Pith. sign in

REVIEW 3 major objections 5 minor 55 references

THBKG: A Temporal Biomedical Knowledge Graph for Decision-Aligned Clinical Advancement Prediction

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Graph propagation over a rewound biomedical knowledge graph outranks direct-evidence baselines at predicting which Phase II programmes advance, reaching 4.3--4.5 times the rest-of-ranking rate at the top ten pairs per therapeutic area.

desk verdict A genuinely useful temporal KG and benchmark whose headline 'decision-aligned' claim needs a node-feature ablation before it fully lands. read the letter →

arxiv 2608.05982 v1 pith:H6PJ7CN2 submitted 2026-08-06 cs.LG q-bio.QM

Pui Chung Siu , Claudia Cabrera , Mani Mudaliar , Arkaitz Zubiaga This is my paper · ORCID
classification cs.LGq-bio.QM
keywords temporalknowledgegraphbiomedicalclinicaladvancementpredictiondecision-alignedevaluationdataleakageneuralnetworksdrugtargetidentification
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Phase II drug programmes fail mainly because the target is not causally linked to the disease, so a method that ranks hypotheses on the evidence available when the decision was made could redirect investment toward the pairs most likely to advance. This paper builds THBKG, a biomedical knowledge graph whose 11.1 million edges each carry the year the underlying observation entered the record, making the graph rewound to any cutoff, and it defines a decision-aligned benchmark that scores each target--disease pair only on evidence datable before that pair's own Phase II entry. On that benchmark, graph propagation over the rewound graph outranks every direct-evidence reference, reaching a relative success of 4.3--4.5 at the top ten pairs per therapeutic area, meaning those pairs advanced at more than four times the rate of the rest. The gain is concentrated on the 72.8% of pairs that have no direct target--disease evidence at their decision point, where a direct-edge model has nothing to read and the encoders still rank five- to sixfold above chance by propagating over the intervening biology. If right, the paper turns retrospective validation into a routine query: any hypothesis with a decision date can be scored on the evidence its maker actually held.

What carries the argument

The load-bearing object is THBKG's event-time index together with its as-of query. Every temporal edge carries the year its evidence changed and is written only in years where the cumulative support score $s_{ij}$ differs from the previous year, so for any cutoff $y$ the graph $G[y]$ is recovered by taking the most recent stored edge with timestamp $<y$; the two static ontology hierarchies carry a sentinel timestamp and are never masked. Each temporal edge also carries a novelty score $n_{ij}$ that peaks shortly after evidence arrives and decays with age, so a pair's profile encodes both how much support exists and when it arrived. The decision-aligned benchmark masks all edges that postdate a pair's Phase II entry, and heterogeneous encoders propagate over the surviving subgraph; an edge-aware variant modulates attention by the score and novelty on each edge. A path-based explainer adapted to the decision-time subgraph then decomposes each prediction into the lowest-cost routes under a learned edge mask, capped at the encoder's receptive field.

What would settle it

Re-run the evaluation with curation-dated edges (affected-pathway bulk ingestion, somatic-mutation updates) re-timestamped to the true publication years of their supporting studies, and with node features recomputed from pre-decision snapshots; if the evidence-free stratum's relative success advantage over the ridge baseline collapses to chance, the central claim is an artefact of timestamp and feature leakage rather than multi-hop propagation.

Watch

Extended reading notes

Core claim

The paper's central claim is that multi-hop propagation over a graph of dated biomedical evidence predicts which target--disease hypotheses advance from Phase II to Phase III better than direct target--disease evidence alone, provided the evaluation is aligned to the decision date. THBKG stores each temporal edge with the year its evidence changed, so the as-of query $G[y]$ returns the subgraph of edges predating cutoff $y$; applying it per pair reconstructs the evidence landscape at that pair's own Phase II entry. Under this protocol the strongest graph encoders reach a therapeutic-area mean relative success of 4.3--4.5 at the top ten pairs, against 2.65 for the ridge-regression baseline and 2.58 for the global association-score reference, with the advantage largest on the 5,239 evaluation pairs (72.8%) carrying no direct target--disease evidence at the decision point. The authors present the twelve encoder/edge-feature conditions as reference points rather than a single winner: which encoder leads depends on the cutoff, edge-feature modulation helps some architectures and not others, and the consistent finding is where the graph helps, not which model helps.

Load-bearing premise

The benchmark's validity rests on the assumption that every piece of evidence and every feature used to score a pair really existed before that pair's Phase II decision, but timestamps sometimes record when a database added evidence rather than when it was discovered, and node features are drawn from present-day resources.

Editorial extensions

If this is right

  • Any target--disease hypothesis with a decision date can be validated retrospectively by scoring it on the evidence that existed at that date and comparing the score with what the programme later did, so the testable set grows as the record accrues.
  • The regime where the method helps is precisely the regime a direct-evidence model cannot see: pairs with no direct target--disease evidence, where propagation over intervening biology ranks the top pairs at five- to sixfold above chance.
  • Because the as-of query has a free cutoff, the same substrate extends to novel target identification and drug repurposing, not just Phase II advancement.
  • No single encoder or edge-feature recipe is separable from the others, so the durable contribution is the graph and the decision-aligned evaluation protocol rather than a specific architecture.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A natural test the paper does not run is to rebuild node features from pre-decision snapshots instead of present-day resources; if the evidence-free stratum's advantage vanishes, the gains are partly an artefact of feature leakage rather than pure propagation.
  • Because timestamps can record curation dates rather than discovery dates, re-dating affected-pathway and somatic-mutation edges by their underlying publication years would isolate whether the relative-success lead survives genuinely decision-time evidence.
  • The first-time-in-Phase-II stratum is too small ($n=170$) to conclude anything, and it is the stratum a discovery programme cares most about; a later ClinicalTrials.gov release with more such pairs could decide whether propagation helps at all where no clinical precedent exists.
  • The paper's own caveat that the model sits inside the loop it learns from implies a reusable risk: a ranking model rewarded for enrichment may sharpen the field's existing focus on well-studied targets, so a guard would be to measure whether the gains persist on understudied targets.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper introduces THBKG, a temporal heterogeneous biomedical knowledge graph containing 110,396 entities and 11.1M edges, with each temporal edge dated by the year its evidence changed. The central methodological proposal is the as-of query G[y] (Equation 3), which reconstructs the subgraph of evidence predating a cutoff, enabling per-instance decision-aligned evaluation. On this substrate the authors define a Phase II-to-III advancement benchmark: for each target--disease pair, predict advancement using only evidence datable before the pair's own Phase II entry. They evaluate four heterogeneous graph encoders, an edge-aware variant (EA-HGT), and two edge-aware GATv2 variants against ridge regression (RDG) and the Open Targets score (OTS), reporting therapeutic-area-mean relative success at top-N. The main empirical claim is that graph propagation reaches RS@10 of 4.3--4.5 versus roughly 2.6 for the direct-evidence baselines, with the gain concentrated on the 72.8% of pairs that have no direct target--disease evidence at the decision point. The paper also adapts PaGE-Link to produce decision-time subgraph explanations and releases the graph, pipeline, splits, and code.

Significance. If the decision-aligned interpretation is sustained, the paper makes a genuinely useful contribution: it is the first heterogeneous biomedical knowledge graph, to my knowledge, that supports per-instance as-of queries with multi-hop propagation, and the benchmark protocol is carefully constructed. The authors deserve credit for releasing the data and code, documenting the resource with a datasheet, reporting multiple aggregation methods to avoid over-reading a macro mean, and explicitly flagging the underpowered first-time-target stratum and the possibility that the model learns the field's prior-attention policy rather than novel biology. The paper is also unusually candid about the limitation that edge timestamps can record curation rather than discovery. However, the central scientific claim -- that the reported enrichment comes from multi-hop propagation over pre-decision evidence -- rests on an untested assumption about node features, detailed below. That assumption is load-bearing for the headline result, so the verdict is major revision rather than acceptance.

major comments (3)
  1. [Section 2.1, Equation (3), Appendix D.2] The decision-aligned protocol masks edges with timestamps at or after the decision year, but it does not time-restrict node features. As documented in Appendix D.2, target features include genetic-constraint scores such as pLI/pLOEUF and expression specificity from single-cell RNA-seq, and disease features are embeddings from a pre-trained biomedical language model; these are present-day resources whose underlying data postdate many Phase II entries in the 2016--2021 evaluation window. Consequently, Equation (3) does not implement the paper's stated claim that a pair is scored 'on evidence datable before that decision': the model can consume post-decision information through node features even when every edge is correctly masked. The paper provides no feature-only baseline and no ablation with time-restricted or removed node features, so the interpretation that the RS@10 advantage of graph encoders over RDG/OTS, especially on the 72.8% evidence-free stratum, reflects multi-hop propagation over pre-decision evidence is not established. I would like to see either (i) node features recomputed from resources available as of each pair's decision year, or (ii) an ablation that removes or masks node features, plus a feature-only model using the same node features.
  2. [Section 5, Appendix D.1] The paper acknowledges that edge timestamps can record curation rather than discovery: affected-pathway evidence spikes with bulk ingestion of datasources in 2018 and 2021, and somatic-mutation evidence declines when the Cancer Gene Census tightens its criteria. This is presented as a limitation, but it is also a correctness risk for the paper's decision-aligned claim, because a pair can be classified as evidence-free at its decision point when the supporting studies had already been published but not yet curated. Since the evidence-free stratum is where the graph encoders show their largest advantage, the manuscript should quantify how many evaluation pairs fall into the evidence-free stratum only because of curation lag, and should include a sensitivity analysis that excludes or re-dates the affected datasources. As written, the reader cannot tell whether part of the headline enrichment is an artifact of curation-date misalignment rather than genuine lack of direct evidence.
  3. [Section 4.2, Table 3] The paper's headline 'outranks every direct-evidence reference' is supported by the therapeutic-area mean of RS@N, but the paper itself notes that this mean is an equally weighted average over 13 high-variance areas and can be dominated by a few areas. The authors do report the per-area median and pooled estimates, and they are appropriately cautious in Section 4.3. My concern is narrower: the paper does not provide confidence intervals or significance tests for the RS@N differences at the specific cutoffs of the abstract's claim (RS@10 = 4.3--4.5 versus 2.6 for RDG/OTS). A paired Wilcoxon is reported only for the pooled-versus-evidence-free comparison, not for the headline RS@10 numbers. Since the abstract's claim is a point estimate at a small cutoff, the manuscript should report a bootstrap or permutation interval for the headline RS@10 difference, or explicitly state that the claim is about the direction and consistency across cutoffs rather than the precise multiple.
minor comments (5)
  1. [Section 4.2, final paragraph] The word 'disussed' should be 'discussed'.
  2. [Section 1, paragraph 4] 'tho no resources are available' appears to be a typo for 'though no resources are available'.
  3. [Section 3.4 and Equation (4)] The symbol N is used both for the evaluation-set size (N = 7,193) and as the cutoff in RS@N; this overloading is confusing in Section 4.2 and in the figures. Please use separate symbols, for example n_eval and k.
  4. [Appendix D.2] Please state the release or access dates of the resources used to construct node features, so that readers can assess the temporal overlap with the 2016--2021 evaluation window directly.
  5. [Section 4.4 and Appendix H] The explanation analysis is demonstrated on a single case study (IL17F to psoriatic arthritis). A small quantitative evaluation, such as the fraction of explained top-ranked pairs whose recovered paths contain only pre-decision edges, would strengthen the claim that the explainer is decision-aligned in practice.

Circularity Check

0 steps flagged · score 2.0 of 10

No load-bearing circularity: the central result is an empirical held-out ranking with the supervision label excluded from message passing, and the only self-citation is a minor, non-load-bearing design inheritance.

full rationale

The paper's derivation chain is self-contained at the equation level. The advancement label is stored as a target-to-disease edge but is explicitly excluded from message passing (Appendix Table 4), and the as-of query G[y] (Equation 3) masks edges by each pair's Phase II entry year. All conditions are scored on the same held-out evaluation pairs (7,193), with RDG and OTS regenerated on the same release and split rather than transcribed, so no fitted parameter is renamed as a prediction. The only author-overlap citation is GATher [29], whose co-author Mani Mudaliar also appears here; it supplies the graph schema and node-feature construction, not the advancement result or the RS@N outcome, so it is a minor self-citation rather than a load-bearing circular step. The static, present-day node features (Appendix D.2) and curation-dated edge timestamps (Section 5, Appendix D.1) are a genuine temporal-leakage and correctness risk for the decision-aligned claim, but that is an empirical validity concern, not an equation-level circularity, and the paper itself discloses the timestamp limitation.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

No new physical or biological entities are postulated. The edge-aware attention variant EA-HGT is a model modification, not an entity. The free parameters listed are the hand-set constants that shape the features and labels used in the central evaluation.

free parameters (3)
  • Novelty score decay constants = k=2, m=3, W=5
    Equation 2; hand-chosen in the paper, control how quickly evidence novelty decays; the ablation shows novelty alone is the strongest single edge-feature condition, so these choices affect the central results.
  • Harmonic sum truncation = min(n,50)
    Equation 1; caps the number of evidence items in the cumulative score following the Open Targets convention; a modeling choice rather than a fitted constant.
  • Advancement label window = 3 years
    Section 3.1; a pair is positive only if Phase III is registered within three years of Phase II entry; follows Czech et al., but the window is a label-defining choice that controls base rates.
assumptions (4)
  • domain assumption Open Targets harmonic-sum convention measures target-disease evidence strength
    Section 2.1; the cumulative edge score s_ij is defined by this convention, and the graph's edge attributes and the OTS baseline both rely on it.
  • ad hoc to paper Static ontology hierarchies are time-invariant
    Section 2.1 and Equation 3: disease, GO and Reactome hierarchies carry sentinel timestamps and are never masked, assuming definitions do not change over the 1995-2025 window.
  • ad hoc to paper Node features may be computed from data that postdate a decision without breaking decision-aligned evaluation
    Appendix D.2: target genetic-constraint scores and disease embeddings come from current resources and pretrained models, with no as-of restriction; this is an unstated premise of the evaluation.
  • domain assumption Advancement labels derived from ClinicalTrials.gov registration are accurate proxies for clinical progression
    Section 3.1; labels exclude ambiguous cases but rely on registry completeness and the three-year window.

how reviews work

0 comments
Cite this review

Pith. "Pith review of THBKG: A Temporal Biomedical Knowledge Graph for Decision-Aligned Clinical Advancement Prediction." pith.science (2026). https://pith.science/paper/H6PJ7CN2

@misc{pith2026260805982,
  author       = {Pith},
  title        = {Pith review of: THBKG: A Temporal Biomedical Knowledge Graph for Decision-Aligned Clinical Advancement Prediction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/H6PJ7CN2}},
  note         = {Machine review of arXiv:2608.05982}
}
read the original abstract

Inadequate target--disease linkage accounts for 40--50\% of Phase~II efficacy failures, so anticipating which programmes will advance would let sponsors back the hypotheses most likely to reach patients. What a programme can be judged on is the evidence that supported its linkage \emph{when it entered the clinic}. No existing biomedical knowledge graph allows that evidence profile to be assembled as of a past date. We present the Temporal Heterogeneous Biomedical Knowledge Graph (THBKG), which describes and predicts therapeutic target--disease links through time: 110,396 entities and 11.1M edges across nineteen relation types, each edge carrying the year its evidence changed, so a pair's profile can be recovered as it stood when its own decision fell due. On this graph we define a decision-aligned benchmark that predicts, for a target--disease pair entering Phase~II, whether it advances to Phase~III on evidence datable before that decision. Graph propagation over the THBKG outranks every direct-evidence reference scored under the same decision-aligned protocol, reaching a relative success of 4.3--4.5 at the top ten pairs per therapeutic area. The gain concentrates on the 72.8\% of pairs with no direct target--disease evidence at their decision point, where a direct-edge model has nothing to read: the encoders still rank five- to sixfold above chance, recovering the signal by propagating over the intervening biology. Adapting a path-based explainer to the decision-time subgraph decomposes each prediction into the evidence landscape behind the hypothesis for explainable prediction. We release the THBKG as a continually updated substrate for studying therapeutic target hypotheses by retrospective validation.

Figures

Figures reproduced from arXiv: 2608.05982 by the authors.

Figure 1
Figure 1. The THBKG. (a) Construction pipeline: dated ev [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Temporal edge distribution of the THBKG. Chord [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Ranking and classification performance of the [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (6 more)
Figure 5
Figure 5. Figure 5: Static edge distribution of the THBKG across evi [PITH_FULL_IMAGE:figures/full_fig_p014_5.png]
Figure 6
Figure 6. Figure 6: Relative success by therapeutic area and top- [PITH_FULL_IMAGE:figures/full_fig_p015_6.png]
Figure 7
Figure 7. Figure 7: Relative success by evidence/clinical-history stratum and top- [PITH_FULL_IMAGE:figures/full_fig_p016_7.png]
Figure 8
Figure 8. Figure 8: Per-therapeutic-area distribution of RS@ [PITH_FULL_IMAGE:figures/full_fig_p017_8.png]
Figure 9
Figure 9. Figure 9: Aggregation-method divergence for the ablation conditions of Table 3. For each model and cutoff ( [PITH_FULL_IMAGE:figures/full_fig_p018_9.png]
Figure 11
Figure 11. Figure 11: Evidence available over time for the prediction of [PITH_FULL_IMAGE:figures/full_fig_p019_11.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

55 extracted references · 24 canonical work pages

  1. [1]

    Rishabh Agarwal, Max Schwarzer, P. S. Castro, Aaron C. Courville, and Marc G. Bellemare. 2021. Deep Reinforcement Learning at the Edge of the Statistical Precipice. InNeural Information Processing Systems, Vol. 34. 29304–29320. doi:10. 48550/arxiv.2108.13264

    arXiv doi:10.48550/arxiv.2108.13264 2021

  2. [2]

    Devansh Arpit, Huan Wang, Yingbo Zhou, and Caiming Xiong. 2021. Ensemble of Averages: Improving Model Selection and Boosting Performance in Domain Generalization. InNeural Information Processing Systems. doi:10.52202/068431- 0601

    doi:10.52202/068431- 2021

  3. [3]

    John Arrowsmith and Philip Miller. 2013. Trial watch: phase II and phase III attrition rates 2011-2012.Nature reviews. Drug discovery12, 8 (2013), 569. doi:10. 1038/nrd4090

    work page 2013

  4. [4]

    M Ashburner, C A Ball, J A Blake, D Botstein, H Butler, J M Cherry, A P Davis, K Dolinski, S S Dwight, J T Eppig, M A Harris, D P Hill, L Issel-Tarver, A Kasarskis, S Lewis, J C Matese, J E Richardson, M Ringwald, G M Rubin, and G Sherlock

    work page

  5. [5]

    Albert-László Barabási, Natali Gulbahce, and Joseph Loscalzo. 2011. Network medicine: a network-based approach to human disease.Nature Reviews Genetics 12, 1 (2011), 56–68. doi:10.1038/nrg2918

    doi:10.1038/nrg2918 2011

  6. [6]

    Anna Breit, Simon Ott, Asan Agibetov, and Matthias Samwald. 2020. OpenBi- oLink: a benchmarking framework for large-scale biomedical link prediction. Bioinformatics36, 13 (2020), 4097–4098. doi:10.1093/bioinformatics/btaa274

    doi:10.1093/bioinformatics/btaa274 2020

  7. [7]

    Shaked Brody, Uri Alon, and Eran Yahav. 2021. How Attentive are Graph Atten- tion Networks?ICLR 2022 - 10th International Conference on Learning Represen- tations(2021). doi:10.48550/arxiv.2105.14491

    arXiv doi:10.48550/arxiv.2105.14491 2021

  8. [8]

    Li Cai, Xin Mao, Yuhao Zhou, Zhaoguang Long, Changxu Wu, and Man Lan

    work page

Show all 55 references
  1. [9]

    Payal Chandak, Kexin Huang, and Marinka Zitnik. 2023. Building a Knowledge Graph to Enable Precision Medicine.Scientific Data10, 1 (2023), 67. doi:10.1038/ s41597-023-01960-3

    2023

  2. [10]

    Eric Czech, Rafal Wojdyla, Daniel Himmelstein, Daniel Frank, Nick Miller, Jack Milwid, Adam Kolom, and Jeff Hammerbacher. 2024. CLINICAL ADVANCE- MENT FORECASTING.medRxiv(2024), 2024.08.02.24311422. doi:10.1101/2024. 08.02.24311422

    2024 doi

  3. [11]

    Jesus de la Fuente, Guillermo Serrano, Uxía Veleiro, Mikel Casals, Laura Vera, Marija Pizurica, Nuria Gómez-Cebrián, Leonor Puchades-Carrasco, Antonio 1http://doi.org/10.5281/zenodo.438045 8 THBKG: A Temporal Biomedical Knowledge Graph for Decision-Aligned Clinical Advancement...

    doi

  4. [12]

    Gritsenko, Zhe Zhao, N

    Mostafa Dehghani, Yi Tay, A. Gritsenko, Zhe Zhao, N. Houlsby, Fernando Diaz, Donald Metzler, and O. Vinyals. 2021. The Benchmark Lottery. InAdvances in Neural Information Processing Systems (NeurIPS). doi:10.48550/arxiv.2107.07002

    2021 doi arXiv

  5. [13]

    Edwards, Ruth Isserlin, Gary D

    Aled M. Edwards, Ruth Isserlin, Gary D. Bader, Stephen V. Frye, Timothy M. Willson, and Frank H. Yu. 2011. Too many roads not taken.Nature470, 7333 (2011), 163–165. doi:10.1038/470163a

    2011 doi

  6. [14]

    Falaguera, Ellen M

    Maria J. Falaguera, Ellen M. McDonagh, David Ochoa, Polina V. Rusina, Juan Maria Roldan-Romero, David G. Hulcoop, Andrew R. Leach, and Ian Dun- ham. 2025. Temporal trends in evidence supporting novel drug target discovery. Nat. Commun.17, 1 (2025), 492–. doi:10.1038/s41467-025-67180-y

    2025 doi

  7. [15]

    Alberto García-Durán, Sebastijan Dumančić, and Mathias Niepert. 2018. Learning Sequence Encoders for Temporal Knowledge Graph Completion. InProceedings of the 2018 Conference on Empirical Methods in Natural Language Processing. 4816–4821. doi:10.18653/v1/d18-1516

    2018 doi

  8. [16]

    Anna Gaulton, Anne Hersey, Michał Nowotka, A Patrícia Bento, Jon Cham- bers, David Mendez, Prudence Mutowo, Francis Atkinson, Louisa J Bellis, Elena Cibrián-Uhalte, Mark Davies, Nathan Dedman, Anneli Karlsson, María Paula Magariños, John P Overington, George Papadatos, Ines Sm...

    2017 doi

  9. [17]

    Timnit Gebru, Jamie Morgenstern, Briana Vecchione, Jennifer Wortman Vaughan, Hanna Wallach, Hal Daumé III, and Kate Crawford. 2021. Datasheets for Datasets. Commun. ACM64, 12 (2021), 86–92. doi:10.1145/3458723

    2021 doi

  10. [18]

    Marc Gillespie, Bijay Jassal, Ralf Stephan, Marija Milacic, Karen Rothfels, Andrea Senff-Ribeiro, Johannes Griss, Cristoffer Sevilla, Lisa Matthews, Chuqiao Gong, Chuan Deng, Thawfeek Varusai, Eliot Ragueneau, Yusra Haider, Bruce May, Veronica Shamovsky, Joel Weiser, Timothy B...

    2022

  11. [19]

    Ishaan Gulrajani and David Lopez-Paz. 2021. In Search of Lost Domain General- ization. InInternational Conference on Learning Representations. doi:10.48550/ arxiv.2007.01434

    2021 doi arXiv

  12. [20]

    Daniel Scott Himmelstein, Antoine Lizee, Christine Hessler, Leo Brueggeman, Sabrina L Chen, Dexter Hadley, Ari Green, Pouya Khankhanian, and Sergio E Baranzini. 2017. Systematic integration of biomedical knowledge prioritizes drugs for repurposing.eLife6, e26726 (2017), e26726...

    2017 doi

  13. [21]

    Hingorani, Valerie Kuan, Chris Finan, Felix A

    Aroon D. Hingorani, Valerie Kuan, Chris Finan, Felix A. Kruger, Anna Gaulton, Sandesh Chopade, Reecha Sofat, Raymond J. MacAllister, John P. Overington, Harry Hemingway, Spiros Denaxas, David Prieto, and Juan Pablo Casas. 2019. Improving the odds of drug development success th...

    2019 doi

  14. [22]

    Zitnik, Yuxiao Dong, Hongyu Ren, Bowen Liu, Michele Catasta, and J

    Weihua Hu, Matthias Fey, M. Zitnik, Yuxiao Dong, Hongyu Ren, Bowen Liu, Michele Catasta, and J. Leskovec. 2020. Open Graph Benchmark: Datasets for Machine Learning on Graphs. InNeural Information Processing Systems. doi:10.48550/arxiv.2005.00687

    2020 doi arXiv

  15. [23]

    Ziniu Hu, Yuxiao Dong, Kuansan Wang, and Yizhou Sun. 2020. Heterogeneous Graph Transformer.The Web Conference 2020 - Proceedings of the World Wide Web Conference, WWW 2020(2020), 2704–2710. doi:10.1145/3366423.3380027

    2020

  16. [24]

    Hwang, Daniel Carpenter, Julie C

    Thomas J. Hwang, Daniel Carpenter, Julie C. Lauffenburger, Bo Wang, Jessica M. Franklin, and Aaron S. Kesselheim. 2016. Failure of Investigational Drugs in Late-Stage Clinical Development and Publication of Trial Results.JAMA Internal Medicine176, 12 (2016), 1826–1833. doi:10....

    2016

  17. [25]

    Woojeong Jin, Meng Qu, Xisen Jin, and Xiang Ren. 2020. Recurrent event network: Autoregressive structure inferenceover temporal knowledge graphs. InProceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP). 6669–6683. doi:10.18653/v1/2020.e...

    2020 doi

  18. [26]

    Sayash Kapoor and Arvind Narayanan. 2023. Leakage and the reproducibility crisis in machine-learning-based science.Patterns (N. Y.)4, 9 (2023), 100804. doi:10.1016/j.patter.2023.100804

    2023

  19. [27]

    Emily A King, J Wade Davis, and Jacob F Degner. 2019. Are drug targets with genetic support twice as likely to be approved? Revised estimates of the impact of genetic support for drug mechanisms on the probability of drug approval. PLoS Genetics15, 12 (2019), e1008489. doi:10....

    2019 doi

  20. [28]

    Painter, Coco Chengliang Dong, and Matthew R

    Eric Vallabh Minikel, Jeffery L. Painter, Coco Chengliang Dong, and Matthew R. Nelson. 2024. Refining the impact of genetic evidence on clinical success.Nature 629, 8012 (2024), 624–629. doi:10.1038/s41586-024-07316-0

    2024 doi

  21. [29]

    Crowther

    David Narganes-Carlon, Anniek Myatt, Mani Mudaliar, and Daniel J. Crowther

  22. [30]

    Nelson, Hannah Tipney, Jeffery L

    Matthew R. Nelson, Hannah Tipney, Jeffery L. Painter, Judong Shen, Paola Nico- letti, Yufeng Shen, Aris Floratos, Pak Chung Sham, Mulin Jun Li, Junwen Wang, Lon R. Cardon, John C. Whittaker, and Philippe Sanseau. 2015. The support of human genetic evidence for approved drug in...

    2015 doi

  23. [31]

    Hulcoop, Ian Dunham, and Ellen M

    David Ochoa, Andrew Hercules, Miguel Carmona, Daniel Suveges, Asier Gonzalez-Uriarte, Cinzia Malangone, Alfredo Miranda, Luca Fumis, Denise Carvalho-Silva, Michaela Spitzer, Jarrod Baker, Javier Ferrer, Arwa Raies, Olesya Razuvayevskaya, Adam Faulconbridge, Eirini Petsalaki, P...

    2021

  24. [32]

    Tudor I Oprea, Cristian G Bologa, Søren Brunak, Allen Campbell, Gregory N Gan, Anna Gaulton, Shawn M Gomez, Rajarshi Guha, Anne Hersey, Jayme Holmes, Ajit Jadhav, Lars Juhl Jensen, Gary L Johnson, Anneli Karlson, Andrew R Leach, Avi Ma’ayan, Anna Malovannaya, Subramani Mani, S...

    2018

  25. [33]

    arXiv:2409.16327 [q-bio.QM] doi:10.48550/arxiv.2409.16327

    GATher: Graph Attention Based Predictions of Gene-Disease Links.arXiv [q-bio.QM](2024). arXiv:2409.16327 [q-bio.QM] doi:10.48550/arxiv.2409.16327

    2024 doi arXiv

  26. [34]

    Przemysław Pobrotyn, Tomasz Bartczak, Mikołaj Synowiec, Radosław Biało- brzeski, and Jarosław Bojar. 2020. Context-Aware Learning to Rank with Self- Attention. arXiv:2005.10084 [cs.IR] doi:10.48550/arxiv.2005.10084

    2020 doi arXiv

  27. [35]

    Chamberlain, Fabrizio Frasca, Davide Eynard, Federico Monti, and M

    Emanuele Rossi, B. Chamberlain, Fabrizio Frasca, Davide Eynard, Federico Monti, and M. Bronstein. 2020. Temporal Graph Networks for Deep Learning on Dy- namic Graphs. InICML 2020 Workshop on Graph Representation Learning and Beyond. doi:10.48550/arxiv.2006.10637

    2020 doi arXiv

  28. [36]

    Scannell, Alex Blanckley, Helen Boldon, and Brian Warrington

    Jack W. Scannell, Alex Blanckley, Helen Boldon, and Brian Warrington. 2012. Diagnosing the decline in pharmaceutical R&D efficiency.Nature reviews. Drug discovery11, 3 (2012), 191–200. doi:10.1038/nrd3681

    2012 doi

  29. [37]

    Sandra Orchard, Mais Ammari, Bruno Aranda, Lionel Breuza, Leonardo Briganti, Fiona Broackes-Carter, Nancy H Campbell, Gayatri Chavali, Carol Chen, Noemi del Toro, Margaret Duesbury, Marine Dumousseau, Eugenia Galeota, Ursula Hinz, Marta Iannuccelli, Sruthi Jagannathan, Rafael ...

    2014

  30. [38]

    Amand F Schmidt, Chris Finan, Maria Gordillo-Marañón, Folkert W Asselbergs, Daniel F Freitag, Riyaz S Patel, Benoît Tyl, Sandesh Chopade, Rupert Faraway, Magdalena Zwierzyna, and Aroon D Hingorani. 2020. Genetic drug target validation using Mendelian randomisation.Nature Commu...

    2020

  31. [39]

    Seligmann, P

    F. Seligmann, P. Becker, Michael Volpp, and G. Neumann. 2023. Beyond Deep En- sembles: A Large-Scale Evaluation of Bayesian Deep Learning under Distribution Shift. InNeural Information Processing Systems. doi:10.48550/arxiv.2306.12306

    2023 doi arXiv

  32. [40]

    Morimoto, and Luís A

    Thomas Stoeger, Martin Gerlach, Richard I. Morimoto, and Luís A. Nunes Amaral

  33. [41]

    Schlichtkrull, Thomas Kipf, Peter Bloem, Rianne van den Berg, Ivan Titov, and M

    M. Schlichtkrull, Thomas Kipf, Peter Bloem, Rianne van den Berg, Ivan Titov, and M. Welling. 2018. Modeling Relational Data with Graph Convolutional Networks. InExtended Semantic Web Conference. Springer, 593–607. doi:10.1007/978-3-319- 93417-4_38

    2018 doi

  34. [42]

    Katerina Trajanoska, Claude Bhérer, Daniel Taliun, Sirui Zhou, J Brent Richards, and Vincent Mooser. 2023. From target discovery to clinical drug development with human genetics.Nature620, 7975 (2023), 737–745. doi:10.1038/s41586-023- 06388-8

    2023 doi

  35. [43]

    Nitin, and P

    Shikhar Vashishth, Soumya Sanyal, V. Nitin, and P. Talukdar. 2020. Composition- based Multi-Relational Graph Convolutional Networks. InInternational Confer- ence on Learning Representations. doi:10.48550/arxiv.1911.03082

    2020 doi arXiv

  36. [44]

    Brian Walsh, Sameh K Mohamed, and Vít Nováček. 2020. BioKG. InProceedings of the 29th ACM International Conference on Information & Knowledge Management. 3173–3180. doi:10.1145/3340531.3412776

    2020

  37. [45]

    Xuanhui Wang, Cheng Li, Nadav Golbandi, Michael Bendersky, and Marc Najork

  38. [46]

    Chi Heem Wong, Kien Wei Siah, and Andrew W. Lo. 2019. Estimation of clinical trial success rates and related parameters.Biostatistics20, 2 (2019), 273–286. 9 Siu et al. doi:10.1093/biostatistics/kxx069

    2019 doi

  39. [47]

    Duxin Sun, Wei Gao, Hongxiang Hu, and Simon Zhou. 2022. Why 90% of clinical drug development fails and how to improve it?Acta Pharmaceutica Sinica. B12, 7 (2022), 3049–3062. doi:10.1016/j.apsb.2022.02.002

    2022 doi

  40. [48]

    Edge fea- tures

    Shichang Zhang, Jiani Zhang, Xiang Song, Soji Adeshina, Da Zheng, C. Faloutsos, and Yizhou Sun. 2023. PaGE-Link: Path-based Graph Neural Network Expla- nation for Heterogeneous Link Prediction. InThe Web Conference. doi:10.1145/ 3543507.3583511 A Ethical Use of Data The THBKG ...

    2016

  41. [52]

    InPro- ceedings of the 27th ACM International Conference on Information and Knowledge Management (CIKM)

    The LambdaLoss Framework for Ranking Metric Optimization. InPro- ceedings of the 27th ACM International Conference on Information and Knowledge Management (CIKM). ACM, 1313–1322. doi:10.1145/3269206.3271784

  42. [54]

    Duzhen Zhang, Zixiao Wang, Zhong-Zhi Li, Yahan Yu, Shuncheng Jia, Jiahua Dong, Haotian Xu, Xing Wu, Yingying Zhang, Tielin Zhang, Jie Yang, Xiuying Chen, and Le Song. 2025. MedKGent: A large language model agent framework for constructing temporally evolving medical knowledge ...

    2025 doi arXiv

  43. [2000]

    The Gene Ontology Consortium.Nature Genetics25, 1 (2000), 25–29

    Gene ontology: tool for the unification of biology. The Gene Ontology Consortium.Nature Genetics25, 1 (2000), 25–29. doi:10.1038/75556

    2000 doi

  44. [2018]

    doi:10.1371/journal.pbio.2006643

    Large-scale investigation of the reasons why potentially important genes are ignored.PLoS Biology16, 9 (2018), e2006643. doi:10.1371/journal.pbio.2006643

    2018 doi

  45. [2024]

    04782(2024)

    A Survey on Temporal Knowledge Graph: Representation Learning and Applications.arXiv preprint arXiv:2403. 04782(2024). arXiv:2403.04782 [cs.CL] doi:10.48550/arxiv.2403.04782

    2024 doi arXiv

  46. [2025]

    doi:10.1038/s42256-025-00987-y

    Towards a more inductive world for drug repurposing approaches.Nature Machine Intelligence7, 3 (2025), 495–508. doi:10.1038/s42256-025-00987-y

    2025 doi

  47. [3255]

    doi:10.1038/s41467-020-16969-0

    doi

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.