered by funding, feasibility, and trends; this risks the system rediscovering latent embedding-space patterns rather than generating novel bridges. A concrete test against expert-proposed directions or controlled forward experiments is needed to validate the proxy.

    Authors: We agree that the retrospective benchmark serves as a proxy and is subject to the biases inherent in published literature, including influences from funding, feasibility, and prevailing trends. This limitation is common to literature-based discovery approaches and does not claim to identify 'intrinsic promise' independent of publication filters. Instead, it measures the system's ability to surface directions that were subsequently realized in the literature. To strengthen the manuscript, we have revised the evaluation section to explicitly discuss these proxy limitations and the potential for rediscovering embedding patterns. We have also added text noting that direct validation against expert-proposed directions or controlled forward experiments would require prospective studies, which we identify as an important avenue for future work. revision: partial

  2. Referee: [Human reader assessment layer] Modest human-agent agreement (noted in the abstract and human assessment layer) indicates that the internal scoring driving the benchmark numbers diverges from expert judgment. This weakens the claim that the system provides reliable support signals; further breakdown of disagreement cases by task type or bundle would clarify whether the 61.0% neighborhood rate reflects genuine utility or benchmark artifacts.

    Authors: We concur that the modest agreement highlights the complementary nature of the system's scoring to expert judgment. In response, we have expanded the human assessment section with a breakdown of disagreement cases, stratified by task type and bundle. This additional analysis reveals that disagreements are more prevalent in bundles involving highly interdisciplinary topics, where expert opinions themselves may vary. We believe this supports the interpretation of the 61.0% future-neighborhood rate as indicating utility in identifying relevant neighborhoods, while acknowledging that the system is intended as a support tool rather than a definitive judge. revision: yes

standing simulated objections not resolved
  • A full prospective validation with controlled forward experiments or direct expert-proposed direction comparisons cannot be addressed within the scope of this revision, as it would necessitate new experimental designs and data collection beyond the current retrospective and human assessment framework.

Circularity Check

0 steps flagged

No significant circularity; retrospective benchmark is external validation

full rationale

The paper's claimed chain proceeds from article embeddings and similarity-graph analysis to sparse frontier extraction, evidence-pack retrieval, and LLM-based hypothesis generation, then evaluates via retrospective realization against actual later literature. This benchmark compares system outputs to independently published future papers under a historical cutoff and is not equivalent to the inputs by construction, nor does it rely on fitted parameters renamed as predictions or load-bearing self-citations. No equations or steps reduce the output to the input definitions; the evaluation uses external gold-standard literature as an independent check. The derivation remains self-contained against this external benchmark.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Based on the abstract alone, the work relies on standard NLP and LLM techniques without introducing new mathematical axioms or invented physical entities; the main addition is the integrated workflow and evaluation protocol.

axioms (1)
  • domain assumption Article embeddings and similarity graphs can reliably identify low-density bridge regions that correspond to promising research frontiers.
    This assumption underpins the sparse frontier extraction step described in the abstract.

pith-pipeline@v0.9.0 · 5874 in / 1303 out tokens · 66092 ms · 2026-05-20T10:34:45.062739+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

41 extracted references · 41 canonical work pages · 1 internal anchor

  1. [1]

    Fishoil,raynaud’ssyndrome,andundiscov- eredpublicknowledge

    D.R.Swanson,“Fishoil,raynaud’ssyndrome,andundiscov- eredpublicknowledge”,PerspectivesinBiologyandMedicine, vol.30,no.1,pp.7–18,1986.doi:10.1353/pbm.1986.0087

    work page doi:10.1353/pbm.1986.0087 1986

  2. [2]

    Using arrowsmith: A computer-assisted approach to formulating and assessing scientific hypotheses

    N. R. Smalheiser and D. R. Swanson, “Using arrowsmith: A computer-assisted approach to formulating and assessing scientific hypotheses”,Computer Methods and Programs in Biomedicine,vol.57,no.3,pp.149–153,1998,issn:0169-2607. doi:https://doi.org/10.1016/S0169-2607(98)00033-9[Online]. Available:https://www.sciencedirect.com/science/article/pii/ S0169260798000339

    work page doi:10.1016/s0169-2607(98)00033-9 1998

  3. [3]

    Usingliterature-baseddiscoverytoidentifydiseasecandidate genes

    D.Hristovski,B.Peterlin,J.A.Mitchell,andS.M.Humphrey, “Usingliterature-baseddiscoverytoidentifydiseasecandidate genes”,International Journal of Medical Informatics, vol. 74, no.2,pp.289–298,2005,MIE2003,issn:1386-5056.doi:https: //doi.org/10.1016/j.ijmedinf.2004.04.024 [Online]. Avail- able: https://www.sciencedirect.com/science/article/pii/ S1386505604001650

    work page doi:10.1016/j.ijmedinf.2004.04.024 2005

  4. [4]

    Citespace ii: Detecting and visualizing emerging trendsandtransientpatternsinscientificliterature

    C. Chen, “Citespace ii: Detecting and visualizing emerging trendsandtransientpatternsinscientificliterature”,Journalof theAmericanSocietyforInformationScienceandTechnology, vol.57,no.3,pp.359–377,2006.doi:https://doi.org/10.1002/ asi.20317eprint:https://onlinelibrary.wiley.com/doi/pdf/10. 1002/asi.20317.[Online].Available:https://onlinelibrary.wiley. com/...

    work page doi:10.1002/asi.20317 2006

  5. [5]

    Leastsquaresquantizationinpcm

    S.Lloyd,“Leastsquaresquantizationinpcm”,IEEETrans.Inf. Theor.,vol.28,no.2,pp.129–137,Sep.2006,issn:0018-9448. doi: 10.1109/TIT.1982.1056489 [Online]. Available: https: //doi.org/10.1109/TIT.1982.1056489

    work page doi:10.1109/tit.1982.1056489 2006

  6. [6]

    Softwaresurvey:Vosviewer,a computerprogramforbibliometricmapping

    N.J.vanEckandL.Waltman,“Softwaresurvey:Vosviewer,a computerprogramforbibliometricmapping”,Scientometrics, vol.84,no.2,pp.523–538,2010.doi:10.1007/s11192-009-0146- 3

    work page doi:10.1007/s11192-009-0146- 2010

  7. [7]

    Textminingandvisualization usingvosviewer

    N.J.vanEckandL.Waltman,“Textminingandvisualization usingvosviewer”,ISSINewsletter,vol.7,no.3,pp.50–54,2011

    work page 2011

  8. [8]

    Context-drivenautomatic subgraphcreationforliterature-baseddiscovery

    D. Cameron, R. Kavuluru, T. C. Rindflesch, A. P. Sheth, K. Thirunarayan,andO.Bodenreider,“Context-drivenautomatic subgraphcreationforliterature-baseddiscovery”,en,J.Biomed. Inform.,vol.54,pp.141–157,Apr.2015.doi:10.1016/j.jbi.2015. 01.014

    work page doi:10.1016/j.jbi.2015 2015

  9. [9]

    Moliere: Automatic biomedical hypothesis generation system

    J. Sybrandt, M. Shtutman, and I. Safro, “Moliere: Automatic biomedical hypothesis generation system”, inProceedings of the23rdACMSIGKDDInternationalConferenceonKnowledge DiscoveryandDataMining,ser.KDD’17,Halifax,NS,Canada: Association for Computing Machinery, 2017, pp. 1633–1642, isbn:9781450348874.doi:10.1145/3097983.3098057[Online]. Available:https://do...

    work page doi:10.1145/3097983.3098057 2017

  10. [10]

    Smart cancer nanomedicine

    R.vanderMeel,E.Sulheim,Y.Shi,F.Kiessling,W.J.M.Mul- der, and T. Lammers, “Smart cancer nanomedicine”,Nature Nanotechnology,vol.14,no.11,pp.1007–1017,Nov.2019,issn: 1748-3395.doi:10.1038/s41565-019-0567-y[Online].Available: https://doi.org/10.1038/s41565-019-0567-y

    work page doi:10.1038/s41565-019-0567-y 2019

  11. [11]

    Fromlouvainto leiden:Guaranteeingwell-connectedcommunities

    V.A.Traag,L.Waltman,andN.J.vanEck,“Fromlouvainto leiden:Guaranteeingwell-connectedcommunities”,Scientific Reports,vol.9,no.1,Mar.2019,issn:2045-2322.doi:10.1038/ s41598-019-41695-z[Online].Available:http://dx.doi.org/10. 1038/s41598-019-41695-z

    work page 2019

  12. [12]

    Retrieval-augmentedgenerationforknowledge- intensive nlp tasks

    P.Lewisetal.,“Retrieval-augmentedgenerationforknowledge- intensive nlp tasks”, inAdvances in Neural Information Processing Systems, H. Larochelle, M. Ranzato, R. Hadsell, M. Balcan, and H. Lin, Eds., vol. 33, Curran Associates, Inc., 2020, pp. 9459–9474. [Online]. Available: https : / / proceedings.neurips.cc/paper_files/paper/2020/file/ 6b493230205f780e1...

    work page 2020

  13. [13]

    A hybrid approach to hierarchical density-based cluster selection

    C. Malzer and M. Baum, “A hybrid approach to hierarchical density-based cluster selection”, in2020 IEEE International ConferenceonMultisensorFusionandIntegrationforIntelligent Systems (MFI), IEEE, Sep. 2020, pp. 223–228.doi: 10.1109/ mfi49285.2020.9235263[Online].Available:http://dx.doi.org/ 10.1109/MFI49285.2020.9235263

    work page doi:10.1109/mfi49285.2020.9235263 2020

  14. [14]

    Agatha: Automatic graph mining and transformer based hypothesis generationapproach

    J. Sybrandt, I. Tyagin, M. Shtutman, and I. Safro, “Agatha: Automatic graph mining and transformer based hypothesis generationapproach”,inProceedingsofthe29thACMInterna- tional Conference on Information & Knowledge Management, ser.CIKM’20,VirtualEvent,Ireland:AssociationforComput- ingMachinery,2020,pp.2757–2764,isbn:9781450368599.doi: 10.1145/3340531.3412...

    work page doi:10.1145/3340531.3412684 2020

  15. [15]

    Development of pharmaceutical nanomedicines: From the bench to the market

    A. A. Halwani, “Development of pharmaceutical nanomedicines: From the bench to the market”,Phar- maceutics, vol. 14, no. 1, 2022,issn: 1999-4923.doi: 10 . 3390 / pharmaceutics14010106 [Online]. Available: https://www.mdpi.com/1999-4923/14/1/106

    work page 2022

  16. [16]

    Artificial intelligence to bring nanomedicine to life

    N. Serov and V. Vinogradov, “Artificial intelligence to bring nanomedicine to life”, en,Adv. Drug Deliv. Rev., vol. 184, no.114194,p.114194,May2022

    work page

  17. [17]

    Forecastingthefutureofartificialintelligence with machine learning-based link prediction in an exponen- tially growing knowledge network

    M.Krennetal.,“Forecastingthefutureofartificialintelligence with machine learning-based link prediction in an exponen- tially growing knowledge network”,Nature Machine Intelli- gence,vol.5,no.11,pp.1326–1335,2023,issn:2522-5839.doi: 10.1038/s42256-023-00735-0[Online].Available:https://doi. org/10.1038/s42256-023-00735-0

    work page doi:10.1038/s42256-023-00735-0 2023

  18. [18]

    arXiv preprint arXiv:2312.07559 , year=

    J.Lála,O.O’Donoghue,A.Shtedritski,S.Cox,S.G.Rodriques, and A. D. White, “Paperqa: Retrieval-augmented generative agentforscientificresearch”,arXivpreprintarXiv:2312.07559, 2023.doi: 10.48550/arXiv.2312.07559 [Online]. Available: https://arxiv.org/abs/2312.07559

    work page doi:10.48550/arxiv.2312.07559 2023

  19. [19]

    Mechanisms and barriers in nanomedicine: Progress in the field and future directions

    T. Anchordoquy et al., “Mechanisms and barriers in nanomedicine: Progress in the field and future directions”, en,ACS Nano, vol. 18, no. 22, pp. 13983–13999, 2024.doi: 10.1021/acsnano.4c00182

    work page doi:10.1021/acsnano.4c00182 2024

  20. [20]

    anthropic.com/news/model-context-protocol,Accessed:2026- 04-12,2024

    Anthropic,Introducingthemodelcontextprotocol,https://www. anthropic.com/news/model-context-protocol,Accessed:2026- 04-12,2024

    work page 2026

  21. [21]

    AcceleratingionizablelipiddiscoveryformRNA deliveryusingmachinelearningandcombinatorialchemistry

    B.Lietal.,“AcceleratingionizablelipiddiscoveryformRNA deliveryusingmachinelearningandcombinatorialchemistry”, NatureMaterials, vol. 23, no. 7, pp. 1002–1008, 2024.doi: 10. 1038/s41563-024-01867-3

    work page 2024

  22. [22]

    Bran, A.; Cox, S.; Schilter, O.; Baldassari, C.; White, A

    A.M.Bran,S.Cox,O.Schilter,C.Baldassari,A.D.White,andP. Schwaller,“Augmentinglargelanguagemodelswithchemistry tools”,NatureMachineIntelligence,vol.6,no.5,pp.525–535, 2024,issn: 2522-5839.doi: 10.1038/s42256-024-00832-8 [Online].Available:https://doi.org/10.1038/s42256-024-00832- 8

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

  23. [23]

    Machinelearning-guidedhighthroughput nanoparticledesign

    A. Ortiz-Perez, D. van Tilborg, R. van der Meel, F. Grisoni, andL.Albertazzi,“Machinelearning-guidedhighthroughput nanoparticledesign”,DigitalDiscovery,vol.3,no.7,pp.1280– 1291,2024.doi:10.1039/D4DD00104D 11–21 Evidence-Grounded Frontier Mapping and Agentic Hypothesis Generation in Nanomedicine Arxiv, May 2026

    work page doi:10.1039/d4dd00104d 2024

  24. [24]

    Skarlinski, Sam Cox, Jon M

    M. D. Skarlinski et al., “Language agents achieve super- human synthesis of scientific knowledge”,arXiv preprint arXiv:2409.13740,2024.doi:10.48550/arXiv.2409.13740[On- line].Available:https://arxiv.org/abs/2409.13740

    work page doi:10.48550/arxiv.2409.13740 2024

  25. [25]

    Wang, Iskandar Sitdikov, Ciro Salcedo, Alireza Seif, and Zlatko K

    J. Yang et al., “Poisoning medical knowledge using large lan- guage models”,Nature Machine Intelligence, vol. 6, no. 10, pp.1156–1168,2024,issn:2522-5839.doi:10.1038/s42256-024- 00899-3[Online].Available:https://doi.org/10.1038/s42256- 024-00899-3

    work page doi:10.1038/s42256-024- 2024

  26. [26]

    Asurveyonhypothesisgenerationforsci- entific discovery in the era of large language models

    A.K.Alkanetal.,“Asurveyonhypothesisgenerationforsci- entific discovery in the era of large language models”,arXiv preprint arXiv:2504.05496, 2025.doi: 10.48550/arXiv.2504. 05496[Online].Available:https://arxiv.org/abs/2504.05496

    work page doi:10.48550/arxiv.2504 2025

  27. [27]

    Scientifichypothesisgenerationandvalida- tion:Methods,datasets,andfuturedirections

    A.Kulkarnietal.,“Scientifichypothesisgenerationandvalida- tion:Methods,datasets,andfuturedirections”,arXivpreprint arXiv:2505.04651,2025.doi:10.48550/arXiv.2505.04651[On- line].Available:https://arxiv.org/abs/2505.04651

    work page doi:10.48550/arxiv.2505.04651 2025

  28. [28]

    ModelContextProtocolContributors,Modelcontextprotocol specification, https://modelcontextprotocol.io/specification/ 2025-11-25,Accessed:2026-04-12,2025

    work page 2025

  29. [29]

    Lucchesi

    S.Ren,P.Jian,Z.Ren,C.Leng,C.Xie,andJ.Zhang,“Towards scientificintelligence:Asurveyofllm-basedscientificagents”, arXiv preprint arXiv:2503.24047, 2025.doi: 10.48550/arXiv. 2503.24047 [Online]. Available: https://arxiv.org/abs/2503. 24047

    work page internal anchor Pith review doi:10.48550/arxiv 2025

  30. [30]

    10896 [cs.CL].[Online].Available:https://arxiv.org/abs/2506

    T.Sounacketal.,Bioclinicalmodernbert:Astate-of-the-artlong- contextencoderforbiomedicalandclinicalnlp,2025.arXiv:2506. 10896 [cs.CL].[Online].Available:https://arxiv.org/abs/2506. 10896

    work page 2025

  31. [31]

    A neural symbolic model for space physics

    J. Ying et al., “A neural symbolic model for space physics”, NatureMachineIntelligence,vol.7,no.10,pp.1726–1741,2025, issn: 2522-5839.doi: 10.1038/s42256-025-01126-3 [Online]. Available:https://doi.org/10.1038/s42256-025-01126-3

    work page doi:10.1038/s42256-025-01126-3 2025

  32. [32]

    arXiv: 2506

    Y.Zhangetal.,Qwen3embedding:Advancingtextembedding and reranking through foundation models, 2025. arXiv: 2506. 05176 [cs.CL].[Online].Available:https://arxiv.org/abs/2506. 05176

    work page 2025

  33. [33]

    Acomprehensivelarge-scalebiomedicalknowl- edge graph for ai-powered data-driven biomedical research

    Y.Zhangetal.,“Acomprehensivelarge-scalebiomedicalknowl- edge graph for ai-powered data-driven biomedical research”, Nature Machine Intelligence, vol. 7, no. 4, pp. 602–614, 2025, issn:2522-5839.doi:10.1038/s42256-025-01014-w [Online]. Available:https://doi.org/10.1038/s42256-025-01014-w

    work page doi:10.1038/s42256-025-01014-w 2025

  34. [34]

    Fromautomationtoautonomy:Asurveyon largelanguagemodelsinscientificdiscovery

    T.Zhengetal.,“Fromautomationtoautonomy:Asurveyon largelanguagemodelsinscientificdiscovery”,inProceedingsof the2025ConferenceonEmpiricalMethodsinNaturalLanguage Processing, Association for Computational Linguistics, 2025, pp. 17733–17750. [Online]. Available: https://aclanthology. org/2025.emnlp-main.895/

    work page 2025

  35. [35]

    Largelanguagemodelsforscientificdiscovery inmolecularpropertyprediction

    Y.Zhengetal.,“Largelanguagemodelsforscientificdiscovery inmolecularpropertyprediction”,NatureMachineIntelligence, vol.7,no.3,pp.437–447,2025,issn:2522-5839.doi:10.1038/ s42256-025-00994-z [Online]. Available: https://doi.org/10. 1038/s42256-025-00994-z

    work page 2025

  36. [36]

    Nature (2026) https://doi.org/10.1038/s41586-025-10072-4

    A.Asaietal.,“Synthesizingscientificliteraturewithretrieval- augmented language models”,Nature, vol. 650, pp. 857–863, 2026.doi: 10.1038/s41586-025-10072-4 [Online]. Available: https://doi.org/10.1038/s41586-025-10072-4

    work page doi:10.1038/s41586-025-10072-4 2026

  37. [37]

    High- throughput platforms for machine learning-guided lipid nanoparticledesign

    A. R. Hanna, D. A. Issadore, and M. J. Mitchell, “High- throughput platforms for machine learning-guided lipid nanoparticledesign”,NatureReviewsMaterials,vol.11,no.1, pp.50–64,Jan.2026,issn:2058-8437.doi:10.1038/s41578-025- 00831-0[Online].Available:https://doi.org/10.1038/s41578- 025-00831-0

    work page doi:10.1038/s41578-025- 2026

  38. [38]

    Predictingnewresearchdirectionsinmate- rialsscienceusinglargelanguagemodelsandconceptgraphs

    T.Marwitzetal.,“Predictingnewresearchdirectionsinmate- rialsscienceusinglargelanguagemodelsandconceptgraphs”, Nature Machine Intelligence, 2026,issn: 2522-5839.doi: 10. 1038/s42256-026-01206-y [Online].Available:https://doi.org/ 10.1038/s42256-026-01206-y

    work page doi:10.1038/s42256-026-01206-y 2026

  39. [39]

    A large-scale randomized study of large language model feedback in peer review

    N. Thakkar et al., “A large-scale randomized study of large language model feedback in peer review”,Nature Machine Intelligence, vol. 8, no. 3, pp. 326–336, 2026,issn: 2522-5839. doi: 10.1038/s42256-026-01188-x [Online]. Available: https: //doi.org/10.1038/s42256-026-01188-x

    work page doi:10.1038/s42256-026-01188-x 2026

  40. [40]

    KG-Registry — kghub.org, https://kghub.org/kg-registry/ resource/semmeddb/semmeddb.html,[Accessed30-03-2026]

    work page 2026

  41. [41]

    cuebeatsnocue

    LangGraph: Agent Orchestration Framework for Reliable AI Agents — langchain.com, https://www.langchain.com/ langgraph,[Accessed27-03-2026]. 12–21 Arxiv, May 2026 Evidence-Grounded Frontier Mapping and Agentic Hypothesis Generation in Nanomedicine A. CORPUS-CONSTRUCTION FLOW Section3givestheformaldefinitionofcorpusassembly,representa- tion,graphconstructio...

    work page 2026