Pith. sign in

REVIEW 3 major objections 5 minor 29 references

Aligning each patient’s personalized gene network with a simulated drug-induced expression profile predicts clinical response more accurately than existing methods and recovers known drug mechanisms.

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

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

Aligning patient-specific gene-regulatory graphs with LINCS-pretrained perturbation embeddings via CLIP-style contrastive learning improves clinical drug-response prediction on TCGA and zero-shot I-SPY2.

T0 review reviewed 2026-07-11 challenge →

load-bearing objection Solid dual-view recipe for patient-level response prediction; absolute numbers on hard splits stay near chance and the TCGA N is small, but the design, ablations, and external check are real enough to read and referee. the 3 major comments →

arxiv 2607.04557 v1 pith:I3ZP5QDG submitted 2026-07-06 cs.LG cs.AIq-bio.QM

Predicting Therapeutic Outcome via Aligning Patient-Specific Knowledge Graph and Gene-Level Perturbation Representations

Dongmin Bang , Sugyun An , Inyoung Sung , Ilho Yun , Sun Kim , Sangseon Lee This is my paper
classification cs.LG cs.AIq-bio.QM
keywords multi-view alignmentknowledge graph modelingtranscriptomic perturbation modelingpatient-specific drug response predictioncontrastive learningprecision oncologygene regulatory networkstransfer learning
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

Predicting whether a cancer patient will respond to a drug from only pre-treatment gene expression is hard: clinical response labels are scarce, and matched post-treatment molecular profiles almost never exist. PREDIKTOR builds two pictures of the same patient–drug pair—one a personalized gene-regulatory network that includes the drug’s known targets, the other a simulated post-treatment expression profile produced by a model pretrained on large cell-line perturbation experiments—then forces the two views to agree in a shared space with contrastive learning before classifying response. On The Cancer Genome Atlas the method outperforms strong baselines under patient, drug, and tissue splits; applied zero-shot to the independent I-SPY2 breast-cancer trial it raises AUROC by 5.6 percent while highlighting genes and pathways already known for paclitaxel. The practical claim is that limited clinical labels can be turned into an interpretable, transferable predictor by coupling patient-specific biology with transferable knowledge of how drugs rewire expression.

Core claim

PREDIKTOR claims that clinical drug response is predicted more accurately and interpretably by aligning two complementary embeddings of the same patient–drug pair: a drug-centric graph embedding of a patient-specific regulatory network (DysRegNet edges plus DrugBank targets) and a frozen CSG2A-generated post-perturbation transcriptomic profile. CLIP-style contrastive learning with drug-context hard negatives aligns the views; their concatenation feeds an end-to-end response classifier. On TCGA the model outperforms state-of-the-art baselines under patient-, drug-, and tissue-split evaluation, and it improves zero-shot AUROC by 5.6% on I-SPY2 while recovering known paclitaxel and breast-cance

What carries the argument

PREDIKTOR multi-view alignment: a GCN-encoded embedding of a patient-specific gene-regulatory graph (DysRegNet + DrugBank) is projected into the same latent space as a frozen CSG2A simulated perturbation profile; a CLIP-style contrastive loss with hard negatives from different drugs forces the two views of each matched pair to agree, after which the concatenated vectors drive response classification.

Load-bearing premise

The method assumes that a frozen cell-line model of drug-induced expression change still yields useful simulated post-treatment profiles for heterogeneous human tumors, and that incomplete drug–target and regulatory edges are enough to ground the patient network.

What would settle it

Train PREDIKTOR on TCGA, then test zero-shot on an external cohort treated with a drug that has neither LINCS L1000 profiles nor reliable DrugBank targets; if AUROC does not beat strong baselines, or if top-attributed genes show no enrichment for that drug’s known mechanisms, the central claim fails.

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

If this is right

  • Patient-level response prediction from pre-treatment transcriptomes improves when static network context and dynamic perturbation signatures are explicitly aligned rather than used alone.
  • Zero-shot transfer to the external I-SPY2 paclitaxel cohort raises AUROC by 5.6% over competing methods without retraining.
  • Gene and pathway attributions from the aligned embeddings recover established mechanisms of paclitaxel and breast-cancer biology.
  • Ablations show both views and the drug-context hard-negative contrastive objective are required for the gains, especially under drug- and tissue-held-out splits.
  • Personalized GRNs outperform a static shared GRN, indicating that patient-specific dysregulation edges carry usable clinical signal.

Where Pith is reading between the lines

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

  • The same alignment recipe could be tried on other scarce-label clinical endpoints (toxicity, acquired resistance) wherever large cell-line perturbation atlases exist but matched post-treatment patient profiles do not.
  • For investigational compounds without documented targets, provisional drug–target edges from chemistry or chemoproteomics would be needed before the network view remains informative.
  • Freezing a large preclinical perturbation encoder may be a general template for transferring omics models into small clinical cohorts without overwriting the pretrained priors.
  • Unexpected genes that rank highly only under multi-view attribution could be prioritized for wet-lab follow-up as candidate biomarkers of response.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. PREDIKTOR predicts binary clinical drug response from pre-treatment patient transcriptomes by aligning two complementary views of each patient–drug pair: (i) a patient-specific knowledge-graph view built from DysRegNet-inferred TF–TG edges (FDR < 0.05, HTRIdb-restricted) plus DrugBank target links, encoded by a two-layer GCN with DREAMwalk initialization and drug-node readout; and (ii) a frozen CSG2A gene–gene attention model pretrained on LINCS L1000 that produces a simulated post-perturbation expression profile over the 978 landmark genes. The two embeddings are projected into a shared space and aligned with a CLIP-style contrastive loss that uses other-drug (drug-context) hard negatives, then concatenated and passed to an MLP trained with BCE + λ_CLIP L_CLIP-context. On a TCGA cohort of 358 patients / 383 pairs / 21 drugs the method is evaluated under patient-, drug-, and tissue-held-out 5-fold CV and is further tested zero-shot on I-SPY2 paclitaxel-treated breast-cancer samples. The paper reports consistent ranking first across most metrics, a 5.6 % AUROC lift on I-SPY2, ablations of each view and of the hard-negative contrastive term, and gene/pathway attributions that recover known paclitaxel and breast-cancer mechanisms.

Significance. If the multi-view alignment truly improves generalization under drug- and tissue-held-out regimes and yields stable, mechanistically coherent attributions, the work would supply a practical template for combining patient-specific regulatory graphs with transferable perturbation priors when matched post-treatment clinical profiles are unavailable. Strengths already present include a clear dual-encoder design, explicit hard-negative strategy, three complementary split protocols, an external zero-shot cohort, component ablations (Fig. 3, Table 2), and quantitative recovery of LINCS paclitaxel consensus genes. These elements make the contribution of genuine interest to the precision-oncology ML community even if absolute performance on the hardest splits remains modest.

major comments (3)
  1. Table 1, drug-split and tissue-split columns: absolute AUROC remains near chance (0.594 ± 0.090 and 0.603 ± 0.085) with large fold variance, while patient-split margins over Random Forest / CSG2A are < 0.01 AUROC. With only 383 labeled pairs and 21 drugs (Methods 2.6.1), the claim of “consistent outperformance” under the hardest generalization regimes is only weakly supported; the ranking-first result can be driven by a few easy drugs/tissues rather than a generalizable multi-view mechanism. Stronger statistical evidence (paired tests across all folds, confidence intervals on the lift, or leave-one-drug-out curves) is needed before the central claim can be considered load-bearing.
  2. Discussion §4 and Methods 2.3.2: the network view is undefined for any compound lacking DrugBank target edges. The paper itself notes that cold-start / investigational drugs therefore lose the graph stream, yet the drug-split evaluation still relies on the remaining drugs that possess targets. This circular dependence undermines the claim that the multi-view design solves the clinical-label scarcity problem for truly novel compounds; either an explicit cold-start protocol (inferred provisional edges or ablation of the drug node) or a clear restriction of the claim is required.
  3. Methods 2.4 / 2.5 and the weakest-assumption note: the frozen CSG2A module is pretrained solely on cell-line LINCS L1000 landmark-gene profiles. No quantitative check is provided that the simulated post-perturbation vectors remain informative for heterogeneous human tumors, especially for the three TCGA drugs absent from LINCS or for tissues poorly represented in L1000. Without such a diagnostic (e.g., correlation of simulated vs. observed post-treatment signatures on any available clinical series, or performance stratified by LINCS coverage), the transferability premise of the perturbation view stays untested.
minor comments (5)
  1. Equation (1) and surrounding text: the dimensions of the concatenated vector (256 + 978) and the subsequent MLP are stated inconsistently with the projection matrices W_G / W_P; a single notation table would help.
  2. Figure 3b: the y-axis label “1 − Absolute Error” is non-standard; reporting mean absolute error or Brier score would be clearer.
  3. Table 3: AUPRC on I-SPY2 is only 0.296; the prevalence of pCR should be stated so readers can judge calibration.
  4. Typographical slips: “DrugBanK”, “Togehter”, “transciptomic”, “grpah” appear in the main text and should be corrected.
  5. Code availability is currently “upon request”; for a methods paper a public repository (or at least a DOI-stamped archive) would strengthen reproducibility claims.

Circularity Check

1 steps flagged

No significant circularity: clinical labels are external, CSG2A/DREAMwalk are frozen pretrained components with independent prior publications, and contrastive alignment is an auxiliary objective rather than a redefinition of the response target.

specific steps
  1. self citation load bearing [Methods 2.4 (Transcriptomic perturbation view) and 2.3.3 (GCN initialization); also Discussion §4]
    "Before training PREDIKTOR, the condition-specific Gene-Gene Attention (CSG2A) module is pretrained on the LINCS L1000 dataset... During the end-to-end learning of PREDIKTOR, the pretrained weights are frozen... Node embeddings... are initialized using DREAMwalk [21]... we kept the pretrained CSG2A and MAT modules frozen to preserve transferable perturbation priors... consistent with prior observations that fine-tuning CSG2A can degrade downstream performance [14]."

    The two core encoders (CSG2A perturbation view and DREAMwalk node initialization) are taken from the authors’ own prior publications and kept frozen. This is ordinary component reuse, not a closed derivation: the clinical response labels and the TCGA/I-SPY2 evaluation remain external, and the CLIP alignment + BCE head are new. The self-citations are therefore not load-bearing for the claimed AUROC gains; they merely supply pretrained feature extractors. Flagged only as a minor self-citation pattern (score contribution ~1).

full rationale

The derivation chain is architectural and empirical, not definitional. Response labels y_p,d come from TCGA clinical metadata and I-SPY2 pCR (Methods 2.6.1); LINCS L1000 pretraining of CSG2A contains no clinical response labels. The network view is built from DysRegNet residual TF–TG edges (FDR<0.05, HTRIdb-restricted) plus DrugBank targets, encoded by a GCN whose node features are initialized by DREAMwalk; the perturbation view is a frozen CSG2A forward pass that produces a simulated post-treatment profile. These two embeddings are projected and aligned by a CLIP-style loss with drug-context hard negatives (Eq. L_CLIP / L_CLIP-context), then concatenated and trained with BCE for binary response. The contrastive term is an auxiliary regularizer (L_total = L_BCE + λ_CLIP L_CLIP-context); it does not redefine or force the clinical label. Self-citations to CSG2A [14] and DREAMwalk [21] are component reuse of previously published modules that are kept frozen; they do not supply a uniqueness theorem or force the reported AUROC by construction. Ablations and external zero-shot evaluation further treat the multi-view design as a testable hypothesis rather than a tautology. Minor self-citation of the authors’ own prior modules is present but not load-bearing for the central claim. Score 1 reflects that single non-circular self-reuse pattern; no fitted-input-called-prediction or self-definitional reduction of the headline metrics exists.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 2 invented entities

The central claim rests on standard ML/graph machinery plus domain assumptions that cell-line perturbation priors and residual-based patient GRNs transfer to clinical response, plus several hand-chosen hyperparameters and the PREDIKTOR multi-view construct itself. No new physical entity is postulated; free parameters are ordinary training/architecture choices.

free parameters (4)
  • lambda_CLIP (contrastive loss weight) = init 5 (trainable)
    Weighted sum L_total = L_BCE + λ_CLIP L_CLIP-context; initially set to 5 and described as trainable; directly trades off alignment vs classification.
  • contrastive temperature tau = learnable
    Learnable temperature in the CLIP-style softmax; shapes the hardness of negatives.
  • GCN/MLP hidden dims, dropout, lr, batch size = 128 / 0.1 / 1e-4 / 32
    Two-layer GCN dim 128, MLP 128-unit hidden, dropout 0.1, Adam 1e-4, batch 32—standard but claim-sensitive architecture choices.
  • DysRegNet FDR threshold and HTRIdb edge prior = FDR < 0.05
    Edges retained at BH FDR < 0.05 and restricted to HTRIdb TF–TG pairs; changes the patient graph topology that the GCN sees.
axioms (5)
  • domain assumption DysRegNet residual deviations from healthy-tissue TF–TG linear baselines identify patient-specific regulatory edges relevant to drug response.
    Methods 2.3.1; entire network view depends on these edges being mechanistically meaningful.
  • domain assumption Frozen CSG2A attention patterns learned on LINCS L1000 cell-line chemical perturbations transfer to human tumor transcriptomes as useful post-treatment proxies.
    Methods 2.4; Discussion notes modules are kept frozen to preserve priors.
  • domain assumption DrugBank target links plus DREAMwalk-initialized heterogeneous GCN embeddings supply drug-centric mechanistic context complementary to expression alone.
    Methods 2.3.2–2.3.3; cold-start drugs without targets break this axiom (Discussion).
  • ad hoc to paper CLIP-style cosine alignment of the two views with other-drug hard negatives improves the joint representation for binary clinical response.
    Methods 2.5.1; Table 2 shows drops without hard negatives or without contrastive loss.
  • standard math Standard GCN message-passing and BCE classification are valid estimators on the constructed graphs and labels.
    Eq. for H^(l+1) and MLP head; conventional ML assumptions.
invented entities (2)
  • PREDIKTOR multi-view patient–drug embedding (aligned network + perturbation views) no independent evidence
    purpose: Unified representation used for end-to-end clinical response classification and attribution.
    The paper’s primary construct; existence is justified only by downstream AUROC/AUPRC and pathway recovery, not by an external physical measurement of the latent itself.
  • Drug-context hard-negative CLIP objective (L_CLIP-context) no independent evidence
    purpose: Force drug-specific inter-view alignment by treating same-patient different-drug pairs as negatives.
    Design choice specific to this paper; Table 2 supports utility but the objective is not independently measured outside the training loop.

reviewed 2026-07-11 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Predicting Therapeutic Outcome via Aligning Patient-Specific Knowledge Graph and Gene-Level Perturbation Representations." pith.science (2026). https://pith.science/paper/I3ZP5QDG

@misc{pith2026260704557,
  author       = {Pith},
  title        = {Pith review of: Predicting Therapeutic Outcome via Aligning Patient-Specific Knowledge Graph and Gene-Level Perturbation Representations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/I3ZP5QDG}},
  note         = {Machine review of arXiv:2607.04557}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Accurate prediction of patient-specific therapeutic response from pre-treatment transcriptomes is hindered by the scarcity of matched clinical response labels and post-treatment molecular profiles. Preclinical transfer-learning models can simulate drug-induced expression changes but are often hard to interpret and unstable, whereas knowledge-graph methods provide mechanistic context yet remain static and fail to capture drug-induced transcriptomic perturbation dynamics. We propose PREDIKTOR, a patient-centered multi-view framework that aligns a personalized network view with a transferable transcriptomic perturbation view to predict clinical drug response. For each patient, we construct an individualized gene regulatory network from tumor expression using DysRegNet and augment it with drug-target links from DrugBank; a graph neural encoder yields a drug-centric, mechanistically grounded embedding. In parallel, a frozen condition-specific gene-gene attention model pretrained on LINCS L1000 generates a simulated post-perturbation transcriptomic profile for the same patient-drug pair. We align the two views in a shared latent space via a CLIP-style contrastive objective with drug-context hard negatives, then concatenate the representations for end-to-end response classification. On TCGA, PREDIKTOR consistently outperforms state-of-the-art baselines under patient-, drug-, and tissue-split evaluations, and transfers zero-shot to the I-SPY2 trial, improving AUROC by 5.6% over competing methods. The aligned embeddings yield stable gene and pathway attributions that recover known mechanisms, supporting actionable and interpretable precision oncology.

Figures

Figures reproduced from arXiv: 2607.04557 by Dongmin Bang, Ilho Yun, Inyoung Sung, Sangseon Lee, Sugyun An, Sun Kim.

Figure 1
Figure 1. Figure 1: Abstract visualization of existing drug response pre [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of PREDIKTOR: Personalized drug Response Estimation via Dual Integration of Knowledge graph and Transferable Omics-driven Representations. a Overall architecture of PREDIKTOR, a patient-centered framework that takes patient transcriptomes as the primary input and predicts patient-specific drug responses. b Contrastive multi-view alignment module, inspired by CLIP, which aligns complementary repres… view at source ↗
Figure 3
Figure 3. Figure 3: Performances of ablation models of PREDIKTOR. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Comparative gene attribution and pathway enrich [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

29 extracted references · 2 linked inside Pith

  1. [1]

    Precision medicine: changing the way we think about healthcare.Clinics, 73:e723, 2018

    Gustavo Rosa Gameiro, Viktor Sinkunas, Gabriel Romero Liguori, and José Otavio Costa Auler-Júnior. Precision medicine: changing the way we think about healthcare.Clinics, 73:e723, 2018

    2018

  2. [2]

    Pharmacogenomics: translating functional genomics into rational therapeutics.science, 286(5439):487–491, 1999

    William E Evans and Mary V Relling. Pharmacogenomics: translating functional genomics into rational therapeutics.science, 286(5439):487–491, 1999

    1999

  3. [3]

    Machine learning approaches to drug response prediction: challenges and recent progress.NPJ precision oncology, 4(1):19, 2020

    George Adam, Ladislav Rampášek, Zhaleh Safikhani, Petr Smirnov, Benjamin Haibe-Kains, and Anna Goldenberg. Machine learning approaches to drug response prediction: challenges and recent progress.NPJ precision oncology, 4(1):19, 2020

    2020

  4. [4]

    The cancer genome atlas pan-cancer analysis project.Nature genetics, 45(10):1113– 1120, 2013

    John N Weinstein, Eric A Collisson, Gordon B Mills, Kenna R Shaw, Brad A Ozenberger, Kyle Ellrott, Ilya Shmulevich, Chris Sander, and Joshua M Stuart. The cancer genome atlas pan-cancer analysis project.Nature genetics, 45(10):1113– 1120, 2013

    2013

  5. [5]

    A landscape of pharmacogenomic interactions in cancer

    Francesco Iorio, Theo A Knijnenburg, Daniel J Vis, Graham R Bignell, Michael P Menden, Michael Schubert, Nanne Aben, Emanuel Gonçalves, Syd Barthorpe, Howard Lightfoot, et al. A landscape of pharmacogenomic interactions in cancer. Cell, 166(3):740–754, 2016

    2016

  6. [6]

    The cancer cell line encyclopedia enables predictive modelling of anticancer drug sensitivity.Nature, 483(7391):603–607, 2012

    Jordi Barretina, Giordano Caponigro, Nicolas Stransky, Kavitha Venkatesan, Adam A Margolin, Sungjoon Kim, Christopher J Wilson, Joseph Lehár, Gre- gory V Kryukov, Dmitriy Sonkin, et al. The cancer cell line encyclopedia enables predictive modelling of anticancer drug sensitivity.Nature, 483(7391):603–607, 2012

    2012

  7. [7]

    Opportunities and challenges in interpretable deep learning for drug sensitivity prediction of cancer cells.Frontiers in Bioinformatics, 2:1036963, 2022

    Bikash Ranjan Samal, Jens Uwe Loers, Vanessa Vermeirssen, and Katleen De Preter. Opportunities and challenges in interpretable deep learning for drug sensitivity prediction of cancer cells.Frontiers in Bioinformatics, 2:1036963, 2022

    2022

  8. [8]

    Deepcdr: a hybrid graph convolutional network for predicting cancer drug response.Bioinformatics, 36(Supplement_2):i911–i918, 2020

    Qiao Liu, Zhiqiang Hu, Rui Jiang, and Mu Zhou. Deepcdr: a hybrid graph convolutional network for predicting cancer drug response.Bioinformatics, 36(Supplement_2):i911–i918, 2020

    2020

  9. [9]

    Deeptta: a transformer-based model for predicting cancer drug response

    Likun Jiang, Changzhi Jiang, Xinyu Yu, Rao Fu, Shuting Jin, and Xiangrong Liu. Deeptta: a transformer-based model for predicting cancer drug response. Briefings in bioinformatics, 23(3), 2022

    2022

  10. [10]

    Drug sensitivity prediction from cell line-based pharmacogenomics data: guidelines for developing machine learning models

    Hossein Sharifi-Noghabi, Soheil Jahangiri-Tazehkand, Petr Smirnov, Casey Hon, Anthony Mammoliti, Sisira Kadambat Nair, Arvind Singh Mer, Martin Ester, and Benjamin Haibe-Kains. Drug sensitivity prediction from cell line-based pharmacogenomics data: guidelines for developing machine learning models. Briefings in bioinformatics, 22(6):bbab294, 2021

    2021

  11. [11]

    Jihye Shin, Yinhua Piao, Dongmin Bang, Sun Kim, and Kyuri Jo. Drpreter: interpretable anticancer drug response prediction using knowledge-guided graph neural networks and transformer.International Journal of Molecular Sciences, 23(22):13919, 2022

    2022

  12. [12]

    Gene expression based inference of cancer drug sensitiv- ity.Nature communications, 13(1):5680, 2022

    Smriti Chawla, Anja Rockstroh, Melanie Lehman, Ellca Ratther, Atishay Jain, Anuneet Anand, Apoorva Gupta, Namrata Bhattacharya, Sarita Poonia, Priyadarshini Rai, et al. Gene expression based inference of cancer drug sensitiv- ity.Nature communications, 13(1):5680, 2022

    2022

  13. [13]

    L1000cds2: Lincs l1000 characteristic direction signatures search engine

    Qiaonan Duan, St Patrick Reid, Neil R Clark, Zichen Wang, Nicolas F Fernandez, Andrew D Rouillard, Ben Readhead, Sarah R Tritsch, Rachel Hodos, Marc Hafner, et al. L1000cds2: Lincs l1000 characteristic direction signatures search engine. NPJ systems biology and applications, 2(1):1–12, 2016

    2016

  14. [14]

    Transfer learning of condition-specific perturbation in gene interactions improves drug response prediction.Bioinfor- matics, 40(Supplement_1):i130–i139, 2024

    Dongmin Bang, Bonil Koo, and Sun Kim. Transfer learning of condition-specific perturbation in gene interactions improves drug response prediction.Bioinfor- matics, 40(Supplement_1):i130–i139, 2024

    2024

  15. [15]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Marina Meila and Tong Zhang, editors, Proceedings of the 38th International Conference on Machine...

    2021

  16. [16]

    I-spy 2: an adaptive breast cancer trial design in the setting of neoadjuvant chemotherapy.Clinical Pharmacology & Therapeutics, 86(1):97–100, 2009

    AD Barker, CC Sigman, Gary J Kelloff, NM Hylton, Donald A Berry, and LJs Esser- man. I-spy 2: an adaptive breast cancer trial design in the setting of neoadjuvant chemotherapy.Clinical Pharmacology & Therapeutics, 86(1):97–100, 2009

    2009

  17. [17]

    Dysregnet: Patient-specific and confounder-aware dysregulated network inference towards precision therapeutics.British journal of pharmacology, 2024

    Johannes Kersting, Olga Lazareva, Zakaria Louadi, Jan Baumbach, David B Blu- menthal, and Markus List. Dysregnet: Patient-specific and confounder-aware dysregulated network inference towards precision therapeutics.British journal of pharmacology, 2024

    2024

  18. [18]

    Htridb: an open-access database for experimentally verified human transcriptional regulation interactions.Nature Precedings, pages 1–1, 2012

    Luiz Bovolenta, Marcio Acencio, and Ney Lemke. Htridb: an open-access database for experimentally verified human transcriptional regulation interactions.Nature Precedings, pages 1–1, 2012

    2012

  19. [19]

    Drugbank 6.0: the drugbank knowledgebase for 2024.Nucleic acids research, 52(D1):D1265–D1275, 2024

    Craig Knox, Mike Wilson, Christen M Klinger, Mark Franklin, Eponine Oler, Alex Wilson, Allison Pon, Jordan Cox, Na Eun Chin, Seth A Strawbridge, et al. Drugbank 6.0: the drugbank knowledgebase for 2024.Nucleic acids research, 52(D1):D1265–D1275, 2024

    2024

  20. [20]

    Semi-supervised classification with graph convolutional networks.arXiv preprint arXiv:1609.02907, 2016

    Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional networks.arXiv preprint arXiv:1609.02907, 2016

    Pith/arXiv arXiv 2016

  21. [21]

    Biomedical knowledge graph learning for drug repurposing by extending guilt-by-association to multiple layers.Nature Communications, 14(1):3570, 2023

    Dongmin Bang, Sangsoo Lim, Sangseon Lee, and Sun Kim. Biomedical knowledge graph learning for drug repurposing by extending guilt-by-association to multiple layers.Nature Communications, 14(1):3570, 2023

    2023

  22. [22]

    Damian Szklarczyk, Rebecca Kirsch, Mikaela Koutrouli, Katerina Nastou, Farrokh Mehryary, Radja Hachilif, Annika L Gable, Tao Fang, Nadezhda T Doncheva, Sampo Pyysalo, et al. The string database in 2023: protein–protein association networks and functional enrichment analyses for any sequenced genome of interest.Nucleic acids research, 51(D1):D638–D646, 2023

    2023

  23. [23]

    Molecule attention transformer.arXiv preprint arXiv:2002.08264, 2020

    Łukasz Maziarka, Tomasz Danel, Sławomir Mucha, Krzysztof Rataj, Jacek Ta- bor, and Stanisław Jastrzębski. Molecule attention transformer.arXiv preprint arXiv:2002.08264, 2020

    Pith/arXiv arXiv 2002

  24. [24]

    Transfer learning enables predictions in network biology

    Christina V Theodoris, Ling Xiao, Anant Chopra, Mark D Chaffin, Zeina R Al Sayed, Matthew C Hill, Helene Mantineo, Elizabeth M Brydon, Zexian Zeng, X Shirley Liu, et al. Transfer learning enables predictions in network biology. Nature, 618(7965):616–624, 2023

    2023

  25. [25]

    From local explanations to global understanding with explainable ai for trees.Nature machine intelligence, 2(1):56–67, 2020

    Scott M Lundberg, Gabriel Erion, Hugh Chen, Alex DeGrave, Jordan M Prutkin, Bala Nair, Ronit Katz, Jonathan Himmelfarb, Nisha Bansal, and Su-In Lee. From local explanations to global understanding with explainable ai for trees.Nature machine intelligence, 2(1):56–67, 2020

    2020

  26. [26]

    A unified approach to interpreting model predictions.Advances in neural information processing systems, 30, 2017

    Scott M Lundberg and Su-In Lee. A unified approach to interpreting model predictions.Advances in neural information processing systems, 30, 2017

    2017

  27. [27]

    Mechanisms of cancer cell death induction by paclitaxel: an updated review.Apoptosis, 27(9):647– 667, 2022

    Shuang Zhao, Yufei Tang, Ruohan Wang, and Masoud Najafi. Mechanisms of cancer cell death induction by paclitaxel: an updated review.Apoptosis, 27(9):647– 667, 2022

    2022

  28. [28]

    Cancer hallmarks, biomarkers and breast cancer molecular subtypes.Journal of cancer, 7(10):1281, 2016

    Xiaofeng Dai, Liangjian Xiang, Ting Li, and Zhonghu Bai. Cancer hallmarks, biomarkers and breast cancer molecular subtypes.Journal of cancer, 7(10):1281, 2016

    2016

  29. [29]

    Gene set knowledge discovery with enrichr.Current protocols, 1(3):e90, 2021

    Zhuorui Xie, Allison Bailey, Maxim V Kuleshov, Daniel JB Clarke, John E Evan- gelista, Sherry L Jenkins, Alexander Lachmann, Megan L Wojciechowicz, Eryk Kropiwnicki, Kathleen M Jagodnik, et al. Gene set knowledge discovery with enrichr.Current protocols, 1(3):e90, 2021. 10 Predicting Therapeutic Outcome via Aligning Patient-Specific Knowledge Graph and Ge...

    2021

This paper was first reviewed by grok-4.5 on July 11, 2026.