Pith. sign in

REVIEW 5 major objections 5 minor 56 references

SOC-DGL: Social Interaction Behavior Inspired Dual Graph Learning Framework for Drug-Target Interaction Identification

T0 review · 5 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 Coupling affinity-driven global graph learning with a balance-theory even-polynomial filter that amplifies even-hop neighbors, plus an adjustable imbalance loss, yields state-of-the-art drug-target interaction prediction across four…

desk verdict A substantial DTI pipeline whose headline gains rest on a load-bearing algebraic error in the EDGL filter and on hyperparameters chosen with test-set peeking. read the letter →

arxiv 2506.01405 v2 pith:NYBLHHBI submitted 2025-06-02 cs.LG

Xiang Zhao , Ruijie Li , Qiao Ning , Shikai Guo , Hui Li , Qian Ma This is my paper
classification cs.LG
keywords drug-targetinteractionpredictiongraphneuralnetworksbalancetheoryeven-polynomialfilterheterogeneouslearningclassimbalancelossdrugrepurposingABL1casestudy
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

The paper sets out to show that drug-target interaction (DTI) prediction improves when a model reads both direct and indirect similarity from a heterogeneous drug-target graph, and proposes SOC-DGL to do so. SOC-DGL combines an affinity-driven module that learns global similarities from an affinity-enhanced network with an equilibrium-driven module that amplifies even-hop neighbors through an even-polynomial graph filter borrowed from social balance theory, then fuses the two views. On four benchmark datasets (KIBA, Davis, BindingDB, DrugBank), under both balanced and imbalanced sampling, the paper claims the framework reaches or surpasses seven state-of-the-art baselines on most metrics; its own tables flag imbalanced Davis as the one setting where a baseline (BCMMDA) scores higher on several metrics. The paper also argues the adjustable imbalance loss is the reason for stable gains when negative samples dominate. If these claims hold, the work offers a reusable two-view graph recipe for drug repurposing and for finding candidate interactions worth experimental testing.

What carries the argument

The load-bearing mechanism is the equilibrium-driven graph learning (EDGL) module's even-polynomial graph filter. After normalizing the bipartite adjacency $A' = [[0, A'_{DT}],[A'^T_{DT}, 0]]$ and setting the propagation matrix $P = D'^{-1/2} A' D'^{-1/2}$, the filter computes $H'' = \sum_{k=0}^{\lfloor K/2\rfloor} \alpha_k P^{2k} X$ with geometric weights $\alpha_k = \alpha(1-\alpha)^k$; retaining only even powers suppresses odd-length paths so that, on a bipartite graph, information returns to its own node type and encodes drug-drug and target-target similarity through shared interaction partners. The ADGL module supplies the global affinity-enhanced graph H built from multi-view drug and target affinity matrices, and the residual connection $\hat H = \omega H' + (1-\omega) H''$ fuses the two views. The adjustable imbalance loss $L = -\frac{1}{n_d n_t}(\varpi \frac{|y^-|}{|y^+|} \sum_{(i,j)\in y^+}\log h^*_{ij} + \sum_{(i,j)\in y^-}\log(1-h^*_{ij}))$ is the third piece of machinery, letting one parameter tune the class-weighting in imbalanced data.

What would settle it

Run the paper's Homo/Hetero/self-attention ablation on KIBA with all filters matched for total degree, number of terms, and $\alpha_k$ weighting, replacing $P^{2k}$ with $P^{2k+1}$ or with a random subset of powers of the same cardinality; if the even-only filter no longer wins on AUROC and AUPR, the balance-theory explanation is undercut. A second check is to measure triadic closure in the benchmark graphs: if the fraction of closed drug-target-drug triangles is statistically indistinguishable from random rewiring, the social-balance analogy is not supported by the data.

Watch

Extended reading notes

Core claim

On its own terms, the central discovery is that treating drug-target prediction as a bipartite friendship problem and applying balance-theory triadic stability — if two relationships in a triad are positive, the third tends to be positive — lets an even-polynomial graph filter extract higher-order homogeneous similarity. The paper constructs a symmetric affinity-enhanced network H from drug-drug, target-target, and drug-target affinity blocks, learns global embeddings with graph convolution in ADGL, then in EDGL zeroes out the homogeneous blocks and propagates features only through even powers of the normalized adjacency: $H'' = \sum_{k=0}^{\lfloor K/2\rfloor} \alpha_k P^{2k} X$, where $\alpha_k = \alpha(1-\alpha)^k$. Dropping odd-degree terms means a drug's signal reaches other drugs through shared targets, and a target's signal reaches other targets through shared drugs. A residual fusion combines the two embeddings and a matrix tri-factorization decoder predicts interaction probabilities, while a weighted binary cross-entropy loss with an adjustable parameter $\varpi$ handles imbalance. The paper reports gains over the compared methods on most metrics across the balanced versions of the four datasets and in most imbalanced settings, and shows that nine of the top ten predicted ABL1 drugs are experimentally confirmed binders.

Load-bearing premise

The load-bearing premise is that balance-theory triadic stability transfers from social networks to drug-target biology, namely that if two relationships in a drug-target-drug or target-drug-target triad are positive, the third tends to be positive, so even-hop neighbors in the bipartite graph carry the biologically meaningful similarity signal; the paper gives no biological or statistical evidence for this triadic closure.

Editorial extensions

If this is right

  • If the reported results hold, SOC-DGL reaches or exceeds the compared state-of-the-art methods on most metrics across the four benchmarks, with imbalanced Davis as the paper's own stated exception.
  • The ablation attributes the gain to the even-polynomial filter: replacing it with an odd-polynomial filter or with self-attention lowers performance on KIBA in both balanced and imbalanced settings.
  • The RLF loss with its adjustable parameter improves over standard, focal, and weighted cross-entropy losses on all four imbalanced datasets.
  • Cold-start experiments on KIBA indicate the model transfers to unseen drugs, with AUROC values clustering around 0.9, and to unseen targets with values around 0.6–0.8.
  • The ABL1 case study supports practical use: nine of the top ten ranked drugs are known binders, and the tenth receives supporting evidence from molecular docking and ligand-similarity analysis.

Reading between the lines

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

  • A natural extension is to test the same dual-module recipe on other bipartite biomedical association problems, such as miRNA-disease or drug-microbe prediction, where even-hop closure has the same intuitive reading.
  • The ablation does not fully isolate parity from depth: an even-only filter and an odd-only filter differ in effective propagation depth, so a matched-depth random-support filter would be needed to confirm that triadic closure, rather than depth or weighting, drives the gain.
  • A close reading of Eq. (10) suggests the adjustable parameter scales the positive-class term rather than directly downweighting negatives; if that reading is correct, its role is to moderate the positive upweighting of the weighted cross-entropy baseline.
  • The unconfirmed ABL1 prediction for DB04367 is an implicit wet-lab claim: a binding assay or high-throughput kinase screen could settle whether the predicted interaction is real.
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

5 major / 5 minor

Summary. The paper proposes SOC-DGL, a dual graph learning framework for drug-target interaction prediction. It consists of an Affinity-Driven Graph Learning (ADGL) module that builds an affinity-enhanced drug-target graph and applies graph convolution, and an Equilibrium-Driven Graph Learning (EDGL) module that uses an even-polynomial graph filter motivated by social balance theory to amplify even-hop neighbors. An adjustable imbalance loss is introduced for imbalanced datasets. The method is evaluated on KIBA, Davis, BindingDB, and DrugBank under balanced and imbalanced settings, with additional cold-start, ablation, hyperparameter, and ABL1 case studies. The paper claims that SOC-DGL consistently outperforms existing state-of-the-art methods across balanced and imbalanced scenarios.

Significance. If the claims were fully supported, SOC-DGL would be a practically useful contribution to DTI prediction, and the availability of source code and data is a strength. The case study on ABL1, including the prediction of a previously unconfirmed interaction and molecular docking evidence, is also a positive feature. However, the central performance claim is contradicted by the paper's own tables, the key EDGL derivation is algebraically invalid, and the evaluation protocol raises circularity concerns. As it stands, the evidence does not establish the claimed contribution.

major comments (5)
  1. [Abstract and Section IV.B, Table II] The abstract and Section IV.B state that SOC-DGL "consistently outperforms existing state-of-the-art methods across both balanced and imbalanced scenarios," but Table II contradicts this. On Davis imbalanced, SOC-DGL achieves AUPR 0.7655, below BCMMDA (0.7800) and MMDG-DTI (0.7787); on KIBA imbalanced, SOC-DGL achieves AUPR 0.7473, below BCMMDA (0.8040). The claim must be weakened to "competitive" or the comparisons must be redone, and the discussion in Section IV.B that acknowledges only the Davis case is incomplete because KIBA imbalanced AUPR also falls short of a compared method.
  2. [Section III.B.3, Eq. (7)-(8)] The derivation of the EDGL filter is mathematically inconsistent. The paper defines g(\tilde L)=Σα_k \tilde L^k and rewrites it as Σα_k (I−\tilde L)^k=Σα_k P^k with P=I−\tilde L, then claims that discarding odd-degree terms in g(\tilde L) yields Eq. (8), H''=Σ α_k P^{2k}X. This is invalid: P^{2k}=(I−\tilde L)^{2k} contains nonzero odd powers of \tilde L, so P^{2k} is not an even polynomial in the Laplacian. If the intended filter is an even-polynomial Laplacian filter, Eq. (8) should use \tilde L^{2k}; if P^{2k} is intended, the balance-theory motivation and the ablation claim in Fig. 3(A) that even-polynomial filtering of homogeneous information is responsible for the gain are not supported by the implemented formula.
  3. [Section IV.E and Section III.B.7] The hyperparameter tuning procedure appears to use the same evaluation folds that produce the headline result. Section IV.E.2 selects K=200 and alpha=0.20 for KIBA balanced because this combination gives AUROC 0.9761, and Section IV.E.3 selects varpi=0.2 for KIBA imbalanced, yet Section III.B.7 reports alpha=0.1 and varpi=0.6 for KIBA. Selecting hyperparameters on the test folds and then reporting the selected configuration's score is circular and likely inflates the reported metrics. The manuscript should either use a separate validation split or nested cross-validation and report the corresponding unbiased estimates.
  4. [Table II and Section IV.B] Table II reports no variance or confidence intervals for any baseline or for SOC-DGL, so the claimed improvements over methods such as MULGA and MMDG-DTI cannot be assessed for statistical significance. The paper's own Table I reports p-values for SOC-DGL folds, but no comparable significance information is given for comparisons against baselines. Baseline configurations, including their hyperparameters and negative sampling protocols, should be described precisely, and repeated runs with standard deviations should be reported.
  5. [Appendix D] The negative sampling procedure is not random and is not applied consistently across comparisons. Appendix D states that negative samples are chosen using drug similarity, treating drugs with low similarity to known positives as negatives. This can systematically alter AUROC and AUPR and may advantage the proposed method if baselines were evaluated with random negatives. The authors should justify this sampling choice, use the same sampling protocol for all methods, or repeat experiments with uniform random negatives to show that the reported ranking is not an artifact of the sampling scheme.
minor comments (5)
  1. [Section IV.B] The text says ten-fold cross-validation was conducted on "four balanced and two imbalanced datasets," but Table I reports both balanced and imbalanced rows for all four datasets, and Table II lists imbalanced results for all four datasets. This wording should be corrected.
  2. [Section IV.E.3 and Fig. 4 caption] The caption of Fig. 4 labels the hyperparameter tuning of varpi as part (E), but the text refers to it as Fig. 4(F); the figure part labels should be made consistent.
  3. [Section VI.B.1] The multi-view learning appendix introduces several variables (A, C1, C2, C3, lambda1-4, mu, beta1, beta2) but does not state the objective function being minimized; adding the optimization problem would make the appendix self-contained and easier to verify.
  4. [Section II.B] The phrase "this study propose the EDGL module" should be "this study proposes the EDGL module".
  5. [Section IV.F] The case study narrative says SOC-DGL predicts the top 9 drugs known to bind ABL1 and analyzes the 10th, but the text around Table IV alternates between "DB01254" and "DB04367" in the explanation of higher-order channels; this should be checked for consistency.

Circularity Check

2 steps flagged · score 6.0 of 10

The flagship KIBA AUROC of 0.9761 is the argmax of a (K, α) grid search run on the same evaluation folds that produce the headline tables, so the central 'prediction' reduces to a fitted maximum; the EDGL balance-theory weighting (α_k = α(1−α)^k) and the imbalance weight ϖ are fold-tuned parameters presented as derived results.

  1. fitted input called prediction [Section IV.E.2 (Optimization of EDGL Hyperparameters) and Table II; filter definition in Section III.B.3]
    "A range of values for K (100 to 500) and α (0.05 to 0.25) were tested to evaluate their impact on SOC-DGL performance. As shown in Fig. 4 (E), optimal performance was achieved with K=200 and α=0.20, resulting in an AUROC of 0.9761. [...] SOC-DGL \ 0.9761 0.9805 0.9350 0.9268 0.9373 0.9133 0.9328. [...] α_k is the learnable weight matrix, and α_k = α×(1−α)^k, with α being a hyperparameter that can be set."

    The headline balanced-KIBA AUROC reported in Table II is exactly the number selected in Fig. 4(E) by grid-searching K and α on the same 10-fold evaluation that produces the headline table. Because the EDGL filter weights are fixed by α_k = α(1−α)^k, the higher-order 'even-polynomial' similarity signal claimed to embody balance theory is a tuned geometric weighting, so the reported performance reduces to the fitted grid maximum by construction. Section III.B.7 lists α=0.1 for the same balanced KIBA setting, confirming that 0.9761 is the tuning run's argmax rather than an independent evaluation of the described configuration. The abstract's claim that SOC-DGL 'consistently outperforms existing state-of-the-art methods' rests on this fold-selected number.

  2. fitted input called prediction [Section IV.E.3 (Optimization of Adjustable Parameter ϖ) vs. Section III.B.7 (Parameter Setting) and Table III]
    "Hyperparameter experiments with ϖ∈{0.1,0.2,0.3,0.4,0.5,0.6} were conducted, and the results in Fig. 4 (F) demonstrated that ϖ=0.2 achieved optimal performance. [...] Under the imbalanced setting, an additional imbalance loss parameter ϖ, used to adjust the weight of negative samples, was set to 0.6."

    The imbalanced KIBA AUROC 0.9364 in Table II is numerically identical to the RLF row in Table III, i.e., it is produced with the adjustable negative-sample weight ϖ that the paper selects on the same evaluation folds, and the paper contradicts itself on the chosen value (0.2 'optimal' in the tuning experiment vs. 0.6 in the parameter-setting section). The claimed advantage of the new RLF loss over SLF/WLF on imbalanced benchmarks is therefore a report of the fold-selected best ϖ, an optimized fit presented as a demonstrated property of the proposed loss function rather than an independent prediction.

full rationale

The genuine circularity is pattern 2 (fitted input called prediction), localized to the headline numbers. The flagship balanced-KIBA AUROC 0.9761 that anchors the abstract's 'consistently outperforms' claim is, by the paper's own account, the argmax of a (K, α) grid search on the same 10-fold evaluation that populates Table II; since the EDGL coefficients are a closed-form function α_k = α(1−α)^k of the tuned α, the balance-theory 'even-polynomial' mechanism reduces to a fitted weighting, and the improvement attributed to it in the Fig. 3(A) ablation inherits that fit. The imbalanced results show the same pattern for ϖ (fold-tuned, with contradictory reported values 0.2 vs 0.6), and Table III's RLF row equals Table II's SOC-DGL imbalanced row exactly. I deliberately do not count several adjacent concerns as circularity: the algebraic invalidity of Eq. (8) (P^{2k} = (I−L)^{2k} contains odd powers of L, so 'discarding odd-degree terms in g(L)' is false as stated) is a correctness risk, not a self-referential reduction; the unvalidated transfer of balance-theory triadic stability to drug-target biology is a burden-of-proof/interpretability issue; the self-citations ([3], [11]) are background and not load-bearing; and the ABL1 case study (9 known hits in the top 10, plus docking and SEA confirmations) is an externally checkable holdout. Because the model is still compared against external baselines and tested in cold-start holdouts, the paper retains independent content, but the flagship performance claims reduce by construction to tuned hyperparameters, giving a partial-circularity score of 6 rather than 0 or 10.

Assumptions & free parameters 7 free parameters · 4 assumptions · 0 invented entities

The framework rests on a transfer of social balance theory to molecular biology, on standard GNN machinery, and on a non-random negative sampling scheme. The free parameters are tuned on the evaluation folds that produce the headline numbers, which limits the independence of the reported gains.

free parameters (7)
  • Affinity binarization threshold = 0.8
    Section III.B.2 sets elements of ADD and ATT greater than or equal to 0.8 to 1 and the rest to 0. The threshold is chosen without sensitivity analysis.
  • EDGL propagation depth K = 200 for KIBA; other values on GitHub
    Tuned in Section IV.E.2 on the KIBA balanced dataset; the final reported AUROC 0.9761 uses K=200.
  • EDGL neighbor contribution ratio alpha = 0.1 balanced, 0.2 as optimal in tuning
    Section III.B.7 sets alpha=0.1 under the balanced setting, while Section IV.E.2 reports alpha=0.20 as optimal on KIBA. The value controls alpha_k in the graph filter.
  • Number of GCN layers N = 2
    Section IV.E.1 selects a two-layer GCN based on experiments; N=1 was slightly better on KIBA in AUROC and AUPR, so the choice is data-dependent.
  • Residual connection weight omega = 0.8
    Section III.B.7 sets the residual connection weight w to 0.8, used in Eq. (21) to fuse ADGL and EDGL outputs.
  • Imbalance loss parameter beta (written as varpi in the paper) = 0.6 in Section III.B.7, 0.2 as optimal in Fig. 4(F)
    The RLF loss weight is set to 0.6 for imbalanced KIBA in the parameter setting, but the tuning experiment in Section IV.E.3 reports 0.2 as optimal. These settings conflict.
  • Multi-view learning hyperparameters = not reported
    Appendix B(1) introduces mu, beta1, beta2, lambda, mu_max, and rho for the affinity matrix construction, but concrete values are not given in the text.
assumptions (4)
  • domain assumption Balance-theory triadic stability applies to drug-target relationships
    Section III states that if two relationships in a triad are positive, the third tends to be positive, and analogizes social friendships to drug-target associations. This justifies the even-hop filter in EDGL.
  • domain assumption Drug structural similarity implies functional similarity, and target similarity reflects functional conservation
    Stated in Section III as the basis for building drug and target affinity matrices. This is a standard DTI assumption but is not independently validated in this paper.
  • domain assumption Negatively sampled pairs constructed from low-similarity drugs are reliable non-interactions
    Appendix D describes selecting negatives using drug similarity: drugs with low similarity to known positives are treated as negative samples. If some of these pairs actually interact, the reported AUROC and AUPR are biased.
  • domain assumption Even-hop neighbor amplification improves DTI graph learning as it does in general GNN robustness
    The even-polynomial graph filter in Eq. (8) is taken from EvenNet [8], where it is shown to improve GNN robustness. The paper assumes the same transfer holds for drug-target graphs without a dedicated biological or statistical justification.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SOC-DGL: Social Interaction Behavior Inspired Dual Graph Learning Framework for Drug-Target Interaction Identification." pith.science (2026). https://pith.science/paper/NYBLHHBI

@misc{pith2026250601405,
  author       = {Pith},
  title        = {Pith review of: SOC-DGL: Social Interaction Behavior Inspired Dual Graph Learning Framework for Drug-Target Interaction Identification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NYBLHHBI}},
  note         = {Machine review of arXiv:2506.01405}
}
read the original abstract

The identification of drug-target interactions (DTI) is critical for drug discovery and repositioning, as it reveals potential therapeutic uses of existing drugs, accelerating development and reducing costs. However, most existing models focus only on direct similarity in homogeneous graphs, failing to exploit the rich similarity in heterogeneous graphs. To address this gap, inspired by real-world social interaction behaviors, we propose SOC-DGL, which comprises two specialized modules: the Affinity-Driven Graph Learning (ADGL) module, learning global similarity through an affinity-enhanced drug-target graph, and the Equilibrium-Driven Graph Learning (EDGL) module, capturing higher-order similarity by amplifying the influence of even-hop neighbors using an even-polynomial graph filter based on balance theory. This dual approach enables SOC-DGL to effectively capture similarity information across multiple interaction scales within affinity and association matrices. To address the issue of imbalance in DTI datasets, we propose an adjustable imbalance loss function that adjusts the weight of negative samples by the parameter. Extensive experiments on four benchmark datasets demonstrate that SOC-DGL consistently outperforms existing state-of-the-art methods across both balanced and imbalanced scenarios. Moreover, SOC-DGL successfully predicts the top 9 drugs known to bind ABL1, and further analyzed the 10th drug, which has not been experimentally confirmed to interact with ABL1, providing supporting evidence for its potential binding.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

56 extracted references · 53 canonical work pages

  1. [1]

    Drug repo- sitioning: a brief overview,

    J.-P. Jourdan, R. Bureau, C. Rochais, and P. Dallemagne, “Drug repo- sitioning: a brief overview,”Journal of Pharmacy and Pharmacology, vol. 72, no. 9, pp. 1145–1151, 2020

    work page 2020

  2. [2]

    Drug repurposing for antimicrobial discovery,

    M. A. Farha and E. D. Brown, “Drug repurposing for antimicrobial discovery,”Nature microbiology, vol. 4, no. 4, pp. 565–577, 2019

    work page 2019

  3. [3]

    A Multi-view Divergence-Convergence Feature Augmentation Framework for Drug-related Microbes Prediction

    X. An, R. Li, Q. Ning, S. Guo, H. Li, and Q. Ma, “A multi- view divergence-convergence feature augmentation framework for drug- related microbes prediction,”arXiv preprint arXiv:2506.18797, 2025

    work page Pith review arXiv 2025

  4. [4]

    Msi- dti: predicting drug-target interaction based on multi-source information and multi-head self-attention,

    W. Zhao, Y . Yu, G. Liu, Y . Liang, D. Xu, X. Feng, and R. Guan, “Msi- dti: predicting drug-target interaction based on multi-source information and multi-head self-attention,”Briefings in bioinformatics, vol. 25, no. 3, p. bbae238, 2024

    work page 2024

  5. [5]

    Artificial neural networks train- ing algorithm integrating invasive weed optimization with differential evolutionary model,

    A. A. Movassagh, J. A. Alzubi, M. Gheisari, M. Rahimi, S. Mohan, A. A. Abbasi, and N. Nabipour, “Artificial neural networks train- ing algorithm integrating invasive weed optimization with differential evolutionary model,”Journal of Ambient Intelligence and Humanized Computing, pp. 1–9, 2023

    work page 2023

  6. [6]

    Auto-metric graph neural network for attack detection on iot-based smart environment and secure data transmission using advanced wild horse standard encryption method,

    R. Yadawad, U. P. Kulkarni, and J. A. Alzubi, “Auto-metric graph neural network for attack detection on iot-based smart environment and secure data transmission using advanced wild horse standard encryption method,”Int. J. Comput. Netw. Inf. Secur., vol. 16, no. 3, pp. 1–15, 2024

    work page 2024

  7. [7]

    A multimodal human-computer interaction for smart learning system,

    T. M. Alzubi, J. A. Alzubi, A. Singh, O. A. Alzubi, and M. Subramanian, “A multimodal human-computer interaction for smart learning system,” International Journal of Human–Computer Interaction, vol. 41, no. 3, pp. 1718–1728, 2025

    work page 2025

  8. [8]

    Evennet: Ignoring odd-hop neighbors improves robustness of graph neural networks,

    R. Lei, Z. Wang, Y . Li, B. Ding, and Z. Wei, “Evennet: Ignoring odd-hop neighbors improves robustness of graph neural networks,”Advances in Neural Information Processing Systems, vol. 35, pp. 4694–4706, 2022

    work page 2022

Show all 56 references
  1. [9]

    Predicting emerging drug interactions using gnns,

    N. Q. K. Le, “Predicting emerging drug interactions using gnns,”Nature Computational Science, vol. 3, no. 12, pp. 1007–1008, 2023

    2023

  2. [10]

    Improved prediction model of protein and peptide toxicity by integrating channel attention into a convolutional neural network and gated recurrent units,

    Z. Zhao, J. Gui, A. Yao, N. Q. K. Le, and M. C. H. Chua, “Improved prediction model of protein and peptide toxicity by integrating channel attention into a convolutional neural network and gated recurrent units,” ACS omega, vol. 7, no. 44, pp. 40 569–40 577, 2022

    2022

  3. [11]

    Mhmda:“similarity- association-similarity

    R. Li, Q. Ning, Y . aomiao Zhao, S. Guo, and H. Li, “Mhmda:“similarity- association-similarity” metapaths and heterogeneous-hyper network learning for mirna-disease association prediction,”IEEE Transactions on Computational Biology and Bioinformatics, 2025

    2025

  4. [12]

    Mmdg-dti: Drug– target interaction prediction via multimodal feature fusion and domain generalization,

    Y . Hua, Z. Feng, X. Song, X.-J. Wu, and J. Kittler, “Mmdg-dti: Drug– target interaction prediction via multimodal feature fusion and domain generalization,”Pattern Recognition, vol. 157, p. 110887, 2025

    2025

  5. [13]

    A review of machine learning-based methods for predicting drug–target interactions,

    W. Shi, H. Yang, L. Xie, X.-X. Yin, and Y . Zhang, “A review of machine learning-based methods for predicting drug–target interactions,”Health Information Science and Systems, vol. 12, no. 1, p. 30, 2024

    2024

  6. [14]

    Dtigems+: drug–target interaction prediction using graph embedding, graph mining, and similarity-based techniques,

    M. A. Thafar, R. S. Olayan, H. Ashoor, S. Albaradei, V . B. Bajic, X. Gao, T. Gojobori, and M. Essack, “Dtigems+: drug–target interaction prediction using graph embedding, graph mining, and similarity-based techniques,”Journal of Cheminformatics, vol. 12, pp. 1–17, 2020

    2020

  7. [15]

    ingnn-dti: prediction of drug– target interaction with interpretable nested graph neural network and pretrained molecule models,

    Y . Sun, Y . Y . Li, C. K. Leung, and P. Hu, “ingnn-dti: prediction of drug– target interaction with interpretable nested graph neural network and pretrained molecule models,”Bioinformatics, vol. 40, no. 3, p. btae135, 2024

    2024

  8. [16]

    Dti- heta: prediction of drug–target interactions based on gcn and gat on heterogeneous graph,

    K. Shao, Y . Zhang, Y . Wen, Z. Zhang, S. He, and X. Bo, “Dti- heta: prediction of drug–target interactions based on gcn and gat on heterogeneous graph,”Briefings in Bioinformatics, vol. 23, no. 3, p. bbac109, 2022

    2022

  9. [17]

    Giae-dti: Predicting drug-target interactions based on heterogeneous network and gin-based graph autoencoder,

    M. Wang, X. Lei, L. Liu, J. Chen, and F.-X. Wu, “Giae-dti: Predicting drug-target interactions based on heterogeneous network and gin-based graph autoencoder,”IEEE Journal of Biomedical and Health Informat- ics, pp. 1–14, 2024

    2024

  10. [18]

    Multi-dimensional clustering through fusion of high-order similarities,

    H. Peng, H. Wang, Y . Hu, W. Zhou, and H. Cai, “Multi-dimensional clustering through fusion of high-order similarities,”Pattern Recogni- tion, vol. 121, p. 108108, 2022

    2022

  11. [19]

    Unsupervised feature selection with high-order similarity learning,

    Y . Mi, H. Chen, C. Luo, S.-J. Horng, and T. Li, “Unsupervised feature selection with high-order similarity learning,”Knowledge-Based Systems, vol. 285, p. 111317, 2024

    2024

  12. [20]

    Disco: graph-based disentangled contrastive learning for cold-start cross-domain recommendation,

    H. Li, Y . Wang, Z. Xiao, J. Yang, C. Zhou, M. Zhang, and W. Ju, “Disco: graph-based disentangled contrastive learning for cold-start cross-domain recommendation,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 39, no. 11, 2025, pp. 12 049–12 057. XIANG ...

    2025

  13. [21]

    schntl: single-cell rna-seq data cluster- ing augmented by high-order neighbors and triplet loss,

    H. Meng, C. Qin, and Z. Long, “schntl: single-cell rna-seq data cluster- ing augmented by high-order neighbors and triplet loss,”Bioinformatics, p. btaf044, 2025

    2025

  14. [22]

    High-order similarity learning based domain adaptation for speech emotion recognition,

    H. Wang, Y . Ji, P. Song, and Z. Liu, “High-order similarity learning based domain adaptation for speech emotion recognition,”Applied Acoustics, vol. 231, p. 110555, 2025

    2025

  15. [23]

    Structural balance: a generalization of heider’s theory

    D. Cartwright and F. Harary, “Structural balance: a generalization of heider’s theory.”Psychological review, vol. 63, no. 5, p. 277, 1956

    1956

  16. [24]

    Similarity-based machine learning methods for predicting drug–target interactions: a brief review,

    H. Ding, I. Takigawa, H. Mamitsuka, and S. Zhu, “Similarity-based machine learning methods for predicting drug–target interactions: a brief review,”Briefings in bioinformatics, vol. 15, no. 5, pp. 734–747, 2014

    2014

  17. [25]

    Prediction of membrane protein types based on the hydrophobic index of amino acids,

    Z.-P. Feng and C.-T. Zhang, “Prediction of membrane protein types based on the hydrophobic index of amino acids,”Journal of protein chemistry, vol. 19, pp. 269–275, 2000

    2000

  18. [26]

    Drug target identification using side-effect similarity,

    M. Campillos, M. Kuhn, A.-C. Gavin, L. J. Jensen, and P. Bork, “Drug target identification using side-effect similarity,”Science, vol. 321, no. 5886, pp. 263–266, 2008

    2008

  19. [27]

    Making sense of large-scale kinase inhibitor bioactivity data sets: a comparative and integrative analysis,

    J. Tang, A. Szwajda, S. Shakyawar, T. Xu, P. Hintsanen, K. Wenner- berg, and T. Aittokallio, “Making sense of large-scale kinase inhibitor bioactivity data sets: a comparative and integrative analysis,”Journal of Chemical Information and Modeling, vol. 54, no. 3, pp. 735–743, 2014

    2014

  20. [28]

    Comprehensive analysis of kinase inhibitor selectivity,

    M. I. Davis, J. P. Hunt, S. Herrgard, P. Ciceri, L. M. Wodicka, G. Pal- lares, M. Hocker, D. K. Treiber, and P. P. Zarrinkar, “Comprehensive analysis of kinase inhibitor selectivity,”Nature biotechnology, vol. 29, no. 11, pp. 1046–1051, 2011

    2011

  21. [29]

    Bindingdb: a web-accessible database of experimentally determined protein–ligand binding affinities,

    T. Liu, Y . Lin, X. Wen, R. N. Jorissen, and M. K. Gilson, “Bindingdb: a web-accessible database of experimentally determined protein–ligand binding affinities,”Nucleic acids research, vol. 35, no. suppl 1, pp. D198–D201, 2007

    2007

  22. [30]

    Drugbank 5.0: a major update to the drugbank database for 2018,

    D. S. Wishart, Y . D. Feunang, A. C. Guo, E. J. Lo, A. Marcu, J. R. Grant, T. Sajed, D. Johnson, C. Li, Z. Sayeedaet al., “Drugbank 5.0: a major update to the drugbank database for 2018,”Nucleic acids research, vol. 46, no. D1, pp. D1074–D1082, 2018

    2018

  23. [31]

    Reopti- mization of mdl keys for use in drug discovery,

    J. L. Durant, B. A. Leland, D. R. Henry, and J. G. Nourse, “Reopti- mization of mdl keys for use in drug discovery,”Journal of chemical information and computer sciences, vol. 42, no. 6, pp. 1273–1280, 2002

    2002

  24. [32]

    Atom pairs as molecular features in structure-activity studies: definition and applica- tions,

    R. E. Carhart, D. H. Smith, and R. Venkataraghavan, “Atom pairs as molecular features in structure-activity studies: definition and applica- tions,”Journal of Chemical Information and Computer Sciences, vol. 25, no. 2, pp. 64–73, 1985

    1985

  25. [33]

    The generation of a unique machine description for chemical structures-a technique developed at chemical abstracts service

    H. L. Morgan, “The generation of a unique machine description for chemical structures-a technique developed at chemical abstracts service.” Journal of chemical documentation, vol. 5, no. 2, pp. 107–113, 1965

    1965

  26. [34]

    Enzyme family classification by support vector machines,

    C. Cai, L. Han, Z. Ji, and Y . Chen, “Enzyme family classification by support vector machines,”Proteins: Structure, Function, and Bioinfor- matics, vol. 55, no. 1, pp. 66–76, 2004

    2004

  27. [35]

    Using amphiphilic pseudo amino acid composition to predict enzyme subfamily classes,

    K.-C. Chou, “Using amphiphilic pseudo amino acid composition to predict enzyme subfamily classes,”Bioinformatics, vol. 21, no. 1, pp. 10–19, 2005

    2005

  28. [36]

    Graphdta: predicting drug–target binding affinity with graph neural networks,

    T. Nguyen, H. Le, T. P. Quinn, T. Nguyen, T. D. Le, and S. Venkatesh, “Graphdta: predicting drug–target binding affinity with graph neural networks,”Bioinformatics, vol. 37, no. 8, pp. 1140–1147, 2021

    2021

  29. [37]

    De novo prediction of drug–target interactions using laplacian regularized schatten p-norm minimization,

    G. Wu, M. Yang, Y . Li, and J. Wang, “De novo prediction of drug–target interactions using laplacian regularized schatten p-norm minimization,” Journal of Computational Biology, vol. 28, no. 7, pp. 660–673, 2021

    2021

  30. [38]

    Drug repositioning based on multi-view learning with matrix completion,

    Y . Yan, M. Yang, H. Zhao, G. Duan, X. Peng, and J. Wang, “Drug repositioning based on multi-view learning with matrix completion,” Briefings in Bioinformatics, vol. 23, no. 3, p. bbac054, 2022

    2022

  31. [39]

    Mulga, a unified multi-view graph autoencoder-based ap- proach for identifying drug–protein interaction and drug repositioning,

    J. Ma, C. Li, Y . Zhang, Z. Wang, S. Li, Y . Guo, L. Zhang, H. Liu, X. Gao, and J. Song, “Mulga, a unified multi-view graph autoencoder-based ap- proach for identifying drug–protein interaction and drug repositioning,” Bioinformatics, vol. 39, no. 9, p. btad524, 2023

    2023

  32. [40]

    Predict- ing drug–protein interactions through branch-chain mining and multi- dimensional attention network,

    Z. Huang, Q. Xiao, T. Xiong, W. Shi, Y . Yang, and G. Li, “Predict- ing drug–protein interactions through branch-chain mining and multi- dimensional attention network,”Computers in Biology and Medicine, vol. 171, p. 108127, 2024

    2024

  33. [41]

    American cancer society,

    M. Amin, “American cancer society,”AJCC cancer staging manual. Eight edition/editor-in-chief, Mahul B. Amin, MD, FCAP, vol. 1024, 2017

    2017

  34. [42]

    Chronic myeloid leukemia: 2018 update on diagnosis, therapy and monitoring,

    E. Jabbour and H. Kantarjian, “Chronic myeloid leukemia: 2018 update on diagnosis, therapy and monitoring,”American journal of hematology, vol. 93, no. 3, pp. 442–459, 2018

    2018

  35. [43]

    Protein tyrosine kinases: their roles and their targeting in leukemia,

    K. K. Bhanumathy, A. Balagopal, F. S. Vizeacoumar, F. J. Vizeacoumar, A. Freywald, and V . Giambra, “Protein tyrosine kinases: their roles and their targeting in leukemia,”Cancers, vol. 13, no. 2, p. 184, 2021

    2021

  36. [44]

    Cellular localization of debromohymenialdisine and hymenialdisine in the marine sponge axinella sp. using a newly developed cell purification protocol,

    Y .-F. Song, Y . Qu, X.-P. Cao, and W. Zhang, “Cellular localization of debromohymenialdisine and hymenialdisine in the marine sponge axinella sp. using a newly developed cell purification protocol,”Marine biotechnology, vol. 13, pp. 868–882, 2011

    2011

  37. [45]

    In vitro pharmacological profiling of r406 identifies molecular targets underlying the clinical effects of fostamatinib,

    M. G. Rolf, J. O. Curwen, M. Veldman-Jones, C. Eberlein, J. Wang, A. Harmer, C. J. Hellawell, and M. Braddock, “In vitro pharmacological profiling of r406 identifies molecular targets underlying the clinical effects of fostamatinib,”Pharmacology research & perspectives, vol. 3...

    2015

  38. [46]

    Drugs, their targets and the nature and number of drug targets,

    P. Imming, C. Sinning, and A. Meyer, “Drugs, their targets and the nature and number of drug targets,”Nature reviews Drug discovery, vol. 5, no. 10, pp. 821–834, 2006

    2006

  39. [47]

    The protein data bank,

    H. M. Berman, J. Westbrook, Z. Feng, G. Gilliland, T. N. Bhat, H. Weissig, I. N. Shindyalov, and P. E. Bourne, “The protein data bank,” Nucleic acids research, vol. 28, no. 1, pp. 235–242, 2000

    2000

  40. [48]

    Brigatinib: first global approval,

    A. Markham, “Brigatinib: first global approval,”Drugs, vol. 77, pp. 1131–1135, 2017

    2017

  41. [49]

    Haberler, E

    C. Haberler, E. Gelpi, C. Marosi, K. R ¨ossler, P. Birner, H. Budka, and J. Hainfellner, “Immunohistochemical analysis of platelet-derived growth factor receptor-α,-β, c-kit, c-abl, and arg proteins in glioblas- toma: Possible implications for patient selection for imatinib me...

    2006

  42. [50]

    Tyrosine kinase inhibitors for the treatment of philadelphia chromosome-positive adult acute lym- phoblastic leukemia,

    P. P. Piccaluga, S. Paolini, and G. Martinelli, “Tyrosine kinase inhibitors for the treatment of philadelphia chromosome-positive adult acute lym- phoblastic leukemia,”Cancer, vol. 110, no. 6, pp. 1178–1186, 2007

    2007

  43. [51]

    Relating protein pharmacology by ligand chemistry,

    M. J. Keiser, B. L. Roth, B. N. Armbruster, P. Ernsberger, J. J. Irwin, and B. K. Shoichet, “Relating protein pharmacology by ligand chemistry,” Nature biotechnology, vol. 25, no. 2, pp. 197–206, 2007

    2007

  44. [52]

    The ewing sarcoma protein regulates dna damage-induced alternative splicing,

    M. P. Paronetto, B. Mi ˜nana, and J. Valc ´arcel, “The ewing sarcoma protein regulates dna damage-induced alternative splicing,”Molecular cell, vol. 43, no. 3, pp. 353–368, 2011

    2011

  45. [53]

    Expression and prognostic value of rad51 in adenocarcinoma at the gastroesophageal junction,

    D. Redati, X. Yang, C. Lei, L. Liu, L. Ge, and H. Wang, “Expression and prognostic value of rad51 in adenocarcinoma at the gastroesophageal junction,”Iranian Journal of Public Health, vol. 51, no. 10, p. 2231, 2022

    2022

  46. [54]

    Suppression of usp7 induces bcr- abl degradation and chronic myelogenous leukemia cell apoptosis,

    S. Jiang, X. Wang, Y . He, H. Huang, B. Cao, Z. Zhang, J. Liu, Q. Wang, Z. Huang, and X. Mao, “Suppression of usp7 induces bcr- abl degradation and chronic myelogenous leukemia cell apoptosis,”Cell Death & Disease, vol. 12, no. 5, p. 456, 2021

    2021

  47. [55]

    Artificial intelligence foundation for therapeutic science,

    K. Huang, T. Fu, W. Gao, Y . Zhao, Y . Roohani, J. Leskovec, C. W. Coley, C. Xiao, J. Sun, and M. Zitnik, “Artificial intelligence foundation for therapeutic science,”Nature chemical biology, vol. 18, no. 10, pp. 1033–1036, 2022

    2022

  48. [56]

    Rdkit: open-source cheminformatics from machine learn- ing to chemical registration,

    G. Landrum, “Rdkit: open-source cheminformatics from machine learn- ing to chemical registration,” inAbstracts of Papers of the American Chemical Society, vol. 258. AMER CHEMICAL SOC 1155 16TH ST, NW, W ASHINGTON, DC 20036 USA, 2019

    2019

Pith tools

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