CoChat
Sign Up
Cited by 249Computational Drug Discovery MethodsMachine Learning in HealthcareBiomedical Text Mining and Ontologies

A foundation model for clinician-centered drug repurposing

Open access
Huang, Kexin, Chandak, Payal, Wang, Qianwen, Havaldar, Shreyas, Vaid, Akhil, Leskovec, Jure, Nadkarni, Girish N., Glicksberg, Benjamin S. +2 more·2024·Nature Medicine
DOI 10.1038/s41591-024-03233-xPMC PMC11645266

Published version of a preprint

Open this paper in CoChat — free accountOpen PDF

Abstract

Drug repurposing – identifying new therapeutic uses for approved drugs – is often serendipitous and opportunistic, expanding the use of drugs for new diseases. The clinical utility of drug repurposing AI models remains limited because the models focus narrowly on diseases for which some drugs already exist. Here, we introduce T x GNN, a graph foundation model for zero-shot drug repurposing, identifying therapeutic candidates even for diseases with limited treatment options or no existing drugs. Trained on a medical knowledge graph, T x GNN utilizes a graph neural network and metric-learning module to rank drugs as potential indications and contraindications across 17,080 diseases. When benchmarked against eight methods, T x GNN improves prediction accuracy for indications by 49.2% and contraindications by 35.1% under stringent zero-shot evaluation. To facilitate model interpretation, T x GNN’s Explainer module offers transparent insights into multi-hop medical knowledge paths that form T x GNN’s predictive rationales. Human evaluation of T x GNN’s Explainer showed that T x GNN’s predictions and explanations perform encouragingly on multiple axes of performance beyond accuracy. Many of T x GNN’s novel predictions align with off-label prescriptions clinicians make in a large healthcare system. T x GNN’s drug repurposing predictions are accurate, consistent with off-label drug use, and can be investigated by human experts through multi-hop interpretable rationales.

AI snapshot

Written from the paper’s full text.

TL;DR

TxGNN is a graph foundation model for zero-shot drug repurposing that predicts indications and contraindications across 17,080 diseases, including diseases with no approved drugs. It combines a graph neural network with disease-similarity metric learning and a GraphMask-based explainer, and its predictions align with off-label prescribing in a large health system.

Key findings

  • Under stringent zero-shot evaluation, TxGNN improved prediction accuracy for indications by 49.2% and contraindications by 35.1% versus 8 benchmarked methods.
  • In a zero-shot holdout where diseases had no approved drugs during training, TxGNN surpassed the best baseline by up to 19% for indications and 23.9% for contraindications.
  • Across 9 disease-area holdouts, TxGNN achieved relative AUPRC gains of 0.5–59.3% (average 25.72%) for indications and 11.8–35.6% (average 18.67%) for contraindications.
  • In EMR analysis, top-ranked predicted drugs had on average 107% higher log(OR) than bottom-50% predictions; the top candidate had log(OR) 2.26 versus 2.92 for FDA-approved indications.
  • In a human study with 12 experts and 192 trials, explanations improved accuracy (+46%, P = 0.0443) and confidence (+49%, P = 0.0041) versus predictions alone.

Methods

  • TxGNN is trained on a heterogeneous medical knowledge graph with 123,527 nodes and 8,063,026 edges, covering 17,080 diseases, 7,957 drugs, 9,388 indications, and 30,675 contraindications.
  • The model uses a heterogeneous GNN encoder, disease-similarity metric learning with disease pooling, pretraining across all relationships, fine-tuning on drug–disease relationships, and a GraphMask explainer for multi-hop rationales.
  • Evaluation used AUPRC against 8 methods with five random data splits; additional analyses included 9 disease-area holdouts, a 12-participant human usability study, and Mount Sinai EMR data from 1,272,085 adults across 478 diseases and 1,290 drugs.

Limitations

  • The paper states TxGNN depends on medical KG quality, which may lack host–pathogen interaction data and genetic-variant pathogenicity information, and may contain data biases or outdated information.
  • The paper states the human evaluation was a pilot with a small number of experts (n = 12), and EMR analyses may be limited by unaccounted confounders and selection biases.
  • The paper states predicted drugs would need extensive screening to establish safety, efficacy, dosage, and treatment timing; no prospective clinical validation is reported.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1038/s41591-024-03233-x

Cites

57 of 57 references verified0 citation errors · 0 not checkable

Statistics

Claims the paper makes

0 of 57 cited sources checked · 0 supported · 0 flagged

  • Findingnone given

    Many of TxGNN's new predictions align well with off-label prescriptions that clinicians previously made in a large healthcare system.

    “Many of TxGNN’s new predictions align well with off-label prescriptions that clinicians previously made in a large healthcare system.”
    • Also stated in: § Main · boosted
      “Many of TxGNN’s predictions have shown alignment with off-label prescriptions made by clinicians in a large healthcare system”
    Quote in the paperSubstantial claim
  • Findingreasoning

    Once trained, TxGNN performs zero-shot inference on new diseases without additional parameters or fine-tuning.

    “Once trained, TxGNN performs zero-shot inference on new diseases without additional parameters or fine-tuning.”
    Quote in the paperSubstantial claim
  • Findingcitation

    The ability of existing machine learning models to identify therapeutic candidates for diseases with incomplete, sparse data and zero-known therapies drops drastically.

    “Unfortunately, the ability of existing machine learning models to identify therapeutic candidates for diseases with incomplete, sparse data and zero-known therapies drops drastically16,22 (as we demonstrate across eight benchmarks in Fig. 2c,d).”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites [16] Gysi, D. M. et al. Network medicine framework for identifying drug-repurposing opportunities for Covid-19. Proc. Natl Acad. Sci. USA 118, e2025581118 (2021). source

      Not checked in the free run.

    • cites [22] Guney, E. Reproducible drug repurposing: when similarity does not suffice. In Pacific Symposium on Biocomputing 132–143 (World Scientific, 2017).

      Not checked in the free run.

  • Findingown data

    92% of 17,080 diseases examined in the study have no indications.

    “92% of 17,080 diseases examined in our study have no indications.”
    Quote in the paperSubstantial claim
  • Findingcitation

    The gold standard labels for evaluating TxGNN come from a previously curated medical KG consisting of 9,388 indications and 30,675 contraindications.

    “The gold standard labels for evaluating such a model come from our previously curated medical KG9 (Fig. 1d and Supplementary Tables 4 and 5), which consists of 9,388 indications and 30,675 contraindications23.”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites [9] Chandak, P., Huang, K. & Zitnik, M. Building a knowledge graph to enable precision medicine. Sci. Data 10, 67 (2023). source

      Not checked in the free run.

    • cites [23] Avram, S. et al. DrugCentral 2021 supports drug discovery and repositioning. Nucleic Acids Res. 49, D1160–D1169 (2021). source

      Not checked in the free run.

  • May combine two claims. 41
  • The quote only partly states this claim. 1
  • Classified as probably not a substantial claim. 32

  • May combine two claims.

    Findingown data

    TxGNN is a graph foundation model for zero-shot drug repurposing that identifies therapeutic candidates for diseases with limited treatment options or no existing drugs, across 17,080 diseases.

    “Here we introduce TxGNN, a graph foundation model for zero-shot drug repurposing, identifying therapeutic candidates even for diseases with limited treatment options or no existing drugs.”
    • Also stated in: § Main · plain
      “In the present study, we introduced TxGNN, a graph foundation model for multi-disease, zero-shot drug repurposing that predicts drug-repurposing candidates across 17,080 diseases, including diseases without treatments (Fig. 1c).”
    Quote in the paperSubstantial claim
    • The authors state this as an assumption: Effective drugs directly target disease-perturbed networks or indirectly propagate therapeutic effects through disease-associated networks.

  • May combine two claims.

    Findingreasoning

    TxGNN is a single pretrained model that adapts across many diseases, instead of training disease-specific models for every disease.

    “Foundation models like TxGNN are transforming deep learning: instead of training disease-specific models for every disease, TxGNN is a single pretrained model that adapts across many diseases.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    When benchmarked against 8 methods, TxGNN improves prediction accuracy for indications by 49.2% and contraindications by 35.1% under stringent zero-shot evaluation.

    “When benchmarked against 8 methods, TxGNN improves prediction accuracy for indications by 49.2% and contraindications by 35.1% under stringent zero-shot evaluation.”
    Quote in the paperSubstantial claim

  • The quote only partly states this claim.

    Findingown data

    Human evaluation of TxGNN's Explainer showed that TxGNN's predictions and explanations perform encouragingly on multiple axes of performance beyond accuracy, including accuracy, trust, usefulness and time efficiency.

    “Human evaluation of TxGNN’s Explainer showed that TxGNN’s predictions and explanations perform encouragingly on multiple axes of performance beyond accuracy.”
    • Also stated in: § Main · plain
      “Our human evaluation of TxGNN’s Explainer showed that TxGNN’s explanations perform encouragingly on multiple axes of performance, including accuracy, trust, usefulness and time efficiency.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: Human experts seek to understand the reasoning behind model predictions to validate the model’s hypotheses and better understand candidate treatment mechanisms.

  • May combine two claims.

    Findingown data

    BioBERT performed best for indication prediction in seven of nine disease areas, whereas RGCN was the best baseline for contraindications in eight of nine.

    “BioBERT performed best for indication prediction in seven of nine disease areas, whereas RGCN was the best baseline for contraindications in eight of nine.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingreasoning

    Established methods often fell short in challenging zero-shot settings, although they achieved satisfactory results in conventional drug-repurposing evaluations.

    “Although established methods achieved satisfactory results in conventional drug-repurposing evaluations, they often fell short in challenging settings.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    TxGNN achieved relative AUPRC gains of 0.5-59.3% (average 25.72%) across 9 disease areas for indications and 11.8-35.6% (average 18.67%) for contraindications.

    “TxGNN achieved relative AUPRC gains of 0.5–59.3% (average 25.72%) across 9 disease areas for indications and 11.8–35.6% (average 18.67%) for contraindications.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    TxGNN Explainer indicated that zolpidem's action on GABRG2 might reduce autism susceptibility and improve prefrontal cortex function.

    “However, TxGNN Explainer indicated that zolpidem’s action on GABRG2 might reduce autism susceptibility and improve prefrontal cortex function.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Tretinoin, transported by albumin (ALB) and targeting ALDH1A2, helps mitigate collagen loss and inflammation, as highlighted in TxGNN's predictive rationale.

    “Tretinoin, transported by albumin (ALB) and targeting ALDH1A2, helps mitigate collagen loss and inflammation, as highlighted in TxGNN’s predictive rationale (Fig. 5c), indicating that TxGNN’s predictive rationale is aligned with medical reasoning.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Focusing on the most predictive relationships (edges with importance scores >0.5, representing an average of 14.9% of edges from the KG), the model's performance showed a slight reduction from AUPRC = 0.890 (s.d. 0.006) to AUPRC = 0.886 (s.d. 0.005), indicating that TxGNN's Explainer effectively identifies key associations and explanations accurately reflect TxGNN's internal reasoning.

    “Focusing on the most predictive relationships (that is, edges with importance scores >0.5, representing an average of 14.9% of edges from the KG), the model’s performance showed a slight reduction from AUPRC = 0.890 (s.d. 0.006) to AUPRC = 0.886 (s.d. 0.005), indicating that TxGNN’s Explainer effectively identifies key associations34 and explanations accurately reflect TxGNN’s internal reasoning.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [34] Agarwal, C., Queen, O., Lakkaraju, H. & Zitnik, M. Evaluating explainability for graph neural networks. Sci. Data 10, 144 (2023). source

      Not checked in the free run.

  • May combine two claims.

    Findingown data

    In the standard benchmarking setting, three of eight existing methods achieved AUPRC > 0.80 and HAN was the best at 0.873 AUPRC.

    “Our experimental results in this setting concur, with three of eight existing methods achieving AUPRC > 0.80 and HAN as the best at 0.873 AUPRC.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    The top therapeutic candidate predicted by TxGNN had a log(OR) of 2.26, approaching the average log(OR) of 2.92 for FDA-approved indications.

    “Although the average log(OR) stands at 1.09, the top therapeutic candidate predicted by TxGNN had a log(OR) of 2.26, approaching the average log(OR) of 2.92 for FDA-approved indications, indicating the enrichment of off-label drug prescriptions among TxGNN’s top-ranked predictions.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The medical KG covers 17,080 diseases, 92% lacking FDA-approved drugs, covering rare and less-understood complex diseases.

    “The medical KG covers 17,080 diseases, 92% lacking FDA-approved drugs, covering rare and less-understood complex diseases.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    TxGNN substantially improves over the next best method in predicting indications (19.0% AUPRC gain) and contraindications (23.9% AUPRC gain).

    “TxGNN substantially improves over the next best method in predicting indications (19.0% AUPRC gain) and contraindications (23.9% AUPRC gain).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    TxGNN consistently ranked newly introduced drugs favorably and, in two instances, placed newly approved medications within the top 5% of predicted drugs.

    “TxGNN consistently ranked newly introduced drugs favorably and, in two instances, placed the newly approved medications within the top 5% of predicted drugs.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    TxGNN outperformed all methods across all nine disease areas for both tasks, demonstrating its broad generalizability and accuracy in zero-shot drug repurposing.

    “However, TxGNN outperformed all methods across all nine disease areas for both tasks, demonstrating its broad generalizability and accuracy in zero-shot drug repurposing.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Findingnone given

    As most diseases do not share underlying pathology, they have low similarity scores.

    “As most diseases do not share underlying pathology, they have low similarity scores.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingown data

    The optimal λ is 0.7.

    “Finally, we used parameter search and found optimal λ = 0.7.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    Drug repurposing leverages existing safety and efficacy data of approved drugs, allowing for faster translation to the clinic and reduced development costs compared to developing new drugs from scratch.

    “Drug repurposing leverages existing safety and efficacy data of approved drugs, allowing for faster translation to the clinic and reduced development costs compared to developing new drugs from scratch4 (Fig. 1a).”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [4] Pushpakom, S. et al. Drug repurposing: progress, challenges and recommendations. Nat. Rev. Drug Discov. 18, 41–58 (2019). source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S2 (checkable).

    Premisecitation

    There is a pressing need to develop therapies for many diseases that currently lack treatments.

    “There is a pressing need to develop therapies for many diseases that currently lack treatments1,2.”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [1] Feigin, V. L. et al. Burden of neurological disorders across the us from 1990-2017: a global burden of disease study. JAMA Neurol. 78, 165–176 (2021). source

      Not checked in the free run.

    • cites [2] O’Connell, D. Neglected diseases. Nature 449, 157–157 (2007). source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Of over 7,000 rare diseases worldwide, only 5-7% have FDA-approved drugs.

    “Of over 7,000 rare diseases worldwide, only 5–7% have Food and Drug Administration (FDA)-approved drugs3.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [3] Rare Disease Day 2021. US Food and Drug Administration fda.gov/news-events/fda-voices/rare-disease-day-2021-fda-shows-sustained-support-rare-disease-product-development-during-public (2023).

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    Approximately 30% of FDA-approved drugs are issued at least one new post-approval indication and many drugs have accrued over ten indications over the years.

    “Approximately 30% of FDA-approved drugs are issued at least one new post-approval indication and many drugs have accrued over ten indications over the years6.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [6] Sahragardjoonegani, B., Beall, R. F., Kesselheim, A. S. & Hollis, A. Repurposing existing drugs for new uses: a cohort study of the frequency of FDA-granted new indication exclusivities since 1997. J. Pharm. Policy Pract. 14, 3 (2021). source

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    Around 95% of rare diseases have no FDA-approved drugs and up to 85% of rare diseases do not have even one drug developed that would show promise in rare disease treatment, diagnosis or prevention.

    “Moreover, around 95% of rare diseases have no FDA-approved drugs and up to 85% of rare diseases do not have even one drug developed that would show promise in rare disease treatment, diagnosis or prevention21.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [21] Fermaglich, L. J. & Miller, K. L. A comprehensive study of the rare diseases and conditions targeted by orphan drug designations and approvals over the forty years of the orphan drug act. Orphanet J. Rare Dis. 18, 1–8 (2023). source

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    Thalidomide was originally developed for morning sickness during pregnancy, repurposed in 1964 for an autoimmune complication of leprosy and again in 2006 for multiple myeloma.

    “Originally developed to help with morning sickness during pregnancy, thalidomide was repurposed in 1964 for an autoimmune complication of leprosy and again in 2006 for multiple myeloma8.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [8] Jourdan, J.-P., Bureau, R., Rochais, C. & Dallemagne, P. Drug repositioning: a brief overview. J. Pharm. Pharmacol. 72, 1145–1151 (2020). source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    Machine learning has been used to analyze high-throughput molecular interactomes to unravel genetic architecture perturbed in disease and help design therapies to target them.

    “Machine learning has been used to analyze high-throughput molecular interactomes to unravel genetic architecture perturbed in disease12,13 and help design therapies to target them14.”
    Quote in the paperNot a substantial claim0 of 3 sources checked
    • cites [12] Ruiz, C., Zitnik, M. & Leskovec, J. Identification of disease treatment mechanisms through the multiscale interactome. Nat. Commun. 12, 1–15 (2021). source

      Not checked in the free run.

    • cites [13] Goh, K.-I. et al. The human disease network. Proc. Natl Acad. Sci. USA 104, 8685–8690 (2007). source

      Not checked in the free run.

    • cites [14] Baraba´si, A.-L., Gulbahce, N. & Loscalzo, J. Network medicine: a network-based approach to human disease. Nat. Rev. Genet. 12, 56–68 (2011). source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    These strategies identify therapeutic candidates based on their impact on cell signaling, gene expression and disease phenotypes.

    “These strategies identify therapeutic candidates based on their impact on cell signaling, gene expression and disease phenotypes5,10–12.”
    Quote in the paperNot a substantial claim0 of 4 sources checked
    • cites [5] Abdelsayed, M., Kort, E. J., Jovinge, S. & Mercola, M. Repurposing drugs to treat cardiovascular disease in the era of precision medicine. Nat. Rev. Cardiol. 19, 751–764 (2022). source

      Not checked in the free run.

    • cites [10] Menche, J. et al. Uncovering disease-disease relationships through the incomplete interactome. Science 347, 1257601 (2015). source

      Not checked in the free run.

    • cites [11] Zitnik, M. et al. Evolution of resilience in protein interactomes across the tree of life. Proc. Natl Acad. Sci. USA 116, 4426–4433 (2019). source

      Not checked in the free run.

    • cites [12] Ruiz, C., Zitnik, M. & Leskovec, J. Identification of disease treatment mechanisms through the multiscale interactome. Nat. Commun. 12, 1–15 (2021). source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S1 (asserted).

    Premisecitation

    These models make predictions for a disease by retrieving drugs from the dataset that appear similar to existing treatments.

    “However, Duran et al.30 reasoned that these models make predictions for a disease by retrieving drugs from the dataset that appear similar to existing treatments.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Duran et al.30 Duran-Frigola, M. et al. Extending the small-molecule similarity principle to all levels of biology with the chemical checker. Nat. Biotechnol. 38, 1087–1096 (2020). source

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    AVPR2 and NPR1 genes play pivotal roles in regulating fluid and electrolyte balance via complementary but distinct pathways; AVPR2 contributes to water retention and urine concentration, whereas NPR1 facilitates vasodilatation, lowers blood pressure and enhances water excretion.

    “AVPR2 and NPR1 genes play pivotal roles in regulating fluid and electrolyte balance via complementary but distinct pathways. AVPR2 contributes to water retention and urine concentration, whereas NPR1 facilitates vasodilatation, lowers blood pressure and enhances water excretion47.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [47] Wishart, D. S. et al. DrugBank 5.0: a major update to the DrugBank database for 2018. Nucleic Acids Res. 46, D1074–D1082 (2018). source

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    Some subtypes of Ehlers-Danlos syndrome have been associated with a pathogenic mutation in the ALB gene and weakly linked to ALDH1AI.

    “Furthermore, some subtypes of Ehlers–Danlos syndrome have been associated with a pathogenic mutation in the ALB gene in ref. 42 and weakly linked to ALDH1AI in ref. 43.”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites [42] Landrum, M. J. et al. Clinvar: improvements to accessing data. Nucleic Acids Res. 48, D835–D844 (2020). source

      Not checked in the free run.

    • cites [43] Javed, S. et al. ALDH1 & CD133 in invasive cervical carcinoma & their association with the outcome of chemoradiation therapy. Indian J. Med. Res. 154, 367 (2021). source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Shortcut learning underlies many deep-learning failures.

    “This phenomenon is known as shortcut learning31 and underlies many deep-learning failures32.”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [31] Bickel, S., Brückner, M. & Scheffer, T. Discriminative learning under covariate shift. J. Mach. Learn. Res. 10, 2137–2155 (2009).

      Not checked in the free run.

    • cites [32] Niven, T. & Kao, H.-Y. Probing neural network comprehension of natural language arguments. In Proc. of the 57th Annual Meeting of the Association for Computational Linguistics 4658–4664 (ACL, 2019). source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    The literature indicates that deferasirox may effectively eliminate hepatic iron.

    “the literature indicates that it may effectively eliminate hepatic iron48.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [48] Seetharaman, J. & Sarma, M. S. Chelation therapy in liver diseases of childhood: current status and response. World J. Hepatol. 13, 1552 (2021). source

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    Zolpidem has shown stimulative effects in neurological conditions, temporarily awakening underactive neurons, suggesting a potential therapeutic use for neurodevelopmental disorders.

    “Surprisingly, zolpidem has shown stimulative effects in neurological conditions, temporarily awakening underactive neurons, suggesting a potential therapeutic use for neurodevelopmental disorders39.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [39] Bomalaski, M. N., Claflin, E. S., Townsend, W. & Peterson, M. D. Zolpidem for the treatment of neurologic disorders: a systematic review. JAMA Neurol. 74, 1130–1139 (2017). source

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    Shortcut decision rules tend to perform well on standard benchmarks but fail on challenging conditions, such as predicting drugs for rare diseases with no treatment options or subtypes of complex diseases with distinct disease mechanisms.

    “Shortcut decision rules tend to perform well on standard benchmarks but fail on challenging conditions33, such as predicting drugs for rare diseases with no treatment options or subtypes of complex diseases with distinct disease mechanisms.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [33] Geirhos, R. et al. Shortcut learning in deep neural networks. Nat. Mach. Intell. 2, 665–673 (2020). source

      Not checked in the free run.

  • May combine two claims.

    Interpretationreasoning

    Amyl nitrite, which targets NPR1, emerges as a potential therapeutic option for NSIAD, confirming the consistency of TxGNN's explanations with medical evidence.

    “Amyl nitrite, which targets the NPR1 gene, emerges as a potential therapeutic option for NSIAD, confirming the consistency of TxGNN’s explanations with medical evidence.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    The standard benchmarking strategy used to evaluate drug-repurposing AI models aligns with the approach predominantly used in the literature.

    “This evaluation method aligns with the approach predominantly used in the literature12,14–16,18–20.”
    Quote in the paperNot a substantial claim0 of 7 sources checked
    • cites [12] Ruiz, C., Zitnik, M. & Leskovec, J. Identification of disease treatment mechanisms through the multiscale interactome. Nat. Commun. 12, 1–15 (2021). source

      Not checked in the free run.

    • cites [14] Baraba´si, A.-L., Gulbahce, N. & Loscalzo, J. Network medicine: a network-based approach to human disease. Nat. Rev. Genet. 12, 56–68 (2011). source

      Not checked in the free run.

    • cites [15] Li, M. M., Huang, K. & Zitnik, M. Graph representation learning in biomedicine and healthcare. Nat. Biomed. Eng. 6, 1353–1369 (2022). source

      Not checked in the free run.

    • cites [16] Gysi, D. M. et al. Network medicine framework for identifying drug-repurposing opportunities for Covid-19. Proc. Natl Acad. Sci. USA 118, e2025581118 (2021). source

      Not checked in the free run.

    • cites [18] Zitnik, M., Agrawal, M. & Leskovec, J. Modeling polypharmacy side effects with graph convolutional networks. Bioinformatics 34, i457–i466 (2018). source

      Not checked in the free run.

    • cites [19] Guney, E., Menche, J., Vidal, M. & Bara´basi, A.-L. Network-based in silico drug efficacy screening. Nat. Commun. 7, 1–13 (2016). source

      Not checked in the free run.

    • cites [20] Cheng, F., Kova´cs, I. A. & Baraba´si, A.-L. Network-based prediction of drug combinations. Nat. Commun. 10, 1–11 (2019). source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Study design and conductnone given

    Reinitializing the decoder weights before fine-tuning mitigates the risk of negative knowledge transfer.

    “It is important to note that the weights in the decoder, specifically for DistMult, wr, are reinitialized before fine-tuning to mitigate the risk of negative knowledge transfer.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Study design and conductnone given

    The research complied with all relevant ethical regulations.

    “We have complied with all relevant ethical regulations.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Study design and conductnone given

    The variant with the highest validation performance was selected for test set evaluation and downstream analyses.

    “The variant with the highest validation performance was selected for test set evaluation and downstream analyses.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Interpretationreasoning

    TxGNN streamlines drug-repurposing prediction, especially when the limited availability of disease-specific datasets hinders drug development.

    “TxGNN streamlines drug-repurposing prediction, especially when the limited availability of disease-specific datasets hinders drug development.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

    Premisecitation

    Rare diseases often offer extensive opportunities for therapeutic advancements.

    “Despite this, rare diseases often offer extensive opportunities for therapeutic advancements3.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [3] Rare Disease Day 2021. US Food and Drug Administration fda.gov/news-events/fda-voices/rare-disease-day-2021-fda-shows-sustained-support-rare-disease-product-development-during-public (2023).

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Study design and conductnone given

    The pilot human evaluation engaged a small number (n = 12) of clinicians and scientists and prioritized an in-depth analysis with a small but qualified group of human experts over a broader study with a larger, potentially less specialized, participant pool.

    “Our pilot human evaluation engaged a small number (n = 12) of clinicians and scientists and prioritized an in-depth analysis with a small but qualified group of human experts over a broader study with a larger, potentially less specialized, participant pool.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationnone given

    TxGNN is a unified model for predicting indications and contraindications across 17,080 diseases, suitable for early drug repurposing beyond single therapeutic areas.

    “TxGNN is a unified model for predicting indications and contraindications across 17,080 diseases, suitable for early drug repurposing beyond single therapeutic areas.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Study design and conductnone given

    Informed consent was waived for analyzing EMRs retrospectively.

    “Informed consent was waived for analyzing EMRs retrospectively.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

    Study design and conductnone given

    Appropriate provisions were taken to ensure the safety and well-being of all participants involved.

    “Appropriate provisions were taken to ensure the safety and well-being of all participants involved.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Study design and conductnone given

    Unaccounted confounders and selection biases might have limited the ability to draw conclusions from the calculated drug enrichment scores.

    “Despite the promising performance of TxGNN’s predictions on medical records, unaccounted confounders and selection biases might have limited the ability to draw conclusions from the calculated drug enrichment scores.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based). May combine two claims.

    Interpretationnone given

    Multi-hop interpretable predictive rationales can enhance transparent use of TxGNN, fostering trust and aiding human experts.

    “Multi-hop interpretable predictive rationales can enhance transparent use of TxGNN, fostering trust and aiding human experts.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning

    TxGNN generalizes to diseases with few treatment options by modeling latent relationships between diseases and performing zero-shot inference for diseases that the model never encountered during training.

    “TxGNN generalizes to diseases with few treatment options by modeling latent relationships between diseases and performing zero-shot inference for diseases that the model never encountered during training.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning

    Use of a learnable attention mechanism is ineffective because it overvalues the original embeddings for well-represented diseases, neglecting the auxiliary embedding.

    “Use of a learnable attention mechanism is ineffective, because it overvalues the original embeddings for well-represented diseases, neglecting the auxiliary embedding.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning

    For well-represented diseases in the KG, the disease-disease metric learning embedding might be unnecessary and could bias the disease embedding.

    “Special consideration is needed because, for well-represented diseases in the KG, the disease–disease metric learning embedding might be unnecessary and could bias the disease embedding.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductnone given

    The sequence in which predictions were presented, along with the conditions, were randomized and counterbalanced across participants and tasks.

    “The sequence in which predictions were presented, along with the conditions (TxGNN Explorer or the baseline approach), were randomized and counterbalanced across participants and tasks.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Interpretationreasoning

    Increasing the extent of masking tends to enlarge the discrepancy between the updated and original predictions.

    “These objectives inherently entail a tradeoff: increasing the extent of masking tends to enlarge the discrepancy between the updated and original predictions.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Premisecitation

    The findings from the pilot study comparison motivated the adoption of path-based explanations, which were preferred based on user feedback.

    “The findings from this comparison motivated the adoption of path-based explanations, which were preferred based on user feedback.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [25] Wang, Q., Huang, K., Chandak, P., Zitnik, M. & Gehlenborg, N. Extending the nested model for user-centric XAI: a design study on gnn-based drug repurposing. IEEE Trans. Vis. Comput. Graph. 29, 1266–1276 (2023). source

      Not checked in the free run.

  • May combine two claims.

    Interpretationreasoning

    Disease representations produced solely by GNNs are insufficient and not ideal for direct similarity computations.

    “Given the insufficiency of disease representations produced solely by GNNs, these representations are not ideal for direct similarity computations.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning

    Using a model to extract predictive information from similar but better-represented diseases in the KG can enrich the target disease embedding and improve model performance for the disease.

    “Using a model to extract predictive information from similar but better-represented diseases in the KG, the target disease embedding can be enriched, improving model performance for the disease.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: Human physiology is an interconnected system where diseases exhibit similarities.

  • Classified as probably not a substantial claim.

    Study design and conductnone given

    All internal EMRs were deidentified before computational analysis and model development.

    “All internal EMRs were deidentified before computational analysis and model development.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

    Study design and conductnone given

    The research did not result in stigmatization, incrimination, discrimination or personal risk to participants.

    “Our research did not result in stigmatization, incrimination, discrimination or personal risk to participants.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationnone given

    TxGNN's zero-shot drug-repurposing capability allows the model to predict drugs for diseases with limited treatment options and scarce information.

    “TxGNN’s zero-shot drug-repurposing capability allows the model to predict drugs for diseases with limited treatment options and scarce information.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    The molecular basis of many rare diseases remains poorly understood.

    “For instance, the molecular basis of many rare diseases remains poorly understood21.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [21] Fermaglich, L. J. & Miller, K. L. A comprehensive study of the rare diseases and conditions targeted by orphan drug designations and approvals over the forty years of the orphan drug act. Orphanet J. Rare Dis. 18, 1–8 (2023). source

      Not checked in the free run.

  • May combine two claims.

    Interpretationreasoning

    After training, edges with gate value 0 can be removed, and the remaining edges serve as explanations for the model's predictions.

    “After the training process is complete, edges (i, j, r) for which = 0 can be removed. The remaining edges serve as explanations for the model’s predictions.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim. May combine two claims.

    Study design and conductnone given

    The Mount Sinai Institutional Review Board approved the study, ensuring that all clinical data were deidentified.

    “The Mount Sinai Institutional Review Board approved the study, ensuring that all clinical data were deidentified.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Premisecitation

    Previous research has emphasized KG completion, optimizing models across the entire spectrum of relationships within a KG.

    “Previous research has emphasized KG completion, optimizing models across the entire spectrum of relationships within a KG58.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [58] Lin, Y., Liu, Z., Sun, M., Liu, Y. & Zhu, X. Learning entity and relation embeddings for knowledge graph completion. In Proc. of the 29th AAAI Conference on Artificial Intelligence Vol. 29, 2181–2187 (2015). source

      Not checked in the free run.

  • May combine two claims.

    Interpretationreasoning

    A higher degree of node connectivity indicates that the disease is better represented in the knowledge and has a denser local network neighborhood, suggesting a reduced reliance on the disease-disease metric learning embedding.

    “A higher degree indicated that the disease was better represented in the knowledge and had a denser local network neighborhood, suggesting a reduced reliance on the disease–disease metric learning embedding and vice versa.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationnone given

    Predicted drugs align with off-label prescription rates in EMRs and match with the medical consensus of human experts.

    “Predicted drugs align with off-label prescription rates in EMRs and match with the medical consensus of human experts.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim. May combine two claims.

    Study design and conductown data

    Two-sided Bonferroni's correction was applied to the P values, identifying statistically significant drug-disease pairs as those with P < 0.005.

    “Two-sided Bonferroni’s correction was applied to the P values using the statsmodels Python library’s multi-test function, identifying statistically significant drug–disease pairs as those with P < 0.005.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Study design and conductnone given

    The Institutional Review Board at Mount Sinai, New York City, USA, approved the retrospective analysis of internal EMRs.

    “The Institutional Review Board at Mount Sinai, New York City, USA, approved the retrospective analysis of internal EMRs.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim. May combine two claims.

    Study design and conductown data

    Predictions were generated for 1,363 diseases, identified by training a KG on 5% of randomly selected drug-disease pairs, serving as a validation set for early stopping.

    “Predictions were generated for 1,363 diseases, identified by training a KG on 5% of randomly selected drug–disease pairs, serving as a validation set for early stopping.”
    Quote in the paperSubstantial claim
66 of 66

  • The authors state this as an assumption.

    Effective drugs directly target disease-perturbed networks or indirectly propagate therapeutic effects through disease-associated networks.

    “Our model for zero-shot drug-repurposing TxGNN (Methods and Supplementary Fig. 2) operates on the principle that effective drugs directly target disease-perturbed networks or indirectly propagate therapeutic effects through disease-associated networks.”

  • The paper relies on this without support.

    Human experts seek to understand the reasoning behind model predictions to validate the model’s hypotheses and better understand candidate treatment mechanisms.

    “Human experts seek to understand the reasoning behind these predictions to validate the model’s hypotheses and better understand candidate treatment mechanisms.”

  • The paper relies on this without support.

    Off-label drug prescription is a longstanding clinical practice.

    “Recognizing the longstanding clinical practice of off-label drug prescription, we used the enrichment of disease–drug pair co-occurrence in a health system’s electronic medical records (EMRs) as a proxy measure for being a potential indication.”

  • The paper relies on this without support.

    There is no established gold standard against which to validate drugs not yet clinically approved for a disease.

    “As these therapies have not yet been approved for treatment, there is no established gold standard against which to validate them.”

  • The paper relies on this without support.

    Kleefstra's syndrome is caused by mutations in the EHMT1 gene and leads to speech delays, autism spectrum disorder and childhood hypotonia, often featuring underdeveloped brains with dormant neuronal pathways.

    “First, we examined TxGNN’s predictions for Kleefstra’s syndrome, a rare disease caused by mutations in the EHMT1 gene. This condition leads to speech delays, autism spectrum disorder and childhood hypotonia, often featuring underdeveloped brains with dormant neuronal pathways.”

  • The paper relies on this without support.

    Ehlers-Danlos syndrome results from mutations in collagen-coding genes (COL1A1 and COL1A2) and is characterized by impaired wound healing and atypical scars.

    “This disorder results from mutations in collagen-coding genes (COL1A1 and COL1A2) and is characterized by impaired wound healing and atypical scars.”

  • The paper relies on this without support.

    Tretinoin stimulates collagen production in the skin.

    “as a result of its ability to stimulate collagen production in the skin.”

  • The paper relies on this without support.

    NSIAD is characterized by water and sodium imbalance caused by a mutation in the AVPR2 gene.

    “This disease is characterized by water and sodium imbalance caused by a mutation in the AVPR2 gene.”

  • The paper relies on this without support.

    Wilson's disease causes excessive copper accumulation that frequently instigates liver cirrhosis in children.

    “Wilson’s disease, a rare disease causing excessive copper accumulation that frequently instigates liver cirrhosis in children (Fig. 6g)”

  • The paper relies on this without support.

    Human physiology is an interconnected system where diseases exhibit similarities.

    “Human physiology is an interconnected system where diseases exhibit similarities.”

  • The paper relies on this without support.

    Drug–disease interactions are governed by complex biological mechanisms.

    “Given that drug–disease interactions are governed by complex biological mechanisms, the extensive range of biomedical relationships in a KG can offer a comprehensive view of biological systems.”

  • The paper relies on this without support.

    The extensive range of biomedical relationships in a KG can offer a comprehensive view of biological systems.

    “Given that drug–disease interactions are governed by complex biological mechanisms, the extensive range of biomedical relationships in a KG can offer a comprehensive view of biological systems.”
12 of 12

Statistics recomputed

Every reported test result, recomputed from its test statistic and degrees of freedom. A result shows as not adding up only once an independent check has confirmed it.

  • Consistentt(11) = 3.64reported p < .01 · recomputed p = .004

    “t(11) = 3.64, p < .01”

1 of 1

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

  • Incorrect percentage for 8 of 12 participants

    Minor

    Eight of twelve participants is 66.7%, not 75.0%.

    Reported 75 · recomputed 66.666667

1 of 1

Cites

Flagged first, then by citations

6 of the 63 references couldn’t be matched to a known paper.

Research works printed but not linked: 6

  1. Schlichtkrull, M. S., De Cao, N. & Titov, I. Interpreting graph neural networks for NLP with differentiable edge masking. In International Conference on Learning Representations (2021).
  2. Sundararajan, M., Taly, A. & Yan, Q. Axiomatic attribution for deep networks. In Proc. of the International Conference on Machine Learning 3319–3328 (PMLR, 2017).
  3. Cheng, J., Dasoulas, G., He, H., Agarwal, C. & Zitnik, M. GNNDelete: a general strategy for unlearning in graph neural networks. In Proc. of the International Conference on Learning Representations (2023).
  4. Huang, K., Jin, Y., Candes, E. & Leskovec, J. Uncertainty quantification over graph with conformalized graph neural networks. Adv. Neural Inf. Process. Syst. 36, 26699–26721 (2024).
  5. Gilmer, J., Schoenholz, S. S., Riley, P. F., Vinyals, O. & Dahl, G. E. Neural message passing for quantum chemistry. In Proc. of the International Conference on Machine Learning 70, 1263–1272 (PMLR, 2017).
  6. Yang, B., Yih, S. W. T., He, X., Gao, J. & Deng, L. Embedding entities and relations for learning and inference in knowledge bases. In Proc. of the International Conference on Learning Representations (2015).

Cited by

Sorted by citations

Related

Authors
Huang, K., Chandak, P., Wang, Q., Havaldar, S., Vaid, A., Leskovec, J., Nadkarni, G. N., Glicksberg, B. S., Gehlenborg, N., & Zitnik, M.
Affiliations
  • Harvard University (US)
  • Harvard–MIT Division of Health Sciences and Technology (US)
  • Icahn School of Medicine at Mount Sinai (US)
  • Stanford University (US)
ORCID
Venue
Nature Medicine
Publisher
Springer Science and Business Media LLC
Published
2024-09-25
Volume
30
Issue
12
Pages
3601-3613
ISSN
  • 1078-8956 (ISSN-L)
  • 1546-170X
DOI
10.1038/s41591-024-03233-x
PMC
PMC11645266
OpenAlex
W4402827768
Semantic Scholar
50b7ffc1793e7fffe937cab79d1c122111a4db12
Cited by
249
References
63
Citations by year
2026: 126 · 2025: 110 · 2024: 11 · 2023: 2
FWCI
52.35
Type
journal article
Language
en
Open access
Licence
Locations
  • Nature Medicine (publishedVersion, cc-by-nc-nd)
  • medRxiv (acceptedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by-nc-nd)
  • PubMed (publishedVersion)
  • PubMed Central (submittedVersion, cc-by)
Funders
  • National Science Foundation (2339524, 2030459, CAREER)
  • U.S. Department of Defense (FA8702-15-D-0001)
  • Pfizer
  • AstraZeneca
  • Sanofi
  • Harvard University
  • National Institutes of Health (-0001)
  • U.S. Air Force (FA8702‐15‐D‐0001)
  • Harvard Data Science Initiative, Harvard University
  • Eunice Kennedy Shriver National Institute of Child Health and Human Development (R01HD108794)
Topics
  • Physical Sciences › Computer Science › Computational Theory and Mathematics › Computational Drug Discovery Methods
  • Physical Sciences › Computer Science › Artificial Intelligence › Machine Learning in Healthcare
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Biomedical Text Mining and Ontologies
Keywords
drug repurposing, graph foundation models, graph neural networks, medical knowledge graph, drug contraindications, explainable AI, off-label drug use

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-09.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-09.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review