Pith. sign in

REVIEW 3 major objections 5 minor 71 references

Structure-Aware Compound-Protein Affinity Prediction via Graph Neural Networks with Group Lasso Regularization

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Training graph neural networks on both scaffold and substituent substructures of activity-cliff pairs, with group lasso or sparse group lasso penalties on the prediction heads, improves per-target pIC50 prediction and feature attribution…

desk verdict Useful combination of AC-aware losses and group lasso for per-kinase pIC50 prediction, but the evaluation protocol allows the same compound in train and test, so the headline gains are not yet convincing. read the letter →

arxiv 2507.03318 v3 pith:QH66CIWV submitted 2025-07-04 cs.LG cs.AI

Zanyu Shi , Yang Wang , Pathum M. Weerawarna , Jie Zhang , Timothy I. Richardson , Yijie Wang , Kun Huang This is my paper · ORCID
classification cs.LGcs.AI
keywords activitycliffscompound-proteinaffinitypredictiongraphneuralnetworksgrouplassosparsefeatureattributionmatchedmolecularpairstyrosinekinases
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 argues that a graph neural network trained on matched-molecular-pair activity cliffs can predict per-target compound–protein affinity more accurately and explainably when it is forced to use both the conserved scaffold atoms and the variable substituent atoms of each pair, and when the loss is penalized with group lasso or sparse group lasso. On six tyrosine kinase targets, this setup lowers root mean square error and raises Pearson correlation relative to a loss that attends only to the uncommon (substituent) atoms, and relative to four of the five comparison baselines. The same regularization makes the models' feature attributions more stable and more concordant with maximum-common-substructure-based ground truth, at both the graph level and the atom level. If the claim holds, the practical payoff is a drug-discovery tool that localizes the moieties driving potency differences during lead optimization on targets where training data are scarce.

What carries the argument

The load-bearing mechanism is a two-branch masked loss with group-level sparsity. After message passing, binary masks $f^{(CN)}$ and $f^{(UCN)}$ split each compound's node embeddings into scaffold and decoration subgraphs; two parallel MLP heads, with weight vectors $\beta^{(CN)}$ and $\beta^{(UCN)}$, each predict the pair's activity difference from its own branch, and the node loss $L_N = L_{CN} + L_{UCN}$ is minimized together with the MSE on absolute pIC50. Group lasso penalizes the $\ell^2$ norms of both head weight vectors so entire branches are driven to zero together; sparse group lasso adds an $\ell^1$ penalty blended by a convex combination $\alpha \in [0,1]$ so individual weights within a surviving branch can also vanish. The sparsity pattern learned by these heads is what later concentrates feature attribution (CAM, Grad-CAM, Gradient × Input, integrated gradients) on the uncommon decoration atoms, yielding the paper's higher global direction scores and better atom-coloring agreement with the MCS-based labels.

What would settle it

Run the same six-kinase benchmark with scaffold-aware splits that forbid any test compound from sharing a Bemis–Murcko scaffold (or the 50% MCS threshold) with a training compound, and compare $L_N$+GL and $L_N$+SGL against $L_{UCN}$ and against ACtriplet. If the RMSE and PCC advantages vanish or fall below significance under these splits, the claim that common-node integration plus sparsity improves generalization fails, and the reported gains would be attributable to near-duplicate analog leakage rather than structure-aware learning.

Watch

Extended reading notes

Core claim

The paper's central claim is that training a message-passing graph neural network on matched-molecular-pair activity cliffs with a structure-aware, sparsity-penalized loss produces more accurate and more explainable target-specific affinity prediction than attending to the variable substituent atoms alone. Each pair's maximum common substructure (MCS) defines two complementary subgraphs — common scaffold nodes and uncommon decoration nodes — and the node loss $L_N = L_{CN} + L_{UCN}$ makes separate prediction heads for the two branches reproduce the observed activity difference $\Delta y_k$ between the paired compounds, alongside the MSE on absolute pIC50. Group lasso ($\lambda(\sqrt{p^{(CN)}}\|\beta^{(CN)}\|_2 + \sqrt{p^{(UCN)}}\|\beta^{(UCN)}\|_2)$) prunes whole branches, while sparse group lasso adds an $\ell^1$ term so individual atoms inside a kept branch can also be zeroed. In the 5-fold cross-validation over six tyrosine kinases (Src-family 1O42, 2H8H, 4MXO; ABL1 2E2B; ALK 3AOX; BTK 3OCS), the paper's named variants SAGGLR+GL and SAGGLR+SGL (NNConv + mean pooling with $L_N$+GL and $L_N$+SGL) report averaged RMSE 0.2665 and 0.2670 with averaged PCC 0.9545 and 0.9519, while the full architecture-by-loss sweep shows its lowest averaged RMSE (0.2427) and highest averaged PCC (0.9618) under GIN + sum with sparse group lasso. The same regularized losses raise graph-level global direction scores and atom-level coloring accuracy against MCS-derived ground truth, which the paper presents as evidence that the model highlights the substructures that actually drive potency differences.

Load-bearing premise

The load-bearing premise, conceded in the paper's own Discussion, is that the random compound-level split keeps highly similar compounds from straddling the training and test sets; if a scaffold- or MCS-aware split were used instead, the reported RMSE and PCC gains — including the already non-significant margin over ACtriplet — could shrink substantially or disappear.

Editorial extensions

If this is right

  • Adding common-scaffold information to the loss, and then group lasso or sparse group lasso penalties, improves RMSE and PCC significantly over the uncommon-node-only loss across all three backbones (NNConv, GAT, GIN), with paired comparisons at P < 0.001 for the headline settings.
  • Against the five external baselines, the SAGGLR variants beat four of them significantly in both RMSE and PCC; the fifth, ACtriplet, is not statistically separable despite a higher mean RMSE (0.3042 versus 0.2665).
  • Graph-level global direction scores rise under group-lasso training for all four attribution methods tested on the Src-family test pairs — up 47.13% for CAM, 14.83% for Grad-CAM, 15.4% for Gradient × Input, and 8.49% for integrated gradients — with Wilcoxon P values below 0.01 for each.
  • Atom-level coloring accuracy, F1, Spearman rank correlation, and attribution AUROC (especially for atoms in the lower-activity compound) improve under the regularized losses, and attribution rankings stay more stable when edges are randomly dropped.
  • The paper states the same structure-aware regularization idea transfers to other GNN chemistry tasks where key-substructure identification matters, including drug synergy prediction, drug–drug interaction, and material property prediction.

Reading between the lines

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

  • Because the ACtriplet margin is not statistically significant, the value of the sparsity penalty itself remains open: an ablation that adds group lasso to the uncommon-only loss $L_{UCN}$ without the common-node branch would isolate whether the gain comes from regularization or from the scaffold branch.
  • The MCS-derived atom labels are a heuristic proxy for causal attribution, so the improvements should be read as 'more consistent with the proxy' rather than proven chemical causality; a prospective assay that modifies the highlighted atoms and checks whether pIC50 moves as predicted would be the decisive test.
  • With pair counts dominated by one target (4MXO has 11,210 pairs versus 679 for 2H8H), the equal-weight averages are sensitive to the smaller datasets, and reporting per-target confidence intervals or rebalancing by scaffold family would show whether the improvement holds on data-lean targets where the method matters most.
  • The group-lasso penalty operates on just two weight groups (the common and uncommon heads), which is coarser than typical feature-group sparsity; applying the same penalty to per-channel or per-bond groups inside the GNN itself could extend the pruning beyond the MCS dichotomy.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The manuscript presents SAGGLR, a graph neural network framework for target-specific compound–protein affinity prediction that leverages activity-cliff (AC) pairs and a structure-aware node-level loss. The loss trains separate linear heads on masked common and uncommon node embeddings to reproduce experimental potency differences, and group lasso or sparse group lasso penalties are added to those heads. The framework is evaluated on six tyrosine kinase datasets with 5-fold cross-validation, comparing against several AC-aware and pretraining baselines, and feature attribution is scored using MCS-based global direction and atom-coloring metrics. The paper reports lower RMSE and higher PCC for SAGGLR+GL/SGL over most baselines, and improved attribution scores with regularization.

Significance. If the claims are sustained, the work would offer a practical way to improve single-target affinity prediction and explanation using limited AC data, and the code release and explicit hyperparameter sensitivity analysis are strengths. The improved methods over the uncommon-node-only baseline and most baselines appear internally consistent. However, the evaluation protocol permits exact compound overlap between training and test sets, and the explainability metrics use ground truth derived from the same MCS masks that supervise the node-level loss. These issues currently prevent the reported predictive and explanatory improvements from being established as real; the central claims are defensible only after re-analysis with a proper split and an independent or clearly framed explanation evaluation.

major comments (3)
  1. [Methods, Data preparation] The pair-level split rule, which states that training and test AC datasets are constructed by retaining only those compound pairs whose first ligand belonged to the corresponding training or test ligand set, permits exactly the same compound to appear in both training and test sets because the second ligand of a pair is unrestricted. The model can therefore memorize a test compound's pIC50 from a training pair containing that compound as the second ligand. The Discussion's caveat about 'highly similar compounds' understates this direct identity leakage. The RMSE/PCC gains in Tables 2 and 3, and the statistical comparisons against ACtriplet, are thus potentially driven by memorization rather than by the structure-aware loss or regularization. A split that assigns all pairs to the fold of both constituent compounds, with pairs spanning folds removed or assigned to one fold, should be used, and the experiments re-run.
  2. [Loss function design, Eqs. (13)-(15) and Model interpretability] The node-level loss LN explicitly trains MLP heads, on embeddings masked by the same MCS-derived common/uncommon masks, to reproduce the activity difference Δy. The explainability evaluation then scores whether model attributions are consistent with those same MCS masks (global direction, atom-coloring accuracy, Spearman correlation). This makes the reported explanation improvements partly a tautology: the model is directly supervised to make the masked common and uncommon substructures predict Δy, so alignment with MCS labels is expected. To support the explainability claim, add an LMSE-only baseline to the quantitative attribution comparisons (Figure 5 shows such a baseline for one example), or evaluate against independent ground truth (e.g., literature SAR series), or re-frame the claims as measuring consistency with the training signal rather than chemical validity.
  3. [Equations (14)-(15)] The group lasso and sparse group lasso penalties are applied to β(CN) and β(UCN), the parameters of the two node-level MLP heads, not to the GNN backbone or to the node embeddings used by the main affinity predictor. The Introduction and Discussion claim that regularization 'prunes and highlights molecular subgraphs' and 'enables the selection or removal of entire subgraphs,' but the penalty only induces group sparsity in the auxiliary heads that produce the node-level loss. This is a mismatch between the described mechanism and the implementation; the mechanism by which this penalty improves graph-level prediction should be clarified, or the penalty should be placed on backbone parameters.
minor comments (5)
  1. [Figure 4 caption] The caption states thresholds 'from 50% to 100% in 5% increments,' whereas the text (Model interpretability) states 'ranging from 50% to 95% in 5% increments'; these should be reconciled.
  2. [Eq. (16) and global direction text] The text says the metric assesses 'average feature attribution across the uncommon nodes,' but the equation applies the masking function to the full node attribution; please clarify how common-node contributions enter the metric.
  3. [Table 3] The caption reports P values computed against LMSE+LUCN within each architecture, but the table lists P values for RMSE and PCC without indicating which test (paired t or Wilcoxon) corresponds to each column; the supplementary tables appear to cover this, but the main table should be self-contained.
  4. [Appendix A] Selected values of λ, α, λ(CN), and λ(UCN) are not reported; providing the chosen values per kinase would improve reproducibility despite the code release.
  5. [Model evaluation] The phrase 'the accuracy and PCC values of atom-level coloring prediction' appears to be a typo; the intended metrics are likely accuracy and F1 score, which are the classification metrics described in the same section.

Circularity Check

1 steps flagged · score 6.0 of 10

Affinity-prediction comparisons are empirical and non-circular, but the explainability claims (global direction and atom-coloring accuracy) are largely forced by the structure-aware loss: the MCS-derived masks used to train the node-level loss are the same masks used as ground truth in the attribution metrics.

  1. fitted input called prediction [Methods, 'Loss function design' (Eq. 13) and 'Evaluation and explanation metrics' (Eq. 16 and mask assignment)]
    "common (CN) and uncommon (UCN) atoms were identified via substructure matching and encoded as binary node-level masks. ... The node-level loss LN ... defined as the sum of the common-structure loss LCN and the uncommon-structure loss LUCN: LN(m_k) = ... ||λ(CN)Δŷ(CN)_k + λ(UCN)Δŷ(UCN)_k − Δy_k||². ... We assigned mask = 0 for common scaffold atoms, mask = 1 for uncommon atoms in the higher-activity compound (i.e., positive ground truth), and mask = −1 for uncommon atoms in the lower-activity compound (i.e., negative ground truth)."

    The same binary masks f(CN)/f(UCN) play two roles. In Eq. 13 they define masked node embeddings and two MLP heads whose weighted sum is trained to equal the observed activity difference Δy_k, so the model is explicitly fit to reconstruct Δy from the common/uncommon mask split. In the evaluation section, the same split is renamed 'ground truth' (mask = 0/1/−1) and used to score atom-coloring accuracy and global direction. Reported improvements in attribution consistency are therefore measuring how well the model's predictions align with the very labels used to construct the training loss; they are a fitted consequence of LN, not an independent explanation of activity differences.

full rationale

The paper's affinity-prediction claim (lower RMSE/higher PCC from LN+GL/SGL) is an empirical comparison against external IC50 data and baselines, so it is not circular in the derivation sense; however the reported explainability results are substantially circular. The node-level loss LN is defined over MCS-derived common/uncommon masks and trained to reproduce the pair activity difference (Eq. 13), and the attribution metrics then use those same masks as ground-truth labels (mask=0/1/−1; Eq. 16). Thus 'global direction' and 'atom-level coloring accuracy' measure consistency with the model's own training labels rather than a causally independent explanation, and the abstract's claim that regularization enhances feature attribution is partly forced by construction. Separately, the evaluation protocol permits exact compound overlap between train and test: pairs are split by 'first ligand,' so a compound assigned to test can appear as the second ligand of a training pair; combined with the paper's own admission that random splits 'likely... place highly similar compounds in both the training and test sets, leading to inflated performance estimates,' this is a correctness risk for the affinity numbers but is not a circularity of the derivation chain. No load-bearing self-citation was found; prior-work citations (refs 23, 45) are external and the method is benchmarked against independent baselines.

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

The central contributions rest on two externally supplied ingredients: per-target IC50 data from BindingDB-derived benchmarks, and MCS-based common/uncommon masks from the RDKit FMCS algorithm. No new physical entities are introduced. The main ledger cost is the MCS ground-truth assumption, which is used in both the loss and the evaluation, plus hyperparameters selected by validation rather than derived from theory.

free parameters (3)
  • lambda (regularization strength for group lasso and sparse group lasso) = Grid-searched per kinase on the validation set; selected values not reported in the paper
    Controls group-level sparsity of the CN/UCN MLP heads. The paper states moderate lambda values are chosen via cross-validation (Appendix A) and that predictive performance depends on this choice.
  • alpha (convex mixing in sparse group lasso) = Grid-searched in {0.1, 0.3, 0.5, 0.7}; exact selected values not reported
    Interpolates between lasso and group lasso penalties in Eq. 15. This parameter defines the LN+SGL variant and is central to the sparse group lasso results.
  • lambda_CN and lambda_UCN (loss weights for common and uncommon nodes) = Tunable; values not reported
    Weight the common-structure loss and uncommon-structure loss in the node-level loss LN (Eq. 13), and the paper says these can be tuned, but no selected values are given.
assumptions (4)
  • domain assumption Atoms unique to the more potent compound of an activity-cliff pair are positive drivers; atoms unique to the less potent compound are negative drivers (MCS-based ground truth).
    Used to construct both the node-level loss LN (through masks and the delta-y decomposition) and the evaluation labels for atom coloring. The Discussion concedes this is a heuristic and not causal ground truth.
  • domain assumption Random compound-level splitting suffices to give unbiased test pairs in the small per-kinase datasets.
    The paper's own Discussion says random splits likely place highly similar compounds in both training and test sets, leading to inflated performance estimates and overly optimistic evaluation results.
  • domain assumption Activity cliffs (pairs with at least 1 log unit pIC50 difference sharing a scaffold) carry the structure-activity signal needed for per-target affinity prediction.
    The entire data construction selects on this property; if AC pairs are not representative of the compound distribution, the trained models may be biased toward cliff pairs and away from typical screening compounds.
  • ad hoc to paper GNN message-passing with 3 layers and 32-dimensional hidden representations is expressive enough to capture substituent effects for these targets.
    The architecture depth and width are design choices tested across three backbones (NNConv, GAT, GIN) but are not derived from data properties or compared against alternative depths and dimensions.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Structure-Aware Compound-Protein Affinity Prediction via Graph Neural Networks with Group Lasso Regularization." pith.science (2026). https://pith.science/paper/QH66CIWV

@misc{pith2026250703318,
  author       = {Pith},
  title        = {Pith review of: Structure-Aware Compound-Protein Affinity Prediction via Graph Neural Networks with Group Lasso Regularization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QH66CIWV}},
  note         = {Machine review of arXiv:2507.03318}
}
read the original abstract

Explainable artificial intelligence approaches accelerate drug discovery by improving molecular representation learning, identifying key molecular structures, and rationalizing drug property prediction. However, developing end-to-end explainable models for target-specific structure-activity relationship modeling remains challenging because compound-protein interaction data are often limited for individual targets, and small changes in chemical substituents or local structural motifs can cause large differences in molecular properties. Therefore, effectively leveraging structural and property information to identify key moieties associated with compound-protein affinity is essential. We propose a graph neural network (GNN) framework that uses property and structural information from activity-cliff molecule pairs targeting specific proteins to predict compound-protein affinity, measured by half-maximal inhibitory concentration (IC50), and explain property differences. To improve explainability, we trained GNNs with structure-aware loss functions using group lasso and sparse group lasso regularization, which prune and highlight molecular subgraphs relevant to activity differences. We applied this framework to activity-cliff data from molecules targeting six tyrosine-protein kinases across the Src, Abl, and Tec families, as well as anaplastic lymphoma kinase. Integrating common- and uncommon-node information with sparse group lasso improved target-specific molecular property prediction, producing lower root mean square errors and higher Pearson correlation coefficients. Regularization also enhanced GNN feature attribution by improving graph-level global direction scores and atom-level coloring accuracy. These results support more interpretable drug discovery pipelines, particularly for identifying critical molecular substructures during lead optimization.

Figures

Figures reproduced from arXiv: 2507.03318 by the authors.

Figure 1
Figure 1. An illustration for an example pair of molecules targeting Src kinase 1O42 with activity cliffs (ΔpKi = 3.85). The paired 2 molecules share common and uncommon substructures. Common substructures consist of uncolored nodes (atoms) and edges (bonds), and uncommon substituents consist of colored nodes and edges (blue and red circles for sites in the pair). Downloaded from https://spj.science.org on June 22, 2026 [PIT… view at source ↗
Figure 2
Figure 2. Model structure illustration. Considering a pair of compounds ci and cj that share a scaffold in the red circle and decorations in the blue circle, graph neural network (GNN) models and the message-passing neural network (MPNN) framework were applied to learn latent node representations for both common and uncommon nodes. Such node￾level information was aggregated to predict graph-level drug–protein binding affinity… view at source ↗
Figure 3
Figure 3. Model performance for the molecules in test sets targeting 6 kinases, respectively, under different model frameworks and loss function settings via 5-fold cross￾validation. Each bar plot indicates the averaged root mean square error (RMSE) or Pearson’s correlation coefficient (PCC) values for each loss function setting, and error bars represent the standard deviation. Wilcoxon signed-rank test P values are provided … view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Comparison of averaged global direction scores for LMSE + LN with and without group lasso by scatterplot with connecting lines. The plot was used for comparing graph￾level global direction scores to show the distribution difference of averaged predicted global directio…
Figure 5
Figure 5. Figure 5: Comparison of the atom-level accuracy in node coloring for ligands binding to the 3 Src family kinases. It compares the (A) ground-truth labels of atom coloring with prediction under 2 other situations: (B) prediction by gradient-weighted class activation mapping (Grad…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

71 extracted references · 60 canonical work pages

  1. [1]

    Machine learning in drug discovery: A review

    Dara S, Dhamercherla S, Jadav SS, Babu CM, Ahsan MJ. Machine learning in drug discovery: A review. Artif Intell Rev. 2022;55(3):1947–1999

    work page 2022

  2. [2]

    Deep learning methods for small molecule drug discovery: A survey

    Hu W , Liu Y , Chen X, Chai W , Chen H, Wang H, Wang G. Deep learning methods for small molecule drug discovery: A survey. IEEE Trans Artif Intell. 2024;5(2):459–479

    work page 2024

  3. [3]

    Neural message passing for quantum chemistry, In: Precup D, Teh YW , editors

    Gilmer J, Schoenholz SS, Riley PF , Vinyals O, Dahl GE. Neural message passing for quantum chemistry, In: Precup D, Teh YW , editors. Proceedings of the 34th international conference on machine learning. Sydney (Australia): PMLR

    work page

  4. [4]

    The graph neural network model

    Scarselli F , Gori M, Tsoi AC, Hagenbuchner M, Monfardini G. The graph neural network model. IEEE Trans Neural Netw. 2009;20(1):61–80

    work page 2009

  5. [5]

    Graph neural networks: A review of methods and applications

    Zhou J, Cui G, Hu S, Zhang Z, Y ang C, Liu Z, Wang L, Li C, Sun M. Graph neural networks: A review of methods and applications. AI Open. 2020;1:57–81

    work page 2020

  6. [6]

    Structure based protein and small molecule generation using EGNN and diffusion models: A comprehensive review

    Soleymani F , Paquet E, Viktor HL, Michalowski W . Structure based protein and small molecule generation using EGNN and diffusion models: A comprehensive review. Comput Struct Biotechnol J. 2024;23:2779–2797

    work page 2024

  7. [7]

    Explainable artificial intelligence for drug discovery and development: A comprehensive survey

    Alizadehsani R, Oyelere SS, Hussain S, Jagatheesaperumal SK, Calixto RR, Rahouti M, Roshanzamir M, De Albuquerque VHC. Explainable artificial intelligence for drug discovery and development: A comprehensive survey. IEEE Access. 2024;12:35796–35812

    work page 2024

  8. [8]

    DTI-Voodoo: Machine learning over interaction networks and ontology-based background knowledge predicts drug–target interactions

    Hinnerichs T, Hoehndorf R. DTI-Voodoo: Machine learning over interaction networks and ontology-based background knowledge predicts drug–target interactions. Bioinformatics. 2021;37(24):4835–4843

    work page 2021

Show all 71 references
  1. [9]

    Enhancing preclinical drug discovery with artificial intelligence

    Vijayan RSK, Kihlberg J, Cross JB, Poongavanam V . Enhancing preclinical drug discovery with artificial intelligence. Drug Discov Today. 2022;27(4):967–984

    2022

  2. [10]

    Bench marking of Machine Learning classifiers on plasma proteomic for COVID-19 severity prediction through interpretable artificial intelligence

    Dimitsaki S, Gavriilidis GI, Dimitriadis VK, Natsiavas P . Bench marking of Machine Learning classifiers on plasma proteomic for COVID-19 severity prediction through interpretable artificial intelligence. Artif Intell Med. 2023;137:Article 102490

    2023

  3. [11]

    Artificial intelligence in drug discovery and development

    Paul D, Sanap G, Shenoy S, Kalyane D, Kalia K, Tekade RK. Artificial intelligence in drug discovery and development. Drug Discov Today. 2021;26:80–93

    2021

  4. [12]

    ChEMBL: A large-scale bioactivity database for drug discovery

    Gaulton A, Bellis LJ, Bento AP , Chambers J, Davies M, Hersey A, Light Y , McGlinchey S, Michalovich D, Al-Lazikani B, et al. ChEMBL: A large-scale bioactivity database for drug discovery. Nucleic Acids Res. 2012;40(D1):D1100–D1107

    2012

  5. [13]

    Quantitative evaluation of explainable graph neural networks for molecular property prediction

    Rao J, Zheng S, Lu Y , Y ang Y . Quantitative evaluation of explainable graph neural networks for molecular property prediction. Patterns. 2022;3(12):Article 100628

    2022

  6. [14]

    GNNExplainer: Generating explanations for graph neural networks

    Ying Z, Bourgeois D, Y ou J, Zitnik M, Leskovec J. GNNExplainer: Generating explanations for graph neural networks. arXiv. 2019. https://doi.org/10.48550/arXiv.1903.03894

    2019 doi arXiv

  7. [15]

    Pushing the boundaries of molecular representation for drug discovery with the graph attention mechanism

    Xiong Z, Wang D, Liu X, Zhong F , Wan X, Li X, Li Z, Luo X, Chen K, Jiang H, et al. Pushing the boundaries of molecular representation for drug discovery with the graph attention mechanism. J Med Chem. 2020;63(16):8749–8760

    2020

  8. [16]

    Parameterized explainer for graph neural network

    Luo D, Cheng W , Xu D, Yu W , Zong B, Chen H, Zhang X. Parameterized explainer for graph neural network. arXiv. 2020. https://doi.org/10.48550/arXiv.2011.04573

    2020 doi arXiv

  9. [17]

    Structure-aware multimodal deep learning for drug–protein interaction prediction

    Wang P , Zheng S, Jiang Y , Li C, Liu J, Wen C, Patronov A, Qian D, Chen H, Y ang Y . Structure-aware multimodal deep learning for drug–protein interaction prediction. J Chem Inf Model. 2022;62(5):1308–1317

    2022

  10. [18]

    MM-DRPNet: A multimodal dynamic radial partitioning network for enhanced protein–ligand binding affinity prediction

    Liu D, Song T, Wang S. MM-DRPNet: A multimodal dynamic radial partitioning network for enhanced protein–ligand binding affinity prediction. Comput Struct Biotechnol J. 2024;23:4396–4405

    2024

  11. [19]

    DrugForm-DTA: Towards real-world drug- target binding affinity model

    Khokhlov I, Tashchilova A, Bugaev-Makarovskiy N, Glushkova O, Yudin V , Keskinov A, Yudin S, Svetlichnyy D, Skvortsova V . DrugForm-DTA: Towards real-world drug- target binding affinity model. Comput Struct Biotechnol J. 2025;27:4106–4120

    2025

  12. [20]

    G–PLIP: Knowledge graph neural network for structure-free protein– ligand bioactivity prediction

    Crouzet SJ, Lieberherr AM, Atz K, Nilsson T, Sach- Peltason L, Müller AT, Dal Peraro M, Zhang JD. G–PLIP: Knowledge graph neural network for structure-free protein– ligand bioactivity prediction. Comput Struct Biotechnol J. 2024;23:2872–2882

    2024

  13. [21]

    Evolving concept of activity cliffs

    Stumpfe D, Hu H, Bajorath J. Evolving concept of activity cliffs. ACS Omega. 2019;4(11):14360–14368

    2019

  14. [22]

    Large-scale prediction of activity cliffs using machine and deep learning methods of increasing complexity

    Tamura S, Miyao T, Bajorath J. Large-scale prediction of activity cliffs using machine and deep learning methods of increasing complexity. J Chem. 2023;15(1):4

    2023

  15. [23]

    Benchmarking molecular feature attribution methods with activity cliffs

    Jiménez-Luna J, Skalic M, Weskamp N. Benchmarking molecular feature attribution methods with activity cliffs. J Chem Inf Model. 2022;62(2):274–283

    2022

  16. [24]

    DIG: A turnkey library for diving into graph deep learning research

    Liu M, Luo Y , Wang L, Xie Y , Yuan H, Gui S, Yu H, Xu Z, Zhang J, Liu Y , et al. DIG: A turnkey library for diving into graph deep learning research. J Mach Learn Res. 2021;22(240):1–9

    2021

  17. [25]

    Coloring molecules with explainable artificial intelligence for preclinical relevance assessment

    Jiménez-Luna J, Skalic M, Weskamp N, Schneider G. Coloring molecules with explainable artificial intelligence for preclinical relevance assessment. J Chem Inf Model. 2021;61(3):1083–1094

    2021

  18. [26]

    Using attribution to decode binding mechanism in neural network models for chemistry

    McCloskey K, Taly A, Monti F , Brenner MP , Colwell LJ. Using attribution to decode binding mechanism in neural network models for chemistry. Proc Natl Acad Sci USA. 2019;116(24):11624–11629

    2019

  19. [27]

    Learning important features through propagating activation differences

    Shrikumar A, Greenside P , Kundaje A. Learning important features through propagating activation differences. arXiv

  20. [28]

    Grad-CAM: Visual explanations from deep networks via gradient-based localization

    Selvaraju RR, Cogswell M, Das A, Vedantam R, Parikh D, Batra D. Grad-CAM: Visual explanations from deep networks via gradient-based localization. Int J Comput Vis. 2020;128:336–359

    2020

  21. [29]

    Src family tyrosine kinases

    Frame MC, Roskoski R. Src family tyrosine kinases. Semantic Scholar. 2001. [accessed 10 Apr 2026] https://api. semanticscholar.org/CorpusID:89575548

    2001

  22. [30]

    Src protein-tyrosine kinase structure, mechanism, and small molecule inhibitors

    Roskoski R. Src protein-tyrosine kinase structure, mechanism, and small molecule inhibitors. Pharmacol Res. 2015;94:9–25

    2015

  23. [31]

    c-Abl tyrosine kinase down regulation as target for memory improvement in Alzheimer’s disease Front Aging Neurosci

    León R, Gutiérrez DA, Pinto C, Morales C, de la Fuente C, Riquelme C, Cortés BI, González-Martin A, Chamorro D, Espinosa N, et al. c-Abl tyrosine kinase down regulation as target for memory improvement in Alzheimer’s disease Front Aging Neurosci. 2023;15:Article 1180987

    2023

  24. [32]

    The potential therapeutic role of Bruton tyrosine kinase inhibition in neurodegenerative diseases

    D’Egidio F , Kacem H, Lombardozzi G, d’ Angelo M, Cimini A, Castelli V . The potential therapeutic role of Bruton tyrosine kinase inhibition in neurodegenerative diseases. Appl Sci. 2025;15(15):Article 8239

    2025

  25. [33]

    Anaplastic lymphoma kinase (ALK) receptor tyrosine kinase: A catalytic receptor with many faces

    Huang H. Anaplastic lymphoma kinase (ALK) receptor tyrosine kinase: A catalytic receptor with many faces. Int J Mol Sci. 2018;19(11):Article 3448

    2018

  26. [34]

    Src family kinases as therapeutic targets in advanced solid tumors: What we have learned so far

    Martellucci S, Clementi L, Sabetta S, Mattei V , Botta L. Src family kinases as therapeutic targets in advanced solid tumors: What we have learned so far. Cancer. 2020;12(6):1448. Downloaded from https://spj.science.org on June 22, 2026 Shi et al. 2026 | https://doi.org/10.341...

    2020 doi

  27. [35]

    Recent developments of protein kinase inhibitors as potential AD therapeutics

    Hilgeroth AP , Tell V . Recent developments of protein kinase inhibitors as potential AD therapeutics. Front Cell Neurosci. 2013;7:Article 189

    2013

  28. [36]

    A sparse-group lasso

    Simon N, Friedman J, Hastie T, Tibshirani R. A sparse-group lasso. J Comput Graph Stat. 2013;22(2):231–245

    2013

  29. [37]

    BindingDB in 2015: A public database for medicinal chemistry, computational chemistry and systems pharmacology

    Gilson MK, Liu T, Baitaluk M, Nicola G, Hwang L, Chong J. BindingDB in 2015: A public database for medicinal chemistry, computational chemistry and systems pharmacology. Nucleic Acids Res. 2016;44(D1):D1045–D1053

    2015

  30. [38]

    The Protein Data Bank

    Berman HM, Westbrook J, Feng Z, Gilliland G, Bhat TN, Weissig H, Shindyalov IN, Bourne PE. The Protein Data Bank. Nucleic Acids Res. 2000;28(1):235–242

    2000

  31. [39]

    Neural message passing for quantum chemistry

    Gilmer J, Schoenholz SS, Riley PF , Vinyals O, Dahl GE. Neural message passing for quantum chemistry. arXiv. 2017. https://doi.org/10.48550/arXiv.1704.01212

    2017 doi arXiv

  32. [40]

    Dynamic edge-conditioned filters in convolutional neural networks on graphs

    Simonovsky M, Komodakis N. Dynamic edge-conditioned filters in convolutional neural networks on graphs. arXiv. 2017. https://doi.org/10.48550/arXiv.1704.02901

    2017 doi arXiv

  33. [41]

    How powerful are graph neural networks? arXiv

    Xu K, Hu W , Leskovec J, Jegelka S. How powerful are graph neural networks? arXiv. 2019. https://doi.org/10.48550/arXiv.1810.00826

    2019 doi arXiv

  34. [42]

    Graph attention networks

    Veličković P , Cucurull G, Casanova A, Romero A, Lio P , Bengio Y . Graph attention networks. arXiv. 2018. https://doi. org/10.48550/arXiv.1710.10903

    2018 doi arXiv

  35. [43]

    Strategies for pre-training graph neural networks

    Hu W , Liu B, Gomes J, Zitnik M, Liang P , Pande V , Leskovec J. Strategies for pre-training graph neural networks. arXiv. 2020. https://doi.org/10.48550/arXiv.1905.12265

    2020 doi arXiv

  36. [44]

    Scheduling techniques for liver segmentation: ReduceLRonPlateau vs OneCycleLR

    Al-Kababji A, Bensaali F , Dakua SP . Scheduling techniques for liver segmentation: ReduceLRonPlateau vs OneCycleLR. arXiv. 2022. https://doi.org/10.48550/arXiv.2202.06373

    2022 doi arXiv

  37. [45]

    Explaining compound activity predictions with a substructure-aware loss for graph neural networks

    Amara K, Rodríguez-Pérez R, Jiménez-Luna J. Explaining compound activity predictions with a substructure-aware loss for graph neural networks. J Chem. 2023;15(1):67

    2023

  38. [46]

    Group lasso regularized deep learning for cancer prognosis from multi- omics and clinical features

    Xie G, Dong C, Kong Y , Zhong JF , Li M, Wang K. Group lasso regularized deep learning for cancer prognosis from multi- omics and clinical features. Genes. 2019;10(3):240

    2019

  39. [47]

    Evaluating attribution for graph neural networks

    Sanchez-Lengeling B, Wei J, Lee B, Reif E, Wang P , Qian W , McCloskey K, Colwell L, Wiltschko A. Evaluating attribution for graph neural networks. Adv Neural Inf Proc Syst. 2020;33:5898–5910

    2020

  40. [48]

    Learning deep features for discriminative localization

    Zhou B, Khosla A, Lapedriza A, Oliva A, Torralba A. Learning deep features for discriminative localization. arXiv. 2015. https://doi.org/10.48550/arXiv.1512.04150

    2015 doi arXiv

  41. [49]

    Axiomatic attribution for deep networks

    Sundararajan M, Taly A, Y an Q. Axiomatic attribution for deep networks. arXiv. 2017. https://doi.org/10.48550/arXiv.1703.01365

    2017 doi arXiv

  42. [50]

    A self-conformation-aware pre- training framework for molecular property prediction with substructure interpretability

    Qiao J, Jin J, Wang D, Teng S, Zhang J, Y ang X, Liu Y , Wang Y , Cui L, Zou Q, et al. A self-conformation-aware pre- training framework for molecular property prediction with substructure interpretability. Nat Commun. 2025;16(1):4382

    2025

  43. [51]

    ACES-GNN: Can graph neural network learn to explain activity cliffs? Digit Discov

    Chen X, Yu D, Zhao L, Liu F . ACES-GNN: Can graph neural network learn to explain activity cliffs? Digit Discov. 2025;4(8):2062–2074

    2025

  44. [52]

    A semi-supervised molecular learning framework for activity cliff estimation

    Wu F . A semi-supervised molecular learning framework for activity cliff estimation. In: Larson K, editor. Proceedings of the thirty-third international joint conference on artificial intelligence (IJCAI-24). Palo Alto (CA): International Joint Conferences on Artificial Intell...

    2024

  45. [53]

    ACGCN: Graph convolutional networks for activity cliff prediction between matched molecular pairs

    Park J, Sung G, Lee S, Kang S, Park C. ACGCN: Graph convolutional networks for activity cliff prediction between matched molecular pairs. J Chem Inf Model. 2022;62(10):2341–2351

    2022

  46. [54]

    ACtriplet: An improved deep learning model for activity cliffs prediction by integrating triplet loss and pre-training

    Yu X, Wang Y , Chen L, Li W , Tang Y , Liu G. ACtriplet: An improved deep learning model for activity cliffs prediction by integrating triplet loss and pre-training. J Pharmaceut Anal. 2025;15:Article 101317

    2025

  47. [55]

    Practically significant method comparison protocols for machine learning in small molecule drug discovery

    Ash JR, Wognum C, Rodríguez-Pérez R, Aldeghi M, Cheng AC, Clevert D-A, Engkvist O, Fang C, Price DJ, Hughes-Oliver JM, et al. Practically significant method comparison protocols for machine learning in small molecule drug discovery. J Chem Inf Model. 2025;65(18):9398–9411

    2025

  48. [56]

    Individual comparisons by ranking methods

    Wilcoxon F . Individual comparisons by ranking methods. Biom Bull. 1945;1(6):80–83

    1945

  49. [57]

    Correlation coefficients: Appropriate use and interpretation

    Schober P , Boer C, Schwarte LA. Correlation coefficients: Appropriate use and interpretation. Anesth Analg. 2018;126(5):1763–1768

    2018

  50. [58]

    SDDSynergy: Learning important molecular substructures for explainable anticancer drug synergy prediction

    Liu Y , Zhang P , Che C, Wei Z. SDDSynergy: Learning important molecular substructures for explainable anticancer drug synergy prediction. J Chem Inf Model. 2024;64(24):9551–9562

    2024

  51. [59]

    MGTNSyn: Molecular structure aware graph transformer network with relational attention for drug synergy prediction

    Liu Y , Zhang P , Li D, Che C, Jin B. MGTNSyn: Molecular structure aware graph transformer network with relational attention for drug synergy prediction. Expert Syst Appl. 2025;282:Article 127699

    2025

  52. [60]

    Key substructure learning with chemical intuition for material property prediction

    Zhang P , Yuan J, Li L, Luo W , Hu J, Li X. Key substructure learning with chemical intuition for material property prediction. In: Onizuka M, Lee J-G, Tong Y , Xiao C, Ishikawa Y , Amer-Y ahia S, Jagadish HV , Lu K, editors. Database systems for advanced applications: 29th in...

    2024

  53. [61]

    KSGTN-DDI: Key substructure- aware graph transformer network for drug-drug interaction prediction

    Zhang P , Liu Y , Shen Z. KSGTN-DDI: Key substructure- aware graph transformer network for drug-drug interaction prediction. Paper presented at: 2023 IEEE International Conference on Bioinformatics and Biomedicine (BIBM); 2023 Dec 5–8; Istanbul, Türkiye

    2023

  54. [62]

    Representational alignment with chemical induced fit for molecular relational learning

    Zhang P , Yuan J, Xie Q, Zhu Y , Li L. Representational alignment with chemical induced fit for molecular relational learning. arXiv. 2025. https://doi.org/10.48550/arXiv.2502.07027

    2025 doi arXiv

  55. [63]

    Subgraph information bottleneck with causal dependency for stable molecular relational learning

    Zhang P , Yuan J, Che C, Zhu Y , Li L. Subgraph information bottleneck with causal dependency for stable molecular relational learning. In: Kwok J, editor. Proceedings of the thirty- fourth international joint conference on artificial intelligence (IJCAI-25). Palo Alto (CA): I...

    2025

  56. [64]

    Matched molecular pair analysis in short: Algorithms, applications and limitations

    Tyrchan C, Evertsson E. Matched molecular pair analysis in short: Algorithms, applications and limitations. Comput Struct Biotechnol J. 2017;15:86–90

    2017

  57. [65]

    A maximum common substructure- based algorithm for searching and predicting drug-like compounds

    Cao Y , Jiang T, Girke T. A maximum common substructure- based algorithm for searching and predicting drug-like compounds. Bioinformatics. 2008;24(13):i366–i374

    2008

  58. [66]

    Macrocycles in drug discovery-learning from the past for the future

    Garcia Jimenez D, Poongavanam V , Kihlberg J. Macrocycles in drug discovery-learning from the past for the future. J Med Chem. 2023;66(8):5377–5396

    2023

  59. [67]

    Targeting Src family kinases in anti-cancer therapies: Turning promise into triumph

    Zhang S, Yu D. Targeting Src family kinases in anti-cancer therapies: Turning promise into triumph. Trends Pharmacol Sci. 2012;33(3):122–128

    2012

  60. [68]

    c-Abl in neurodegenerative disease

    Schlatterer SD, Acker CM, Davies P . c-Abl in neurodegenerative disease. J Mol Neurosci. 2011;45(3):445–452

    2011

  61. [69]

    TEC family kinases in health and disease—Loss- of-function of BTK and ITK and the gain-of-function fusions ITK–SYK and BTK–SYK

    Hussain A, Yu L, Faryal R, Mohammad DK, Mohamed AJ, Smith CIE. TEC family kinases in health and disease—Loss- of-function of BTK and ITK and the gain-of-function fusions ITK–SYK and BTK–SYK. FEBS J. 2011;278(12):2001–2010

    2011

  62. [70]

    Aberrant role of ALK in tau proteinopathy through autophagosomal dysregulation

    Park J, Choi H, Kim YD, Kim SH, Kim Y , Gwon Y , Lee DY , Park SH, Heo WD, Jung YK. Aberrant role of ALK in tau proteinopathy through autophagosomal dysregulation. Mol Psychiatry. 2021;26(10):5542–5556. Downloaded from https://spj.science.org on June 22, 2026 Structure-Aware C...

    2021

  63. [2019]

    https://doi.org/10.48550/arXiv.1704.02685

    doi arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.