Pith. sign in

REVIEW 4 major objections 12 minor 69 references

A graph network pre-trained on ~100 million molecules predicts ADMET properties better than strong baselines and holds up on new compounds.

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 ↗

T0 review · grok-4.5

2026-07-31 18:23 UTC pith:WNEOHNFO

load-bearing objection Solid engineering bundle with real wet-lab follow-up, but the 18-drug HLMC external table is internally corrupted and cannot support the main generalization claim as written. the 4 major comments →

arxiv 2607.24314 v1 pith:WNEOHNFO submitted 2026-07-27 cs.LG

MEGA-CL: A Molecular Foundation Model for Generalizable ADMET Prediction through Graph External Attention and Contrastive Learning

Tinghui Jin , Kedu Jin , Ying Li , Guanghui Ren , Jingzhi Xue , Shiyu Zhou , Xiaoli Dai , Li-bin Wei
show 3 more authors
Xijing Chen Di Zhao Jinfeng Liu
This is my paper
classification cs.LG
keywords ADMET predictiongraph neural networkscontrastive learninggraph external attentionmolecular foundation modelpharmacokineticsCYP450 inhibitionhuman liver microsomal clearance
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.

Drug candidates often fail because absorption, distribution, metabolism, excretion, and toxicity (ADMET) are hard to predict from structure alone, and labeled data are scarce and noisy. This paper introduces MEGA-CL, a foundation-style graph neural network that first learns molecular representations by contrastive pre-training on roughly 100 million unlabeled compounds, then fine-tunes on specific ADMET tasks. Its encoder pairs a deeper residual message-passing backbone with multi-head graph external attention so the model can use both local chemistry inside a molecule and shared patterns across related molecules. On MoleculeNet benchmarks and a curated suite of 21 ADMET endpoints it beats conventional machine-learning and graph baselines, especially on hard continuous pharmacokinetic quantities such as clearance and volume of distribution. External checks on recent FDA drugs and a small prospective lab study on three preclinical candidates show many predictions within clinically useful fold-error ranges, supporting the claim that the approach can speed early in silico ADMET screening.

Core claim

MEGA-CL shows that contrastive pre-training on a very large unlabeled chemical set, combined with residual graph convolution and multi-head graph external attention, yields transferable molecular representations that outperform strong baselines across 13 MoleculeNet datasets and 21 ADMET tasks, and that those representations remain useful on independent FDA-drug clearance data and on three prospectively assayed preclinical compounds (HLMC within 2.5-fold; 11/15 CYP450 inhibition labels correct).

What carries the argument

MEGA-CL encoder: GCN+ (residual, normalized message passing that limits over-smoothing) plus multi-head graph external attention (learnable external memory that mixes inter-molecule chemical prototypes), trained with NT-Xent contrastive loss on augmented molecular graphs, then fine-tuned with task heads.

Load-bearing premise

That strong scaffold-split scores plus a handful of external drugs and only three prospectively tested compounds are enough to show the model will generalize for real early-stage ADMET decisions when many labels are noisy and system-level rather than pure structure–property maps.

What would settle it

On a larger, fully held-out set of new scaffolds with measured human liver microsomal clearance and CYP450 inhibition, check whether MEGA-CL still keeps most continuous predictions inside 2–3-fold error and classification accuracy near the reported 73% while beating the same baselines; a clear drop below those margins would falsify the practical-generalization claim.

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

If this is right

  • Early discovery teams could rank candidates on clearance, VDss, and CYP liability in silico before committing to microsomal and inhibition assays.
  • Hard continuous PK endpoints (CL, VDss, HLM) become more tractable when local substructure encoding is paired with cross-molecule attention over large chemical space.
  • The same pre-trained encoder can be reused across many ADMET endpoints via light fine-tuning rather than training separate models from scratch.
  • Prospective use is framed as guiding which in vitro metabolism and CYP assays to run, not replacing them.

Where Pith is reading between the lines

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

  • If external attention mainly helps by sharing signal across scaffold neighbors, gains should shrink on truly isolated chemotypes far from the pre-training and fine-tuning manifolds.
  • System-level endpoints such as half-life and mean residence time will likely stay noisier than binding-like properties (e.g., Fu) unless multi-task or physiology-aware heads are added.
  • A natural next stress test is time-split or organization-holdout evaluation on industrial ADMET decks to separate chemical-space generalization from assay-protocol familiarity.

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

4 major / 12 minor

Summary. The manuscript presents MEGA-CL, a graph neural network for molecular property and ADMET prediction that combines a GCN+ message-passing backbone (residual connections, layer normalization) with a multi-head "graph external attention" (GEA) module over learned memory matrices, pre-trained with NT-Xent contrastive learning on unlabeled PubChem molecules and fine-tuned on downstream tasks. The authors evaluate on 13 MoleculeNet benchmarks, a curated 21-task ADMET benchmark (14 classification, 7 regression, including in vivo PK endpoints from Lombardo et al.), nine Pharmabench tasks, an 18-compound external set of FDA-approved drugs for HLMC prediction, and a prospective wet-lab validation on three preclinical compounds (Oroxylin A, Icariin, and an undisclosed Compound B) with measured microsomal clearance and CYP450 inhibition profiles. The central claims are consistent superiority over ML/GNN baselines and clinically useful fold-error accuracy (majority of predictions within 2–3-fold), supported by the external and prospective experiments.

Significance. If the results hold, this is a useful contribution at the intersection of molecular representation learning and preclinical DMPK. Genuine strengths that should be credited: (i) all downstream experiments are repeated three times with mean±SD reported, which is better practice than most leaderboard-style papers; (ii) scaffold splitting (8:1:1) is used throughout, a more honest protocol than random splits; (iii) the fold-error analysis (Fig. 3) stratifies endpoints by pharmacokinetic complexity rather than reporting a single aggregate number, and the discussion of Fu vs. CL/HL/MRT is scientifically sensible; (iv) the prospective wet-lab component — actual microsomal stability and CYP inhibition assays on three compounds, with full SOP, MRM parameters, and kinetic calculations in the SI — is rare in this literature and substantially raises the evidentiary bar; (v) code (GitHub) and pre-training data (Zenodo) are deposited, supporting reproducibility. The architecture itself (external attention over learned memory prototypes grafted onto a MolCLR-style contrastive pipeline) is an incremental but reasonable combination of known components. However, the strongest claim — real-world gener

major comments (4)
  1. [External validation on clinically approved drugs; Supplementary Table S8] Table S8 contains duplicated prediction values that a continuous regressor cannot plausibly produce, which undermines the 16/18-within-2-fold HLMC statistic that anchors the external-validation claim (Fig. 7D and the abstract's '>50% within 2-fold'). Specifically: (a) the four-value block 11.61 / 20.15 / 18.56 / 8.57 appears twice — for Tandutinib, Veliparib, Ensartinib, Niraparib and again for Olaparib, Infigratinib, Zorifertinib, Talazoparib; (b) Lenvatinib's predicted HLMC (47.93) matches Oroxylin A's prospective prediction (47.9291, Table 3) and Sorafenib's (52.48) matches Compound B's (52.4807, Table 3); (c) Veliparib appears twice with different predictions (20.15 and 23.32), so the '18 compounds' contain at most 17 unique molecules. Independent continuous predictions agreeing to 4+ significant figures across structurally distinct molecules do not occur by chance; at minimum the ta
  2. [External validation on clinically approved drugs; Methods, Dataset compilation] No train/test decontamination is described for the external FDA set, and contamination is plausible. The external set is described as drugs 'approved or newly indicated within the past five years', but it includes sorafenib (approved 2005), nebivolol (2007), and lenvatinib (2015). The HLMC model was fine-tuned on 2,284 Pharmabench HLMC records, a large fraction of the marketed-drug universe with measured microsomal clearance; several of the 18 external drugs (sorafenib, olaparib, niraparib, veliparib, lenvatinib) are exactly the kind of well-studied compounds likely to appear in such a collection. If any external compounds (or close analogues) are in the fine-tuning or pre-training data, the 16/18-within-2-fold result is not evidence of generalization. The authors should report an explicit overlap check (exact structure and scaffold-level) between the 18 external compounds and all traini
  3. [Abstract; Fig. 1C; Methods (Dataset compilation); Supplementary Fig. S1 caption] The pre-training corpus size is stated inconsistently in a way that bears on the 'foundation model' framing. The abstract, Fig. 1C, and the Fig. S1 caption claim pre-training on ~100 million unique molecules, while Methods (Dataset compilation and preprocessing) states that filtering of 1.7×10^8 PubChem entries 'yielded approximately 10 million unique, unannotated molecular SMILES strings'. These differ by an order of magnitude. Since the scale of pre-training is a central selling point (the abstract opens with it), the correct number must be established and used consistently; if the true corpus is ~10M, comparisons to MolCLR (pre-trained on ~10M) should also be reframed, since the claimed gains would no longer be attributable to larger pre-training scale.
  4. [Abstract; Results (fold-error analysis; external validation); Fig. 7 caption; Tables 1–2, S2] Several headline accuracy statistics are mutually inconsistent across the abstract, main text, and figure captions, and the 'consistently outperforms' claim is overstated relative to the authors' own tables. (a) The abstract claims '>75% of predictions within a 3-fold error range', but the fold-error section states 'more than 75% of predictions falling within a 5-fold error range for each dataset' — these are very different claims. (b) The abstract claims 'over 50% within 2-fold' for the 18-drug external set while the text reports 16/18 (89%). (c) Fig. 7's caption says '21 independent drugs' where the text and Table S8 describe 18 (and the conclusions merge the 18 FDA drugs with the 3 prospective compounds to get 21). (d) 'Consistently outperforms state-of-the-art baselines' (abstract) is not supported by Table S2, where MEGA-CL loses to at least one baseline on CYP1A2_sub, CYP2C19_sub,
minor comments (12)
  1. [Methods, Fine-tuning and evaluation metrics] Eq. (8) defines MAE but the formula shown is the mean of squared errors (MSE); the squared term should be an absolute value.
  2. [Methods, Fine-tuning and evaluation metrics] Eq. (9): the denominator of R² is typeset as ∑(y_i − y_i)², which is identically zero; the mean of the experimental values should appear in the second term. The sentence following the equations ('represent the experimental value, predicted value, and the mean...') also lacks subjects.
  3. [Methods, Multi-Head Graph External Attention] The GEA description states 'two learnable external memory matrices' but only M_K is defined; M_V, the head dimension split, and the symbol d_v are never defined. Please give M_K, M_V, S, H, and d_k/d_v explicit definitions and typical values.
  4. [Supplementary Information, §3.2 (IC50 Calculations)] The CYP inhibition classification logic is inverted in the SI: 'compounds with an IC50 less than 10 μM were considered to exhibit no significant inhibitory activity' — per the main text (threshold 10 μM defines inhibitors), this should be 'greater than'.
  5. [External prospective experimental validation; SI §1.1] Compound B's structure, purity, and provenance are withheld. This limits independent verification of the prospective validation, which is one of the paper's main strengths. If disclosure is genuinely impossible, the authors should state so explicitly in the main text and discuss the limitation; ideally a blinded structure (e.g., scaffold class) could be given.
  6. [Results (ADMET benchmark construction); Methods; Table S1] Task counts are inconsistent: the Results describe 21 tasks (14 classification + 7 regression), Methods says '19 tasks' curated and later '12 classification tasks and 7 regression tasks', while Table S1 lists 14 classification + 7 regression. Please reconcile.
  7. [Table S3; Methods, Baseline] Table S3 lists 'Logistics' (logistic regression) as a baseline for regression tasks. Logistic regression is a classifier; if linear regression was intended, please correct and clarify throughout (including the Baseline paragraph).
  8. [Table 3; Table S5] Table 3 reports Compound B experimental HLMC as 30.15, while the main text and Table S5 report 31.5 (and the text's stated fold error 1.66 uses 31.50). Please reconcile.
  9. [Results, Framework of MEGA-CL] Fig. 1A text: 'yielding G̃_{2n−1} and G̃_{2n−1}' — the two augmented views should have distinct indices (2n−1 and 2n).
  10. [Throughout] Typographical: 'Compoud B' (Table 3 title and column headers); 'ADMETT predictions' (Methods); duplicated internal standard 'caffeine, caffine' (SI §1.2); '𝒱 enotes' (Methods); inconsistent use of HLM vs. HLMC for the same endpoint.
  11. [References] Reference 45 (Berkson, logistic function) is cited with a 2012 date for a 1944 paper; several arXiv preprint citations (e.g., refs. 14, 15, 36) have since appeared in peer-reviewed venues and should be updated.
  12. [Abstract; Fig. 3] The abstract's 'more than 75% of predictions falling within a 3-fold error range' cannot be verified from Fig. 3A as described in the text; please state the exact within-3-fold percentages per endpoint in the text or a table.

Circularity Check

0 steps flagged

No derivation-chain circularity: standard pretrain–finetune ML evaluated on held-out and external labels, not predictions forced by construction.

full rationale

MEGA-CL’s load-bearing chain is (1) contrastive pre-training on ~100M unlabeled PubChem SMILES with NT-Xent, GCN+, and multi-head GEA, then (2) task-specific fine-tuning and scoring on MoleculeNet, curated ADMET sets, Pharmabench, an 18-drug FDA HLMC set, and three prospective compounds with new wet-lab assays. None of these steps defines the reported metrics in terms of the quantities being predicted: pre-training does not use ADMET labels; GEA external memories and residual GCN+ updates are architectural priors, not algebraic restatements of CL/VDss/HLMC; ROC-AUC/RMSE/fold-error are computed after supervised fine-tuning on held-out splits. Citations (MolCLR, external attention, Chemprop, Lombardo PK tables, PubChem AID 1851) are third-party methods or public data, not author uniqueness theorems that force the result. Apparent integrity problems in Table S8 (duplicated predicted values, caption n mismatches) and possible train/external overlap are data-quality or leakage concerns, not circular reduction of a claimed derivation to its inputs. Under the stated criteria the paper is self-contained against external benchmarks with no self-definitional or fitted-input-as-prediction loop.

Axiom & Free-Parameter Ledger

6 free parameters · 6 axioms · 2 invented entities

The claim rests on standard GNN/contrastive learning assumptions, the hypothesis that external memory attention captures useful inter-molecule chemistry, large-scale PubChem pre-training transfer, and the suitability of noisy public ADMET labels plus small prospective n for asserting practical utility. Free parameters are ordinary deep-learning hyperparameters and assay thresholds rather than physical constants.

free parameters (6)
  • NT-Xent temperature τ
    Scales similarity in the contrastive loss; chosen as a training hyperparameter, not derived.
  • Graph augmentation probabilities (atom mask p_m, bond drop p_b, subgraph removal)
    Stochastic view generation controls what invariance is learned; set by authors following MolCLR-style practice.
  • External memory size S and number of attention heads H
    Define capacity of GEA global prototypes; architectural choices that affect representation power.
  • Pre-training LR schedule (5e-4, warmup, cosine), weight decay 1e-5, batch 512, 100 epochs = lr=5e-4, wd=1e-5, batch=512, 100 epochs
    Optimization settings that determine the encoder actually obtained.
  • CYP inhibitor IC50 threshold 10 μM = 10 μM
    Binarization boundary taken from PubChem AID 1851 / study convention; directly defines classification labels and prospective accuracy.
  • Fine-tuning classification threshold 0.5 = 0.5
    Maps model probabilities to inhibitor/non-inhibitor calls in external validation Table 3.
axioms (6)
  • domain assumption Molecules can be faithfully represented as atom–bond graphs with standard RDKit featurization for ADMET-relevant learning.
    Stated in molecular graph construction; underpins the entire GNN pipeline.
  • domain assumption Contrastive agreement under atom mask / bond delete / subgraph removal yields transferable chemical representations.
    Inherited from MolCLR-style pre-training; core of the self-supervised stage.
  • domain assumption Residual GCN+ with BN/FFN/Dropout sufficiently mitigates over-smoothing for the depths used.
    Methods equations (1)–(3); justifies deeper message passing claim.
  • ad hoc to paper Multi-head attention over learned external memory matrices captures inter-graph chemical correlations useful for PK/ADMET.
    GEA module motivation in Introduction/Methods; not independently proven outside this architecture’s ablations (limited).
  • domain assumption Scaffold splits (8:1:1) and chosen public ADMET compilations are adequate proxies for prospective discovery generalization.
    Dataset compilation section; load-bearing for ‘generalizable’ claims.
  • standard math NT-Xent, mean pooling, and MLP heads are appropriate training/readout objectives for the reported metrics.
    Standard SSL and GNN practice cited via loss equations (5)–(9).
invented entities (2)
  • MEGA-CL encoder (GCN+ backbone + multi-head GEA) no independent evidence
    purpose: Joint local message passing and global inter-graph prototype attention for molecular embeddings.
    Named architecture combining existing pieces; evaluated only inside this paper’s experiments.
  • Learnable external memory matrices M_K, M_V (graph prototypes) no independent evidence
    purpose: Provide cross-molecule keys/values so attention is not limited to intra-graph nodes.
    Adapted from external attention literature; no external physical meaning or independent assay of the prototypes.

pith-pipeline@v1.2.0-grok45-kimik3 · 30310 in / 3852 out tokens · 66522 ms · 2026-07-31T18:23:53.725404+00:00 · methodology

0 comments
read the original abstract

Predicting the absorption, distribution, metabolism, excretion and toxicity (ADMET) properties of small molecules remains a major challenge in drug discovery. Here, we present MEGA-CL, a foundation graph neural network framework for universal molecular ADMET prediction. MEGA-CL integrates self-supervised contrastive learning with a multi-head external attention mechanism and an enhanced message-passing architecture, enabling simultaneous modeling of local chemical substructures and global inter-graph relationships while mitigating over-smoothing effects commonly observed in deep graph networks. Across 13 benchmark datasets and 21 downstream ADMET tasks, MEGA-CL consistently outperforms state-of-the-art baseline models. In particular, the framework demonstrates robust performance on challenging regression tasks, including clearance (CL) and steady-state volume of distribution (VDss), while maintaining strong generalization ability in independent external validation. Clinically relevant predictive accuracy was achieved, with more than 75% of predictions falling within a 3-fold error range. In an external evaluation on 18 novel compounds derived from recently approved FDA drugs, over 50% of human liver microsome clearance (HLMC) predictions were within a 2-fold error range. To further assess its practical applicability, MEGA-CL was prospectively evaluated on three preclinical drug candidates using in vitro hepatic microsomal metabolism assays and CYP450 inhibition assays guided by model predictions. The predicted HLMC values for all candidates were within 2.5-fold of the experimentally measured values, and 73.3% of CYP450 inhibition endpoints (11/15) were correctly classified. These results demonstrate the potential of MEGA-CL as a generalizable framework for accelerating in silico ADMET evaluation and early-stage drug candidate optimization.

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

69 extracted references · 4 canonical work pages · 2 internal anchors

  1. [1]

    & Philpott, K

    Hughes, J., Rees, S., Kalindjian, S. & Philpott, K. Principles of early drug discovery. Br. J. Ph armacol. 162, 1239 –1249 (2011)

    2011

  2. [2]

    Zhang, K. et al. Artificial intelligence in drug development. Nat. Med. 31, 45 –59 (2025). 25

    2025

  3. [3]

    Yu, L. X. Pharmaceutical Quality by Design: Product and Process Development, Understanding, and Control. Pharm. Res. 25, 781 –791 (2008)

    2008

  4. [4]

    Waring, M. J. et al. An analysis of the attrition of drug candidates from four major pharmaceut ical companies. Nat. Rev. Drug Discovery 14, 475 –486 (2015)

    2015

  5. [5]

    & Zhou, S

    Sun, D., Gao, W., Hu, H. & Zhou, S. Why 90% of clinical drug development fails and how to improve it? Acta Pharm. Sin., B 12, 3049 –3062 (2022)

    2022

  6. [6]

    Kerns, E. H. & Di, L. Drug -like Properties: Concepts, Structure Design and Methods: From AD ME to Toxicity Optimization. (Academic Press, Amsterdam ; Boston, 2008)

    2008

  7. [7]

    & Houston, J

    Ito, K. & Houston, J. B. Comparison of the Use of Liver Models for Predicting Drug Clearanc e Using in Vitro Kinetic Data from Hepatic Microsomes and Isolated Hepatocytes. Pharm. Res. 21, 785 –792 (2004)

    2004

  8. [8]

    Qiang, B. et al. Bridging the gap between chemical reaction pretraining and conditional molecul e generation with a unified model. Nat Mach Intell 5, 1476 –1485 (2023)

    2023

  9. [9]

    Wang, W. et al. An Integrated AI -PBPK Platform for Predicting Drug In Vivo Fate and Tissue Distribution in Human and Inter -Species Extrapolation. Clin. Pharmacol. Ther. 118, 865 –875 (2 025)

  10. [10]

    & Lowe, E

    Sliwoski, G., Kothiwale, S., Meiler, J. & Lowe, E. W. Computational Methods in Drug Discove ry. Pharmacol. Rev. 66, 334 –395 (2014)

    2014

  11. [11]

    Sadybekov, A. V. & Katritch, V. Computational approaches streamlining drug discovery. Nature 616, 673 –685 (2023)

    2023

  12. [12]

    & Maddi, S

    Venkataraman, M., Rao, G., Reddi, J. & Maddi, S. Leveraging machine learning models in eval uating ADMET properties for drug discovery and development. ADMET DMPK 13, 2772 (202 5)

  13. [13]

    & Vaid, R

    Suvidhi, M., Kumar, S., Sumanshu, M. & Vaid, R. In silico admet predictions: enhancing drug development through qsar modeling. in 41 –52 (2024). doi:10.58532/V3BCAG21P2CH1

    work page doi:10.58532/v3bcag21p2ch1 2024

  14. [14]

    S., Riley, P

    Gilmer, J., Schoenholz, S. S., Riley, P. F., Vinyals, O. & Dahl, G. E. Neural Message Passing for Quantum Chemistry. Preprint at https://arxiv.org/abs/1704.01212v2 (2017)

    Pith/arXiv arXiv 2017

  15. [15]

    Schü tt, K. T. et al. SchNet: A continuous -filter convolutional neural network for modeling quan tum interactions. Preprint at https://arxiv.org/abs/1706.08566v5 (2017)

    Pith/arXiv arXiv 2017

  16. [16]

    Guo, Z. et al. Graph-based molecular representation learning. Preprint at https://doi.org/10.48550 /arXiv.2207.04869 (2023)

    Pith/arXiv arXiv doi:10.48550/arxiv.2207.04869 2023

  17. [17]

    -H., Le, M

    Dinh Pham, L. -H., Le, M. -T. & Thai, K. -M. Improved ADME Prediction by Multitask Pretraini ng on Predicted Data: Insights from the ASAP -Polaris-OpenADMET Blind Challenge. J. Chem. Inf. Model. 66, 395 –405 (2026)

    2026

  18. [18]

    Putrama, I. M. & Martinek, P. Heterogeneous data integration: Challenges and opportunities. Da ta Brief 56, 110853 (2024)

    2024

  19. [19]

    & Chandrasekaran, S

    Pallikkavaliyaveetil, N. & Chandrasekaran, S. Small data, big challenges: Machine - and deep -lea rning strategies for data -limited drug discovery. Adv. Drug Delivery Rev. 229, 115762 (2026). 26

    2026

  20. [20]

    & Xie, L

    Liu, Y., Lim, H. & Xie, L. Exploration of chemical space with partial labeled noisy student sel f-training and self -supervised graph embedding. BMC Bioinf. 23, 158 (2022)

    2022

  21. [21]

    & Barati Farimani, A

    Wang, Y., Wang, J., Cao, Z. & Barati Farimani, A. Molecular contrastive learning of representa tions via graph neural networks. Nat. Mach. Intell. 4, 279 –287 (2022)

    2022

  22. [22]

    You, Y. et al. Graph Contrastive Learning with Augmentations. Preprint at https://arxiv.org/abs/ 2010.13902 (2020)

    Pith/arXiv arXiv 2010

  23. [23]

    Chen, Y. et al. Self-supervised learning in drug discovery. Sci. China Inf. Sci. 68, 170103 (202 5)

  24. [24]

    Xia, J., Zhu, Y., Du, Y. & Li, S. Z. Pre -training Graph Neural Networks for Molecular Represe ntations: Retrospect and Prospect. in ICML 2022 2nd AI for Science Workshop (2022)

    2022

  25. [25]

    Li, X. et al. Neural atoms: propagating long -range interaction in molecular graphs through effici ent communication channel. Preprint at https://doi.org/10.48550/arXiv.2311.01276 (2024)

    Pith/arXiv arXiv doi:10.48550/arxiv.2311.01276 2024

  26. [26]

    Godwin, J. et al. Simple GNN regularisation for 3D molecular property prediction & beyond. P reprint at https://doi.org/10.48550/arXiv.2106.07971 (2022)

    Pith/arXiv arXiv doi:10.48550/arxiv.2106.07971 2022

  27. [27]

    & Zhou, S

    Zhao, B., Xu, W., Guan, J. & Zhou, S. Molecular property prediction based on graph structure learning. Bioinformatics 40, btae304 (2024)

    2024

  28. [28]

    Concepts and applications of molecular similarity, by Mark A

    Klopmand, G. Concepts and applications of molecular similarity, by Mark A. Johnson and Gera ld M. Maggiora, eds., John Wiley & Sons, New York, 1990, 393 pp. Price: $65.00. J. Comput. Chem. 13, 539 –540 (1992)

    1990

  29. [29]

    & Ramsundar, B

    Chithrananda, S., Grand, G. & Ramsundar, B. ChemBERTa: Large -Scale Self -Supervised Pretrai ning for Molecular Property Prediction. Preprint at https://api.semanticscholar.org/CorpusID:2248 03102 (2020)

    2020

  30. [30]

    DeepGCNs: Can GCNs Go as Deep as CNNs?

    Li, G., Mü ller, M., Thabet, A. & Ghanem, B. DeepGCNs: can GCNs go as deep as CNNs? Pre print at https://doi.org/10.48550/arXiv.1904.03751 (2019)

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1904.03751 1904

  31. [31]

    & Liu, X

    Li, Y., Wan, Y. & Liu, X. Semi -supervised Learning with Graph Convolutional Networks Base d on Hypergraph. Neural Process. Lett. 54, 2629 –2644 (2022)

    2022

  32. [32]

    -H., Liu, Z

    Guo, M. -H., Liu, Z. -N., Mu, T. -J. & Hu, S. -M. Beyond Self -Attention: External Attention Usin g Two Linear Layers for Visual Tasks. IEEE Trans. Pattern Anal. Mach. Intell. 45, 5436 –5447 (2023)

    2023

  33. [33]

    Mé ndez-Lucio, O., Nicolaou, C. A. & Earnshaw, B. MolE: a foundation model for molecular gr aphs using disentangled attention. Nat. Commun. 15, 9431 (2024)

    2024

  34. [34]

    Liu, J. et al. Towards Graph Foundation Models: A Survey and Beyond. Preprint at https://doi. org/10.48550/arXiv.2310.11829 (2023)

    Pith/arXiv arXiv doi:10.48550/arxiv.2310.11829 2023

  35. [35]

    Wu, Z. et al. MoleculeNet: A Benchmark for Molecular Machine Learning. Preprint at https://ar xiv.org/abs/1703.00564 (2018)

    Pith/arXiv arXiv 2018

  36. [36]

    Lu, C. et al. Molecular Property Prediction: A Multilevel Quantum Interactions Modeling Persp ective. Preprint at https://arxiv.org/abs/1906.11081 (2019). 27

    Pith/arXiv arXiv 1906

  37. [37]

    Yang, K. et al. Analyzing Learned Molecular Representations for Property Prediction. J. Chem. Inf. Model. 59, 3370 –3388 (2019)

    2019

  38. [38]

    & Micheli, A

    Errica, F., Podda, M., Bacciu, D. & Micheli, A. A Fair Comparison of Graph Neural Networks for Graph Classification. Preprint at https://arxiv.org/abs/1912.09893 (2022)

    Pith/arXiv arXiv 1912

  39. [39]

    Liu, S., Demirel, M. F. & Liang, Y. N -gram graph: simple unsupervised representation for grap hs, with applications to molecules. Preprint at https://doi.org/10.48550/arXiv.1806.09206 (2019)

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1806.09206 2019

  40. [40]

    Mobley, D. L. & Guthrie, J. P. FreeSolv: a database of experimental and calculated hydration f ree energies, with input files. J. Comput. -Aided Mol. Des. 28, 711 –720 (2014)

    2014

  41. [41]

    Hu, W. et al. Open graph benchmark: datasets for machine learning on graphs. Preprint at http s://doi.org/10.48550/arXiv.2005.00687 (2021)

    Pith/arXiv arXiv doi:10.48550/arxiv.2005.00687 2005

  42. [42]

    Random Forests

    Breiman, L. Random Forests. Mach. Learn. 45, 5 –32 (2001)

    2001

  43. [43]

    & Vapnik, V

    Cortes, C. & Vapnik, V. Support -vector networks. Mach. Learn. 20, 273 –297 (1995)

    1995

  44. [44]

    & Guestrin, C

    Chen, T. & Guestrin, C. XGBoost: A Scalable Tree Boosting System. in Proceedings of the 22 nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining 785–794 (Association for Computing Machinery, New York, NY, USA, 2016). doi:10.1145/2939672.293 9785

    work page doi:10.1145/2939672.293 2016

  45. [45]

    Application of the logistic function to bio -assay

    Berkson, J. Application of the logistic function to bio -assay. J. Am. Stat. Assoc. 39, 357 –365 (2 012)

  46. [46]

    Heid, E. et al. Chemprop: a machine learning package for chemical property prediction. J. Che m. Inf. Model. 64, 9 –17 (2024)

    2024

  47. [47]

    PubChem BioAssay: Quantitative High -Through put Screening for Inhibitors of Human Cytochrome P450 Isoforms 1A2, 2C9, 2C19, 2D6, and 3 A4 (AID: 1851)

    National Center for Biotechnology Information. PubChem BioAssay: Quantitative High -Through put Screening for Inhibitors of Human Cytochrome P450 Isoforms 1A2, 2C9, 2C19, 2D6, and 3 A4 (AID: 1851). (2009)

    2009

  48. [48]

    In Vitro Drug Interaction Studies — Cytochrome P450 Enzyme - and Transporter -Mediated Dru g Interactions Guidance for Industry. (U.S. Food and Drug Administration, 2020)

    2020

  49. [49]

    W., AlRabiah, H., Mostafa, G

    Attwa, M. W., AlRabiah, H., Mostafa, G. A. E., Bakheit, A. H. & Kadi, A. A. Assessment of I n Silico and In Vitro Selpercatinib Metabolic Stability in Human Liver Microsomes Using a Va lidated LC -MS/MS Method. Molecules 28, 2618 (2023)

    2023

  50. [50]

    Sodhi, J. K. & Benet, L. Z. Successful and Unsuccessful Prediction of Human Hepatic Clearanc e for Lead Optimization. J. Med. Chem. 64, 3546 –3559 (2021)

    2021

  51. [51]

    Obach, R. S. et al. The Prediction of Human Pharmacokinetic Parameters from Preclinical and In Vitro Metabolism Data. J. Pharmacol. Exp. Ther. 283, 46 –58 (1997)

    1997

  52. [52]

    Niu, Z. et al. PharmaBench: Enhancing ADMET benchmarks with large language models. Sci. Data 11, 985 (2024)

    2024

  53. [53]

    B., Mathé , E

    Shah, P., Siramshetty, V. B., Mathé , E. & Xu, X. Developing Robust Human Liver Microsomal Stability Prediction Models: Leveraging Inter -Species Correlation with Rat Data. Pharmaceutics 16, 1257 (2024). 28

    2024

  54. [54]

    Ryu, J. Y. et al. PredMS: a random forest model for predicting metabolic stability of drug cand idates in human liver microsomes. Bioinformatics 38, 364 –368 (2022)

    2022

  55. [55]

    Abadi, M. et al. TensorFlow: Large-Scale Machine Learning on Heterogeneous Distributed Syst ems. Preprint at https://arxiv.org/abs/1603.04467 (2016)

    Pith/arXiv arXiv 2016

  56. [56]

    Paszke, A. et al. PyTorch: An Imperative Style, High -Performance Deep Learning Library. Prep rint at https://arxiv.org/abs/1912.01703 (2019)

    Pith/arXiv arXiv 1912

  57. [57]

    & Lenssen, J

    Fey, M. & Lenssen, J. E. Fast Graph Representation Learning with PyTorch Geometric. Preprin t at https://arxiv.org/abs/1903.02428 (2019)

    Pith/arXiv arXiv 1903

  58. [58]

    Scikit -learn: Machine Learning in Python

    Pedregosa, F. Scikit -learn: Machine Learning in Python. J. Mach. Learn. Res. 12, 2825 –2830 (2 011)

  59. [59]

    & Obach, R

    Lombardo, F., Berellini, G. & Obach, R. S. Trend Analysis of a Database of Intravenous Phar macokinetic Parameters in Humans for 1352 Drug Compounds. Drug Metab. Dispos. 46, 1466 – 1477 (2018)

    2018

  60. [60]

    Fang, J. et al. Prediction of Cytochrome P450 Substrates Using the Explainable Multitask Deep Learning Models. Chem. Res. Toxicol. 37, 1535 –1548 (2024)

    2024

  61. [61]

    & Mahanta, H

    Parasar, U., Baruah, O., Saikia, D., Bharali, P. & Mahanta, H. J. Interpretable machine learning and graph attention network based model for predicting PAMPA permeability. J. Mol. Graphics Model. 138, 109050 (2025)

    2025

  62. [62]

    Thakkar, S. et al. Drug-induced liver injury severity and toxicity (DILIst): binary classification of 1279 drugs by human hepatotoxicity. Drug Discovery Today 25, 201 –208 (2020)

    2020

  63. [63]

    & Kalutarage, H

    Wickramasinghe, I. & Kalutarage, H. Naive bayes: applications, variations and vulnerabilities: a review of literature with code snippets for implementation. Soft Comput. 25, 2277 –2293 (2021). Supplementary Material for MEGA-CL: A Molecular Foundation Model for Generalizable AD- MET Prediction through Graph External Attention and Contrastive Learning Ting...

    2021

  64. [64]

    This compound was hermetically sealed, stored away from light at 2 –8 °C, and had a shelf life of 2 years

    Reagents, Chemicals and Hepatic Microsome Source 1.1 Test Articles Oroxylin A (purity, 99.20%; batch number: 20171106) was purchased from China Pharmaceutical University. This compound was hermetically sealed, stored away from light at 2 –8 °C, and had a shelf life of 2 years. Its stock solution was prepared in dimethyl sulfoxide (DMSO) and further serial...

  65. [65]

    For Oroxylin A, eight gradient concentrations including 0.1, 0.2, 0.5, 1, 2, 5, 10, and 30 μM were adopted universally for all eight CYP isoforms tested

    Phase I NADPH-Dependent Microsomal Metabolic Stability Assay 2.1 In Vitro Incubation Conditions A series of concentration gradients was applied for the test compounds in the incubation system. For Oroxylin A, eight gradient concentrations including 0.1, 0.2, 0.5, 1, 2, 5, 10, and 30 μM were adopted universally for all eight CYP isoforms tested. For I cari...

  66. [66]

    The metabolic half-life (t) was calculated using the formula t = ln2 / k

    The first-order elimination constant (k) was determined via linear regression based on the equation ln(Ct) = −kt + ln(C0), where Ct represents the substrate concentration at time t and C0 refers to the ini- tial substrate concentration at t = 0; specifically, k (units: ×10 min) corresponds to the negative slope of the curve plotted with the natural logari...

  67. [67]

    Human CYP450 Isoform Inhibition Assay 3.1 In Vitro Incubation Conditions The standard 200 μL in vitro incubation matrix consisted of 0.05 M PBS (pH 7.2 –7.4), 5 mM MgCl, 1 mM NADPH cofactor, and human liver microsomes at a final protein concentration of 0.2 mg/mL, supplemented with fixed isoform-specific probe substrates and serially diluted test compound...

  68. [68]

    Representative column station- ary phases: Agilent Extend-C18 (1.8 μm), Waters ACQUITY BEH C18 (1.7 μm), with matching guard pre-columns

    UPLC-MS/MS Analytical Methods (SI Standardized Full Chromatography & MS MRM Parameters) All analytes (parent test compounds, CYP probe metabolites, internal standards) were quantified via dedicated UPLC-MS/MS MRM ESI positive/negative ion methods; cone voltage, collision energy, and diagnostic ion transitions were documented in supplementary Table S10. Re...

  69. [69]

    Kinetic parameters (k, t1/2, CLint) and IC50 values were reported as mean values of n=3 biological replicates with standard deviation where applicable

    Data Visualization & Statistical Analysis All metabolic stability curves (relative residual substrate vs incubation time), first -order ln-linear ki- netic plots, and CYP inhibition dose -response curves were generated in GraphPad Prism 8.0. Kinetic parameters (k, t1/2, CLint) and IC50 values were reported as mean values of n=3 biological replicates with ...

    1911