Pith. sign in

REVIEW 3 major objections 5 minor 87 references

VECTOR+ claims that contrastive learning plus Gaussian-mixture sampling turns tiny inhibitor datasets into novel drug-like libraries, with a PD-L1 candidate docking at -17.6 kcal/mol versus -15.4 for the best reference.

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 ↗

VECTOR+ combines contrastive learning and Gaussian mixture sampling to generate novel, synthetically plausible inhibitors from low-data datasets, with improved docking scores over known compounds.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection Worth a look for the curated PD-L1 dataset and the honest framing, but the central GMM procedure is described two incompatible ways and the docking claims need a cleaner baseline. the 3 major comments →

arxiv 2509.00684 v1 pith:6YPPTKAD submitted 2025-08-31 cs.LG cs.AI

Valid Property-Enhanced Contrastive Learning for Targeted Optimization & Resampling for Novel Drug Design

Amartya Banerjee , Somnath Kar , Anirban Pal , Debabrata Maiti This is my paper
classification cs.LG cs.AI
keywords contrastive learningmolecular generationGaussian mixture modellow-data drug discoveryPD-L1 inhibitorskinase inhibitorsmolecular dockingSMILES generation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

VECTOR+ tries to show that data scarcity is not a fatal obstacle for generative molecular design if the latent space is organized by biological function before sampling. It couples a pretrained SMILES encoder with a contrastive loss that pulls same-activity molecules together and pushes different-activity molecules apart, then fits a Gaussian mixture model to the structured embeddings and samples from the component aligned to the desired activity class. The paper reports that, with only 296 PD-L1 training molecules, this produces 8,374 valid novel candidates, 100 with docking scores below -15.0 kcal/mol and a best of -17.6 kcal/mol, surpassing the best reference inhibitor (-15.4). It also reports transfer to kinase inhibitors, where 47 allosteric training examples yield 2,500 new class-conditioned molecules. If these results hold, property-conditioned generation in low-data drug discovery can be achieved with contrastive representation learning plus a simple Gaussian sampler, without large pretraining corpora or heavy reinforcement-learning loops.

Core claim

VECTOR+ claims that low-data molecular generation fails not because datasets are small but because the latent space is not organized by function. Contrastive training on a pretrained SMILES encoder pulls same-activity molecules together and pushes others apart; a Gaussian mixture fitted to the resulting clusters, with components matched to classes by Hungarian alignment, is then treated as an optimal surrogate for sampling. On 296 curated PD-L1 inhibitors this yields 8,374 valid novel molecules; 100 dock below -15.0 kcal/mol and the best is -17.6, beating the best reference inhibitor at -15.4. On 2,056 kinase inhibitors, including 47 allosteric examples, it generates 2,500 molecules per clas

What carries the argument

The load-bearing object is the contrastively structured latent space plus the GMM surrogate. The contrastive loss (pairwise L1 distance with margin m=1) pulls same-class molecules together and pushes different-class molecules apart, making per-class clusters exist. A K-component Gaussian mixture model is then fit by expectation-maximization, and each component is assigned to a property class by maximum-weight bipartite matching using the Hungarian algorithm. The theoretical anchor is Theorem 1: for any distribution with finite mean and covariance, the Gaussian with those moments uniquely minimizes forward KL divergence, so the per-class Gaussian is the best Gaussian surrogate for the unknown

Load-bearing premise

The generation step assumes that after contrastive training each property class occupies one roughly Gaussian, unimodal cluster in the latent space, so sampling from the single aligned Gaussian component yields molecules of the intended class; the paper shows 2D projections but does not quantitatively test unimodality or Gaussianity.

What would settle it

Take the trained encoder's validation-set embeddings for each activity class, run a standard multivariate normality test, and compare a one-component Gaussian per class against a two-component mixture by log-likelihood or BIC; if the class-conditional latent is clearly multimodal, single-Gaussian sampling should produce molecules whose predicted IC50 distribution does not match the high-activity training distribution. A complementary decisive experiment is to synthesize the top-scoring candidate and measure its IC50 in the PD-L1 assay.

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

If this is right

  • With 296 PD-L1 training molecules, the framework generated 8,374 valid unique candidates, 100 of which dock below -15.0 kcal/mol, a threshold only the best reference inhibitor reaches.
  • Starting from 47 allosteric kinase inhibitors, the framework produced 2,500 novel class-conditioned molecules, suggesting the pipeline can operate when a training class is very small.
  • Generated PD-L1 candidates retain the conserved biphenyl pharmacophore while introducing motifs absent from the training set, such as an eight-membered ring, indicating novelty beyond interpolation between known scaffolds.
  • On benchmarking metrics, VECTOR+ matched the training distribution better than JT-VAE and MolGPT on molecular weight and logP, achieved higher Tanimoto similarity to the original ligands, and its top-50 average docking score was better than both baselines.

Where Pith is reading between the lines

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

  • The headline result, docking scores beyond known inhibitors, is a computational forecast rather than a measured biological outcome; the paper itself notes that docking scores do not map directly to IC50, so synthesis and assay of the top candidates would be the decisive test.
  • The method conditions on a coarse median split for PD-L1, so the high-activity class spans a wide range of potencies; coupling the contrastive embedding with a continuous log(IC50) regressor and sampling from an isopotency contour could sharpen target specificity beyond two bins.
  • The generation stage depends heavily on the SMILES decoder, so invalid or repetitive outputs are attributed to decoding rather than representation; a graph-based or fragment-based decoder might preserve the same latent structure while improving validity and synthetic accessibility.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. VECTOR+ is presented as a low-data generative framework that couples property-guided contrastive representation learning with Gaussian-mixture latent sampling. A ChemBERTa encoder with a projection head is trained with a pairwise margin contrastive loss on discrete property labels; the resulting latent space is modeled with a GMM, components are aligned to classes via Hungarian matching, and a GRU decoder maps newly sampled latent vectors to SMILES strings. The method is evaluated on a curated PD-L1 inhibitor set (296 compounds, split at the median of log IC50) and a kinase inhibitor set (2,056 compounds in four binding-mode classes). The headline results are that 100 of 8,374 generated PD-L1 molecules achieve docking scores below -15.0 kcal/mol against PDB 5J89, with the top candidate at -17.6 kcal/mol versus -15.4 kcal/mol for the best reference inhibitor, and that analogous kinase candidates beat known drugs such as brigatinib and sorafenib. Benchmarking against JT-VAE and MolGPT is reported, together with 250 ns MD simulations for selected hits. The paper also contributes a curated PD-L1 dataset and releases code and data.

Significance. If the central claims are sustained, the contribution is useful: a reproducible, data-efficient pipeline for targeted molecular generation in a low-data regime, a new curated PD-L1 dataset, and an external in-silico evaluation pipeline. The information-theoretic theorem in the Methods is correct and cleanly proved, although it is not by itself sufficient to justify the specific GMM sampling procedure. The strongest results, however, rest on two currently unsupported pillars: the consistency between the described GMM procedure and the one actually used, and the fairness of comparing order statistics of a large generated library with a small reference set. Both are fixable with additional analysis and clarification, so the underlying framework remains plausible. The paper should be credited for making code and data available and for benchmarking against standard baselines, but the headline docking claims cannot be accepted as stated.

major comments (3)
  1. [Methods, 'Latent Space Modeling and Targeted Generation' (Eqs. 2-8; Algs. 1-2) vs. Results, 'GMM-Based Reactivity Cluste] The methodology specifies a single global GMM fitted to all embeddings with K=C and Hungarian alignment of components to classes (Eqs. 2-8, Algorithm 1), and Algorithm 2 samples from the aligned component N(μ_k*, Σ_k*). The Results section instead states that for PD-L1 'we trained a GMM on the high-activity cluster' and for kinase 'trained separate GMMs for each cluster.' These are different algorithms, and they are not equivalent in general: only when each class is a well-separated single Gaussian component does a globally fitted, aligned component coincide with the class-conditional distribution. The paper itself acknowledges in the 'Justification for the GMM-based Surrogate' that Theorem 1 'does not imply that EM-fitted mixture components will exactly coincide with the latent class-conditionals.' Since the reported generated molecules are attributed to a specific sampling procedure, t
  2. [Results, 'PD-L1 Ligands' and 'Kinase Ligands'] The comparison of generated molecules with known inhibitors is based on extreme order statistics. For PD-L1, the top 50 of 8,374 generated molecules score between -15.2 and -17.6 kcal/mol, while the reference set of 296 known inhibitors ranges from -12.2 to -15.4 kcal/mol. The maximum of a much larger sample is expected to be more extreme even under the same score distribution, so the abstract's claim that the top candidate 'outperforms known inhibitors' is not established. Similarly, kinase comparisons are made against single reference drugs (brigatinib, sorafenib, etc.) rather than against the distribution of a comparable-sized reference library. Please report distributional comparisons, such as top-k values from equal-sized random subsamples of the generated and reference sets, or enrichment/quantile statistics of generated scores relative to the reference distribution.
  3. [Methods, 'Latent Space Modeling and Targeted Generation' and Appendix 'Model details'] The GMM fitting procedure is underspecified in a way that affects the central sampling claim. The encoder output is 768-dimensional, the PD-L1 training set has n=296, and no covariance parameterization (full, diagonal, tied), regularization, or dimensionality reduction is reported. EM with full 768×768 covariance matrices on 296 points is ill-posed, and the sampled distribution depends critically on the chosen constraint. Please report the exact GMM implementation, including covariance type, regularization, and any dimension reduction, and provide validation of the fitted model (e.g., BIC, held-out log-likelihood, or a two-sample test between GMM samples and true class embeddings).
minor comments (5)
  1. [Eq. (1)] The text says the final loss is computed by averaging over the upper triangle, but the displayed sum is over all ordered pairs i≠j. Please define the normalization explicitly to avoid double-counting ambiguity.
  2. [Algorithm 2] The optional hill-climbing procedure is described in the algorithm but the Results do not state whether it was used in any of the reported experiments. Please clarify whether the 8,374 PD-L1 and 2,500-per-class kinase molecules were generated with or without hill-climbing; if it was not used, say so explicitly.
  3. [Table 3] The column heading 'Avg. ΔR2 = (ΔR2_Contr − ΔR2_Base)' is confusingly not the difference of R² values between embedding types but a difference of differences. Please clarify the notation and the computation.
  4. [Benchmarking section] For JT-VAE and MolGPT, please report the number of valid molecules used for docking evaluation, the training epochs, and whether the same data split and preprocessing were used. This is important for interpreting the validity and docking comparisons.
  5. [Figure 3 and Appendix figures] Docking scores are displayed as 'D.S.' numbers without units in the figures; the units (kcal/mol) appear only in the text. Please add units to the figure labels for clarity.

Circularity Check

0 steps flagged

No material circularity: the docking-based central claim is externally evaluated and was not used in training; self-citations are peripheral and non-load-bearing.

full rationale

The central claim—that VECTOR+ generates molecules with docking scores surpassing known inhibitors against PD-L1 (5J89) and kinase targets—does not reduce to the method's inputs. The contrastive encoder is trained with Eq. 1 using SMILES and IC50-derived/property class labels; the GMM is fit to the resulting latent embeddings (Eqs. 2–8); the decoder maps sampled latent vectors to SMILES. Docking scores from Glide XP are computed only after generation and are not part of any loss, assignment, or GMM-fitting objective, so there is no fitted-input-called-prediction loop. The GMM surrogate is justified by an internally proven Theorem 1, and the paper explicitly concedes that this theorem 'does not imply that EM-fitted mixture components will exactly coincide with the latent class-conditionals' (Theoretical Justification section), which is an honest limitation rather than a circular premise. The manuscript's self-citations (refs. 61 and 87) are used only as contextual support in Related Work and an appendix comparison, not as load-bearing evidence for the docking result. The inconsistency between the global-GMM-with-Hungarian-alignment procedure in Methods/Algorithm 1 and the per-cluster GMM description in Results is a reproducibility/consistency concern, but in neither described procedure is the docking score used to fit or select the sampler. Therefore no circular step is established.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 0 invented entities

The central result rests on hand-set hyperparameters (m, K, decoding temperature, hill-climbing parameters) and on domain assumptions about latent geometry and the validity of docking as an evaluation metric. The paper's theoretical justification (Theorem 1) is a classical result and does not uniquely fix the GMM procedure actually used.

free parameters (5)
  • Contrastive margin m = 1.0
    Set by hand in the contrastive loss (Eq. 1).
  • Number of GMM components K = C (2 for PD-L1, 4 for kinase)
    Set equal to the number of property classes; no model selection is performed.
  • Decoding temperature = not specified
    Used in temperature-controlled softmax sampling in the GRU decoder; value not reported.
  • Hill-climbing hyperparameters (H, k, alpha) = not specified
    Used in Algorithm 2, but no default values are given in the paper.
  • Median split for PD-L1 activity bins = median log(IC50)
    Defines the two activity classes; a different split would change the contrastive labels and all downstream results.
axioms (4)
  • domain assumption Contrastive loss with margin creates well-separated, unimodal class clusters in latent space
    The GMM surrogate and Hungarian alignment rely on this; the paper shows UMAP separation but does not test unimodality.
  • domain assumption Docking score (Glide XP) is a meaningful proxy for inhibitor potential
    The central evaluation metric; the paper itself states there is no direct mathematical correlation between docking scores and biological outcomes.
  • domain assumption Pretrained ChemBERTa embeddings provide a suitable base representation for low-data optimization
    Relies on transfer learning from a large SMILES corpus to a small domain-specific dataset.
  • domain assumption Sampling from a Gaussian fitted to latent embeddings and decoding with a GRU yields valid, novel, target-relevant molecules
    The decoder is trained on the same small dataset; coverage of the latent space outside training regions is not guaranteed.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Valid Property-Enhanced Contrastive Learning for Targeted Optimization & Resampling for Novel Drug Design." pith.science (2026). https://pith.science/paper/6YPPTKAD

@misc{pith2026250900684,
  author       = {Pith},
  title        = {Pith review of: Valid Property-Enhanced Contrastive Learning for Targeted Optimization & Resampling for Novel Drug Design},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6YPPTKAD}},
  note         = {Machine review of arXiv:2509.00684}
}
Share X Bluesky LinkedIn Reddit HN
abstract

Efficiently steering generative models toward pharmacologically relevant regions of chemical space remains a major obstacle in molecular drug discovery under low-data regimes. We present VECTOR+: Valid-property-Enhanced Contrastive Learning for Targeted Optimization and Resampling, a framework that couples property-guided representation learning with controllable molecule generation. VECTOR+ applies to both regression and classification tasks and enables interpretable, data-efficient exploration of functional chemical space. We evaluate on two datasets: a curated PD-L1 inhibitor set (296 compounds with experimental $IC_{50}$ values) and a receptor kinase inhibitor set (2,056 molecules by binding mode). Despite limited training data, VECTOR+ generates novel, synthetically tractable candidates. Against PD-L1 (PDB 5J89), 100 of 8,374 generated molecules surpass a docking threshold of $-15.0$ kcal/mol, with the best scoring $-17.6$ kcal/mol compared to the top reference inhibitor ($-15.4$ kcal/mol). The best-performing molecules retain the conserved biphenyl pharmacophore while introducing novel motifs. Molecular dynamics (250 ns) confirm binding stability (ligand RMSD < $2.5$ angstroms). VECTOR+ generalizes to kinase inhibitors, producing compounds with stronger docking scores than established drugs such as brigatinib and sorafenib. Benchmarking against JT-VAE and MolGPT across docking, novelty, uniqueness, and Tanimoto similarity highlights the superior performance of our method. These results position our work as a robust, extensible approach for property-conditioned molecular design in low-data settings, bridging contrastive learning and generative modeling for reproducible, AI-accelerated discovery.

Figures

Figures reproduced from arXiv: 2509.00684 by Amartya Banerjee, Anirban Pal, Debabrata Maiti, Somnath Kar.

Figure 1
Figure 1. Figure 1: Workflow for GMM-based sampling and SMILES generation. Contrastive embeddings from ChemBERTa are clustered using a GMM. vectors are sampled per cluster, decoded into SMILES, filtered using RDKit, and saved for downstream applications. Methodology The VECTOR+ framework is a pipeline designed to learn property-centric molecular representations and leverage them for targeted de novo generation. The methodolog… view at source ↗
Figure 2
Figure 2. Figure 2: This figure shows dimensionality reduction from 768-dimensional ChemBERTa embeddings to 2D UMAP space. (A-B) PD-L1 inhibitors classified into two activity groups based on log(IC50) values (Class 1: below median; Class 2: above median). In base embeddings (A), molecular representations show significant overlap between activity classes, while contrastive learning (B) successfully separates compounds into dis… view at source ↗
Figure 3
Figure 3. Figure 3: In silico studies were performed on the generated molecules from both the PD-L1 and kinase datasets. This figure displays the top-ranked candidate structures from each target along with their corresponding docking scores (in kcal/mol). The consistently high docking scores highlight the model’s ability to generate ligands with strong predicted binding affinity. Notably, structural analysis of the PD-L1 cand… view at source ↗
Figure 4
Figure 4. Figure 4: This figure shows the comparative evaluation of molecular properties and generative performance between the models. (A) Molecular weight distributions of original PD-L1 ligands and molecules generated by MolGPT, JT-VAE, and VECTOR+, showing VECTOR+’s better alignment with the original distribution. (B) logP distribution comparison, where VECTOR+ captures the lipophilicity profile of the original molecules … view at source ↗
Figure 5
Figure 5. Figure 5: This figure showed molecular docking and MD simulation of top-ranked generated molecules. A total of 8,374 AI-generated molecules were virtually screened against PD-L1 (PDB ID: 5J89). (A) and (B) depict 2D interaction diagrams of two top-scoring candidates, illustrating multiple stabilizing interactions with key amino acid residues within the binding pocket. (C) and (D) present the 250 ns molecular dynamic… view at source ↗
Figure 6
Figure 6. Figure 6: Molecular docking and MD simulations of generated kinase inhibitors. A total of 2,500 novel molecules were generated for each inhibitor class and subjected to virtual screening. (A) and (B) show 2D interaction diagrams of the top-ranked molecules from the Type I and Type II inhibitor classes, respectively, highlighting key interactions with residues in the kinase active site. (C) and (D) present 250 ns mol… view at source ↗
Figure 7
Figure 7. Figure 7: The figure showed the data segmentation and preprocessing applied prior to machine learning. (A) PD-L1 inhibitors dataset was categorized based on their biological activity. (B) Kinase inhibitor dataset was segmented according to the inhibitors’ mode of binding to the target protein. (C) The raw IC50 values of the PD-L1 inhibitors exhibited a right-skewed distribution. (D) Log transformation of the IC50 va… view at source ↗
Figure 8
Figure 8. Figure 8: A total of 8,374 molecules were generated using our method for the PD-L1 dataset. In silico studies on these compounds identified several potential top hits. This figure presents the structures of a few representative examples. 22/26 [PITH_FULL_IMAGE:figures/full_fig_p022_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: A total of 2500 molecules were generated using our method for each class of the Kinase dataset. In silico studies on these compounds identified several potential top hits. This figure presents the structures of a few representative examples for Type I and Type II classes. 23/26 [PITH_FULL_IMAGE:figures/full_fig_p023_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: This figure presents the structures of a few representative examples for Type I1/2 and Allosteric inhibitors 24/26 [PITH_FULL_IMAGE:figures/full_fig_p024_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: MAE comparison across 7 regressors. For each model, blue bars show the mean absolute error (MAE) using base embeddings, and red bars show MAE using contrastive embeddings; error bars denote one standard deviation over 50 seeds. 25/26 [PITH_FULL_IMAGE:figures/full_fig_p025_11.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

87 extracted references · 72 canonical work pages · 5 internal anchors

  1. [1]

    J., McKee, M

    Wouters, O. J., McKee, M. & Luyten, J. Estimated research and development investment needed to bring a new medicine to market, 2009-2018.Jama 323, 844–853 (2020)

    2009

  2. [2]

    S., Babu, C

    Dara, S., Dhamercherla, S., Jadav, S. S., Babu, C. M. & Ahsan, M. J. Machine learning in drug discovery: a review. Artif. intelligence review55, 1947–1999 (2022)

    1947

  3. [3]

    & Blaschke, T

    Chen, H., Engkvist, O., Wang, Y., Olivecrona, M. & Blaschke, T. The rise of deep learning in drug discovery. Drug discovery today23, 1241–1250 (2018)

    2018

  4. [4]

    Artificial intelligence to deep learning: machine intelligence approach for drug discovery.Mol

    Gupta, R.et al. Artificial intelligence to deep learning: machine intelligence approach for drug discovery.Mol. diversity 25, 1315–1360 (2021)

    2021

  5. [5]

    & Chakrabarti, S

    Biswas, N. & Chakrabarti, S. Artificial intelligence (ai)-based systems biology approaches in multi-omics data analysis of cancer.Front. oncology10, 588221 (2020)

    2020

  6. [6]

    Pitt, W. R.et al. Real-world applications and experiences of ai/ml deployment for drug discovery (2025)

    2025

  7. [7]

    & Kar, S

    Yang, S. & Kar, S. Application of artificial intelligence and machine learning in early detection of adverse drug reactions (adrs) and drug-induced toxicity.Artif. Intell. Chem.1, 100011 (2023)

    2023

  8. [8]

    & Sil, P

    Sinha, K., Ghosh, N. & Sil, P. C. A review on the recent applications of deep learning in predictive drug toxicological studies. Chem. Res. Toxicol.36, 1174–1205 (2023)

    work page 2023

  9. [9]

    Tran, T. T. V., Surya Wibowo, A., Tayara, H. & Chong, K. T. Artificial intelligence in drug toxicity prediction: recent advances, challenges, and future perspectives.J. chemical information modeling63, 2628–2643 (2023)

    work page 2023

  10. [10]

    & Lavecchia, A

    Gangwal, A. & Lavecchia, A. Unlocking the potential of generative ai in drug discovery.Drug Discov. Today 103992 (2024)

    work page 2024

  11. [11]

    P., Welling, M.et al

    Kingma, D. P., Welling, M.et al. Auto-encoding variational bayes (2013)

    work page 2013

  12. [12]

    Goodfellow, I. J.et al. Generative adversarial nets.Adv. neural information processing systems27 (2014)

    work page 2014

  13. [13]

    & Jereesh, A

    Haroon, S., Hafsath, C. & Jereesh, A. Generative pre-trained transformer (gpt) based model with relative attention for de novo drug design.Comput. Biol. Chem.106, 107911 (2023)

    work page 2023

  14. [14]

    Dou, B. et al. Machine learning methods for small data challenges in molecular science.Chem. Rev. 123, 8736–8780 (2023)

    work page 2023

  15. [15]

    Parvatikar, P. P.et al. Artificial intelligence: Machine learning approach for screening large database and drug discovery. Antivir. Res.220, 105740 (2023)

    work page 2023

  16. [16]

    Tingle, B. I.et al. Zinc-22- a free multi-billion-scale database of tangible compounds for ligand discovery.J. chemical information modeling63, 1166–1176 (2023)

    work page 2023

  17. [17]

    van Tilborg, D.et al.Deep learning for low-data drug discovery: Hurdles and opportunities.Curr. Opin. Struct. Biol. 86, 102818 (2024)

    work page 2024

  18. [18]

    Zaorsky, N. G.et al. Causes of death among cancer patients.Annals oncology28, 400–407 (2017)

    work page 2017

  19. [19]

    & Soerjomataram, I

    Bray, F., Laversanne, M., Weiderpass, E. & Soerjomataram, I. The ever-increasing importance of cancer as a leading cause of premature death worldwide.Cancer 127, 3029–3030 (2021)

    work page 2021

  20. [20]

    Current challenges in cancer treatment.Clin

    Zugazagoitia, J.et al. Current challenges in cancer treatment.Clin. therapeutics 38, 1551–1566 (2016)

    work page 2016

  21. [21]

    Combining immunotherapy and radiotherapy for cancer treatment: current challenges and future directions

    Wang, Y.et al. Combining immunotherapy and radiotherapy for cancer treatment: current challenges and future directions. Front. pharmacology9, 185 (2018)

    work page 2018

  22. [22]

    From chemotherapy to biological therapy: A review of novel concepts to reduce the side effects of systemic cancer treatment.Int

    Schirrmacher, V. From chemotherapy to biological therapy: A review of novel concepts to reduce the side effects of systemic cancer treatment.Int. journal oncology54, 407–419 (2019)

    work page 2019

  23. [23]

    M., Komninos, D

    van den Boogaard, W. M., Komninos, D. S. & Vermeij, W. P. Chemotherapy side-effects: not all dna damage is equal. Cancers 14, 627 (2022)

    work page 2022

  24. [24]

    Wang, X., Teng, F., Kong, L. & Yu, J. Pd-l1 expression in human cancers and its association with clinical outcomes. OncoTargets therapy5023–5039 (2016). 17/26

    work page 2016

  25. [25]

    Pd-l1 promotes tumor growth and progression by activating wip andβ-catenin signaling pathways and predicts poor prognosis in lung cancer.Cell death & disease11, 506 (2020)

    Yu, W.et al. Pd-l1 promotes tumor growth and progression by activating wip andβ-catenin signaling pathways and predicts poor prognosis in lung cancer.Cell death & disease11, 506 (2020)

    work page 2020

  26. [26]

    M., Rennert, P

    Mahoney, K. M., Rennert, P. D. & Freeman, G. J. Combination cancer immunotherapy and new immunomodu- latory targets. Nat. reviews Drug discovery14, 561–584 (2015)

    work page 2015

  27. [27]

    Yi, M., Niu, M., Xu, L., Luo, S. & Wu, K. Regulation of pd-l1 expression in the tumor microenvironment.J. hematology & oncology14, 1–13 (2021)

    work page 2021

  28. [28]

    & Ohmori, T

    Yamaoka, T., Kusumoto, S., Ando, K., Ohba, M. & Ohmori, T. Receptor tyrosine kinase-targeted cancer therapy. Int. journal molecular sciences19, 3491 (2018)

    work page 2018

  29. [29]

    Therapeutic advances of targeting receptor tyrosine kinases in cancer.Signal Transduct

    Tomuleasa, C.et al. Therapeutic advances of targeting receptor tyrosine kinases in cancer.Signal Transduct. Target. Ther.9, 201 (2024)

    work page 2024

  30. [30]

    Hsu, J. L. & Hung, M.-C. The role of her2, egfr, and other receptor tyrosine kinases in breast cancer.Cancer Metastasis Rev. 35, 575–588 (2016)

    work page 2016

  31. [31]

    Regulatory mechanisms of pd-1/pd-l1 in cancers.Mol

    Lin, X.et al. Regulatory mechanisms of pd-1/pd-l1 in cancers.Mol. Cancer 23, 108 (2024)

    work page 2024

  32. [32]

    & Lovly, C

    Du, Z. & Lovly, C. M. Mechanisms of receptor tyrosine kinase activation in cancer.Mol. cancer 17, 1–13 (2018)

    work page 2018

  33. [33]

    & Pichika, M

    Mak, K.-K., Wong, Y.-H. & Pichika, M. R. Artificial intelligence in drug discovery and development.Drug discovery evaluation: safety pharmacokinetic assays1461–1498 (2024)

    work page 2024

  34. [34]

    H., Kogej, T., Tyrchan, C

    Segler, M. H., Kogej, T., Tyrchan, C. & Waller, M. P. Generating focused molecule libraries for drug discovery with recurrent neural networks.ACS central science4, 120–131 (2018)

    work page 2018

  35. [35]

    Bjerrum, E. J. & Threlfall, R. Molecular generation with recurrent neural networks (rnns).arXiv preprint arXiv:1705.04612 (2017)

    Pith/arXiv arXiv 2017

  36. [36]

    & Chen, H

    Olivecrona, M., Blaschke, T., Engkvist, O. & Chen, H. Molecular de-novo design through deep reinforcement learning. J. cheminformatics 9, 1–14 (2017)

    work page 2017

  37. [37]

    Gómez-Bombarelli, R. et al. Automatic chemical design using a data-driven continuous representation of molecules. ACS central science4, 268–276 (2018)

    work page 2018

  38. [38]

    & Sawant, S

    Rathod, V., Gadilohar, J., Pawar, S., Joshi, A. & Sawant, S. Unlocking new possibilities in drug discovery: A gan-based approach. InArtificial Intelligence-based Healthcare Systems, 135–144 (Springer, 2023)

    work page 2023

  39. [39]

    & Wang, F

    Deng, J., Yang, Z., Ojima, I., Samaras, D. & Wang, F. Artificial intelligence in drug discovery: applications and techniques. Briefings Bioinforma. 23 (2022)

    work page 2022

  40. [40]

    Lim, J., Ryu, S., Kim, J. W. & Kim, W. Y. Molecular generative model based on conditional variational autoencoder for de novo molecular design.J. cheminformatics 10, 1–9 (2018)

    work page 2018

  41. [41]

    & Cho, K

    Kang, S. & Cho, K. Conditional molecular design with deep generative models.J. chemical information modeling 59, 43–52 (2018)

    work page 2018

  42. [42]

    & Jaakkola, T

    Jin, W., Barzilay, R. & Jaakkola, T. Junction tree variational autoencoder for molecular graph generation. In International conference on machine learning, 2323–2332 (PMLR, 2018)

    work page 2018

  43. [43]

    & Zhu, S

    Liao, Z., Xie, L., Mamitsuka, H. & Zhu, S. Sc2mol: a scaffold-based two-step molecule generator with variational autoencoder and transformer.Bioinformatics 39, btac814 (2023)

    work page 2023

  44. [44]

    & Liu, X

    Liu, H., Tian, S. & Liu, X. Phenotypic profile-informed generation of drug-like molecules via dual-channel variational autoencoders. arXiv preprint arXiv:2506.02051(2025)

    Pith/arXiv arXiv 2025

  45. [45]

    Reinforced adversarial neural computer for de novo molecular design.J

    Putin, E.et al. Reinforced adversarial neural computer for de novo molecular design.J. chemical information modeling 58, 1194–1204 (2018)

    work page 2018

  46. [46]

    Zhavoronkov, A. et al. Deep learning enables rapid identification of potent ddr1 kinase inhibitors. Nat. biotechnology 37, 1038–1040 (2019)

    work page 2019

  47. [47]

    & Grisoni, F

    Rossen, L., Sirockin, F., Schneider, N. & Grisoni, F. Scaffold hopping with generative reinforcement learning.J. Chem. Inf. Model.(2024)

    work page 2024

  48. [48]

    Direct steering of de novo molecular generation with descriptor conditional recurrent neural networks

    Kotsias, P.-C.et al. Direct steering of de novo molecular generation with descriptor conditional recurrent neural networks. Nat. Mach. Intell. 2, 254–265 (2020)

    work page 2020

  49. [49]

    Transformer-based generative model accelerating the development of novel braf inhibitors.ACS omega 6, 33864–33873 (2021)

    Yang, L.et al. Transformer-based generative model accelerating the development of novel braf inhibitors.ACS omega 6, 33864–33873 (2021). 18/26

    work page 2021

  50. [50]

    Loeffler, H. H.et al. Reinvent 4: Modern ai–driven generative molecule design.J. Cheminformatics 16, 20 (2024)

    work page 2024

  51. [51]

    & Lee, S.-S

    Chakraborty, C., Bhattacharya, M. & Lee, S.-S. Artificial intelligence enabled chatgpt and large language models in drug target discovery, drug discovery, and development.Mol. therapy Nucleic acids33, 866–868 (2023)

    work page 2023

  52. [52]

    Lee, N.et al.Rag-enhanced collaborative llm agents for drug discovery.arXiv preprint arXiv:2502.17506(2025)

    arXiv 2025

  53. [53]

    arXiv preprint arXiv:2411.15692(2024)

    Liu, S.et al.Drugagent: Automating ai-aided drug discovery programming through llm multi-agent collaboration. arXiv preprint arXiv:2411.15692(2024)

    Pith/arXiv arXiv 2024

  54. [54]

    & Priyakumar, U

    Bagal, V., Aggarwal, R., Vinod, P. & Priyakumar, U. D. Molgpt: molecular generation using a transformer- decoder model. J. chemical information modeling62, 2064–2076 (2021)

    work page 2064

  55. [55]

    & Coley, C

    Gao, W., Raghavan, P., Shprints, R. & Coley, C. W. Revealing the relationship between publication bias and chemical reactivity with contrastive learning.J. Am. Chem. Soc.147, 8959–8968 (2025)

    work page 2025

  56. [56]

    Shrivastava, A. D. & Kell, D. B. Fragnet, a contrastive learning-based transformer model for clustering, interpreting, visualizing, and navigating chemical space.Molecules 26, 2065 (2021)

    work page 2065

  57. [57]

    & Zhang, Q

    Qian, Y., Shi, M. & Zhang, Q. Consmi: contrastive learning in the simplified molecular input line entry system helps generate better molecules.Molecules 29, 495 (2024)

    work page 2024

  58. [58]

    Xu, M. et al. Geodiff: A geometric diffusion model for molecular conformation generation.arXiv preprint arXiv:2203.02923 (2022)

    Pith/arXiv arXiv 2022

  59. [59]

    A structure-based framework for selective inhibitor design and optimization.Commun

    Zou, Y.et al. A structure-based framework for selective inhibitor design and optimization.Commun. Biol. 8, 422 (2025)

    work page 2025

  60. [60]

    Levy, A. et al. Solving inverse problems in protein space using diffusion-based priors. arXiv preprint arXiv:2406.04239 (2024)

    Pith/arXiv arXiv 2024

  61. [61]

    Adaptive Multimodal Protein Plug-and-Play with Diffusion-Based Priors

    Banerjee, A., Xu, X., Moosmüller, C. & Lee, H. Adaptive multimodal protein plug-and-play with diffusion-based priors. arXiv preprint arXiv:2507.21260(2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  62. [62]

    & Ramsundar, B

    Chithrananda, S., Grand, G. & Ramsundar, B. Chemberta: large-scale self-supervised pretraining for molecular property prediction. arXiv preprint arXiv:2010.09885(2020)

    Pith/arXiv arXiv 2010

  63. [63]

    Algorithms for the assignment and transportation problems.J

    Munkres, J. Algorithms for the assignment and transportation problems.J. society for industrial applied mathematics 5, 32–38 (1957)

    work page 1957

  64. [64]

    & Hoi, S

    Li, J., Zhou, P., Xiong, C. & Hoi, S. C. Prototypical contrastive learning of unsupervised representations.arXiv preprint arXiv:2005.04966 (2020)

    Pith/arXiv arXiv 2005

  65. [65]

    & LeCun, Y

    Balestriero, R. & LeCun, Y. Contrastive and non-contrastive self-supervised learning recover global and local spectral embedding methods. In Koyejo, S.et al.(eds.) Advances in Neural Information Processing Systems, vol. 35, 26671–26685 (Curran Associates, Inc., 2022)

    work page 2022

  66. [66]

    & Sanghavi, S

    Bansal, P., Kavis, A. & Sanghavi, S. Understanding self-supervised learning via gaussian mixture models.arXiv preprint arXiv:2411.03517 (2024)

    work page arXiv 2024

  67. [67]

    Discovery of small-molecule inhibitors of the pd-1/pd-l1 axis that promote pd-l1 internalization and degradation

    Wang, T.et al. Discovery of small-molecule inhibitors of the pd-1/pd-l1 axis that promote pd-l1 internalization and degradation. J. medicinal chemistry65, 3879–3893 (2022)

    work page 2022

  68. [68]

    & Feng, Z

    Yang, Y., Wang, K., Chen, H. & Feng, Z. Design, synthesis, evaluation, and sar of 4-phenylindoline derivatives, a novel class of small-molecule inhibitors of the programmed cell death-1/programmed cell death-ligand 1 (pd-1/pd-l1) interaction. Eur. J. Medicinal Chem.211, 113001 (2021)

    work page 2021

  69. [69]

    Guo, J.et al. Design, synthesis, and biological evaluation of linear aliphatic amine-linked triaryl derivatives as potent small-molecule inhibitors of the programmed cell death-1/programmed cell death-ligand 1 interaction with promising antitumor effects in vivo.J. Medicinal Chem.63, 13825–13850 (2020)

    work page 2020

  70. [70]

    Design, synthesis, and evaluation of o-(biphenyl-3-ylmethoxy) nitrophenyl derivatives as pd-1/pd-l1 inhibitors with potent anticancer efficacy in vivo.J

    OuYang, Y.et al. Design, synthesis, and evaluation of o-(biphenyl-3-ylmethoxy) nitrophenyl derivatives as pd-1/pd-l1 inhibitors with potent anticancer efficacy in vivo.J. Medicinal Chem.64, 7646–7666 (2021)

    work page 2021

  71. [71]

    Design, synthesis, and pharmacological evaluation of biaryl-containing pd-1/pd-l1 interaction inhibitors bearing a unique difluoromethyleneoxy linkage.J

    Song, Z.et al. Design, synthesis, and pharmacological evaluation of biaryl-containing pd-1/pd-l1 interaction inhibitors bearing a unique difluoromethyleneoxy linkage.J. Medicinal Chem.64, 16687–16702 (2021)

    work page 2021

  72. [72]

    Discovery of [1, 2, 4] triazolo [4, 3-a] pyridines as potent inhibitors targeting the programmed cell death-1/programmed cell death-ligand 1 interaction.J

    Qin, M.et al. Discovery of [1, 2, 4] triazolo [4, 3-a] pyridines as potent inhibitors targeting the programmed cell death-1/programmed cell death-ligand 1 interaction.J. medicinal chemistry62, 4703–4715 (2019). 19/26

    work page 2019

  73. [73]

    Discovery of 4-arylindolines containing a thiazole moiety as potential antitumor agents inhibiting the programmed cell death-1/programmed cell death-ligand 1 interaction.J

    Qin, M.et al. Discovery of 4-arylindolines containing a thiazole moiety as potential antitumor agents inhibiting the programmed cell death-1/programmed cell death-ligand 1 interaction.J. medicinal chemistry64, 5519–5534 (2021)

    work page 2021

  74. [74]

    & Bajorath, J

    Miljkovic, F., Rodríguez-Pérez, R. & Bajorath, J. Machine learning models for accurate prediction of kinase inhibitors with different binding modes.J. medicinal chemistry63, 8738–8748 (2019)

    work page 2019

  75. [75]

    P., Kooistra, A

    Van Linden, O. P., Kooistra, A. J., Leurs, R., De Esch, I. J. & De Graaf, C. Klifs: a knowledge-based structural database to navigate kinase–ligand interaction space.J. medicinal chemistry57, 249–277 (2014)

    work page 2014

  76. [76]

    J.et al.Klifs: a structural kinase-ligand interaction database.Nucleic acids research44, D365–D371 (2016)

    Kooistra, A. J.et al.Klifs: a structural kinase-ligand interaction database.Nucleic acids research44, D365–D371 (2016)

    work page 2016

  77. [77]

    Gavrin, L. K. & Saiah, E. Approaches to discover non-atp site kinase inhibitors.MedChemComm 4, 41–51 (2013)

    work page 2013

  78. [78]

    & Gray, N

    Liu, Y. & Gray, N. S. Rational design of inhibitors that bind to inactive kinase conformations.Nat. chemical biology 2, 358–364 (2006)

    work page 2006

  79. [79]

    Koeberle, S. C.et al. Skepinone-l is a selective p38 mitogen-activated protein kinase inhibitor.Nat. chemical biology 8, 141–143 (2012)

    work page 2012

  80. [80]

    & Héberger, K

    Bajusz, D., Rácz, A. & Héberger, K. Why is tanimoto index an appropriate choice for fingerprint-based similarity calculations? J. cheminformatics 7, 20 (2015)

    work page 2015

Showing first 80 references.

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.