Pith. sign in

REVIEW 3 major objections 2 minor 12 references

CombiMOTS: Combinatorial Multi-Objective Tree Search for Dual-Target Molecule Generation

T0 review · 3 major / 2 minor · reviewed 2026-05-08 · grok-4.3

Challenge this review Download PDF Read on arXiv ↗

Pith's one-line read A Pareto Monte Carlo Tree Search generates dual-target molecules by exploring synthesizable fragments under vectorized multi-objective constraints.

desk verdict CombiMOTS combines Pareto MCTS with synthesizable fragments and vectorized constraints for dual-target generation, but its results stay in silico with docking as the main evidence. read the letter →

arxiv 2604.23307 v1 submitted 2026-04-25 cs.LG cs.AI

Thibaud Southiratn , Bonil Koo , Yijingxiu Lu , Sun Kim This is my paper · ORCID
classification cs.LGcs.AI
keywords dual-targetmoleculegenerationmulti-objectiveoptimizationMonteCarlotreesearchmolecularsynthesizablefragmentsdrugdiscoveryPareto
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

Dual-target molecule generation aims to find compounds that engage two proteins simultaneously for better therapeutic effects. Existing methods often collapse multiple goals into one combined score and overlook how molecules would actually be synthesized. CombiMOTS instead runs a tree search over combinations of chemical fragments while maintaining separate objectives for binding affinity and physical properties. Experiments on real databases show the resulting molecules achieve strong docking scores to both targets, greater variety among candidates, and more balanced drug-like traits.

What carries the argument

Pareto Monte Carlo Tree Search (PMCTS) that maintains non-dominated solution sets across vectorized objectives during combinatorial fragment assembly.

What would settle it

Synthesize the highest-scoring molecules produced by the method and measure their actual binding affinities to both target proteins in laboratory assays.

Watch

Extended reading notes

Core claim

CombiMOTS is a Pareto Monte Carlo Tree Search framework that generates dual-target molecules by exploring a synthesizable fragment space while employing vectorized optimization constraints to encapsulate target affinity and physicochemical properties, producing novel compounds with high docking scores, enhanced diversity, and balanced pharmacological characteristics.

Load-bearing premise

Computational docking scores and physicochemical calculations reliably indicate real biological activity, and the fragment library adequately covers practical drug-like molecules.

Editorial extensions

If this is right

  • Multi-objective problems in molecule design can be addressed without reducing them to single scalar values.
  • Synthetic feasibility becomes part of the search rather than a separate post-processing step.
  • Generated sets exhibit greater structural diversity while still satisfying multiple simultaneous constraints.

Reading between the lines

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

  • The approach could extend to generating molecules with more than two targets or additional constraints such as toxicity avoidance.
  • Embedding synthesis awareness during search may shorten the timeline between computational design and experimental testing.
  • Maintaining Pareto fronts rather than scalarized scores may help surface unexpected but useful trade-offs in polypharmacology.
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 / 2 minor

Summary. The manuscript proposes CombiMOTS, a Pareto Monte Carlo Tree Search (PMCTS) framework for dual-target molecule generation. It explores a synthesizable fragment space while employing vectorized optimization constraints to balance target affinities and physicochemical properties. Experiments on real-world databases are reported to show that CombiMOTS produces novel dual-target molecules with high docking scores, enhanced diversity, and balanced pharmacological characteristics.

Significance. If the results hold under more rigorous validation, CombiMOTS could advance dual-target drug discovery by addressing multi-objective trade-offs without scalarization and by integrating synthetic accessibility into the generative process. The public release of code and data is a clear strength supporting reproducibility.

major comments (3)
  1. [Abstract and Experimental Results] Abstract and Experimental Results: The headline claims rest on docking scores and computed physicochemical properties as proxies for actual dual-target binding and pharmacological utility. Docking is known to produce false positives (particularly in multi-objective settings), yet no orthogonal validation (MD simulations, SPR, or wet-lab assays) or retrosynthetic feasibility checks beyond internal rules are described; this directly undermines the assertion that the molecules are 'useful' for drug discovery.
  2. [Methods] Methods: The precise mechanism by which vectorized constraints are enforced inside the PMCTS (e.g., how Pareto dominance is maintained across the affinity/property vector during node expansion and selection) is not specified with sufficient algorithmic detail or pseudocode, making it impossible to assess whether the reported diversity and balance improvements arise from the method itself or from post-hoc filtering.
  3. [Results] Results: Quantitative claims of 'enhanced diversity' and 'high docking scores' are presented without reported statistical tests, error bars across multiple runs, or explicit baseline implementations; this prevents evaluation of whether the improvements are significant or merely artifacts of the chosen fragment space and scoring functions.
minor comments (2)
  1. [Methods] The notation used for the vectorized objective function and constraint set should be introduced with an explicit equation early in the Methods section to improve readability.
  2. [Figures] Figure captions describing molecule visualizations or Pareto fronts would benefit from additional labels indicating the specific targets and property values for each example.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the detailed and constructive feedback. We address each major comment below and have revised the manuscript accordingly to improve clarity, rigor, and balance in the presentation of results.

read point-by-point responses

  1. Referee: [Abstract and Experimental Results] The headline claims rest on docking scores and computed physicochemical properties as proxies for actual dual-target binding and pharmacological utility. Docking is known to produce false positives (particularly in multi-objective settings), yet no orthogonal validation (MD simulations, SPR, or wet-lab assays) or retrosynthetic feasibility checks beyond internal rules are described; this directly undermines the assertion that the molecules are 'useful' for drug discovery.

    Authors: We agree that docking scores and computed properties are computational proxies subject to false positives and do not substitute for experimental validation. The manuscript is a computational study focused on the generative algorithm and its multi-objective search capabilities. In the revised version we have added an explicit Limitations section that acknowledges these points, tones down claims from 'useful' to 'promising candidates warranting further experimental investigation', and discusses the role of the fragment-based retrosynthetic constraints. We have also clarified that additional MD or wet-lab validation lies outside the current scope but would be a natural next step. revision: partial

  2. Referee: [Methods] The precise mechanism by which vectorized constraints are enforced inside the PMCTS (e.g., how Pareto dominance is maintained across the affinity/property vector during node expansion and selection) is not specified with sufficient algorithmic detail or pseudocode, making it impossible to assess whether the reported diversity and balance improvements arise from the method itself or from post-hoc filtering.

    Authors: We appreciate this observation. The original description focused on the overall framework without sufficient low-level detail on Pareto handling. The revised manuscript now includes a new Algorithm 1 box that specifies the node expansion, selection, and back-propagation steps, including the exact Pareto dominance comparison across the multi-dimensional objective vector and the pruning rules applied during search. This makes clear that the reported improvements are produced by the integrated PMCTS procedure rather than post-hoc filtering. revision: yes

  3. Referee: [Results] Quantitative claims of 'enhanced diversity' and 'high docking scores' are presented without reported statistical tests, error bars across multiple runs, or explicit baseline implementations; this prevents evaluation of whether the improvements are significant or merely artifacts of the chosen fragment space and scoring functions.

    Authors: We concur that statistical reporting is necessary. The revised Results section now reports means and standard deviations over five independent runs for all key metrics, includes p-values from paired t-tests against each baseline, and provides additional implementation details (hyper-parameters, random seeds, and exact baseline configurations) to enable direct reproduction and fair comparison. These additions allow readers to assess the significance of the observed gains. revision: yes

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity detected; algorithmic framework evaluated externally

full rationale

The paper introduces CombiMOTS as a new Pareto Monte Carlo Tree Search (PMCTS) algorithm that explores a synthesizable fragment space under vectorized multi-objective constraints for dual-target molecule generation. Claims rest on experimental results from real-world databases showing novel molecules with high docking scores, diversity, and balanced properties. No load-bearing steps reduce by construction to fitted inputs, self-definitions, or self-citation chains; the method is presented as an independent search procedure without renaming known results or smuggling ansatzes. The derivation is self-contained and externally benchmarked.

Assumptions & free parameters 0 free parameters · 2 assumptions · 0 invented entities

The central claim rests on standard computational chemistry assumptions about simulation fidelity and fragment validity rather than new postulates.

assumptions (2)
  • domain assumption Docking scores from simulations correlate sufficiently with real target binding affinity
    Used to claim high target engagement for generated molecules.
  • domain assumption Molecules assembled from the selected fragment library are chemically synthesizable
    Underpins the synthesis-aware claim of the method.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CombiMOTS: Combinatorial Multi-Objective Tree Search for Dual-Target Molecule Generation." pith.science (2026). https://pith.science/paper/2604.23307

@misc{pith2026260423307,
  author       = {Pith},
  title        = {Pith review of: CombiMOTS: Combinatorial Multi-Objective Tree Search for Dual-Target Molecule Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2604.23307}},
  note         = {Machine review of arXiv:2604.23307}
}
read the original abstract

Dual-target molecule generation, which focuses on discovering compounds capable of interacting with two target proteins, has garnered significant attention due to its potential for improving therapeutic efficiency, safety and resistance mitigation. Existing approaches face two critical challenges. First, by simplifying the complex dual-target optimization problem to scalarized combinations of individual objectives, they fail to capture important trade-offs between target engagement and molecular properties. Second, they typically do not integrate synthetic planning into the generative process. This highlights a need for more appropriate objective function design and synthesis-aware methodologies tailored to the dual-target molecule generation task. In this work, we propose CombiMOTS, a Pareto Monte Carlo Tree Search (PMCTS) framework that generates dual-target molecules. CombiMOTS is designed to explore a synthesizable fragment space while employing vectorized optimization constraints to encapsulate target affinity and physicochemical properties. Extensive experiments on real-world databases demonstrate that CombiMOTS produces novel dual-target molecules with high docking scores, enhanced diversity, and balanced pharmacological characteristics, showcasing its potential as a powerful tool for dual-target drug discovery. The code and data is accessible through https://github.com/Tibogoss/CombiMOTS.

Figures

Figures reproduced from arXiv: 2604.23307 by the authors.

Figure 1
Figure 1. Search Space Reduction. Property-aware fragments are extracted for each target using Fragment-wise Graph Information Bottleneck (Lee et al., 2023). By applying a Tanimoto similarity threshold to a synthesizable search space, we curate a final set of informed industry-ready building blocks. Fragment Extraction To extract chemically meaningful fragments from active molecules, we employ Fragment-wise Graph Information … view at source ↗
Figure 2
Figure 2. One full iteration of the CombiMOTS Algorithm. (i) Selection from local Pareto fronts occurs until a leaf node to expand is found. (ii) During expansion, two types of nodes are created by consulting the reaction templates: those representing products from current building blocks and those incorporating compatible reactants. Upon creation, oracles predict properties for all child nodes to establish their local Pareto… view at source ↗
Figure 3
Figure 3. Normalized distributions of the generated molecules across dual-docking scores, QED and SA scores on the GSK3β￾JNK3 target pair. QED Diversity Validity Uniqueness Novelty Dock GSK3B Dock JNK3 SA 0 0.2 0.4 0.6 0.8 1.0 QED Diversity Validity Uniqueness Novelty Dock EGFR Dock MET SA 0 0.2 0.4 0.6 0.8 1.0 QED Diversity Validity Uniqueness Novelty Dock PIK3CA Dock mTOR SA 0 0.2 0.4 0.6 0.8 1.0 CombiMOTS MARS RationaleRL … view at source ↗
Figures from the paper (21 more)
Figure 4
Figure 4. Figure 4: Radar charts summarizing performance on the GSK3β￾JNK3 (left), EGFR-MET (middle) and PIK3CA-mTOR (right) tasks. We report average values of originality metrics and median values of quality metrics, normalized to the best score on each axis. 5.7. Ablation Studies To dem…
Figure 5
Figure 5. Figure 5: Examples of molecules generated by CombiMOTS on the GSK3β-JNK3 target pair. They both yield high drug-likeness and synthesizability metrics, and different (colored) parts of the compounds are positioned into each protein pocket. We show in
Figure 6
Figure 6. Figure 6: compares the predicted toxicities of 10,000 CombiMOTS-generated molecules for the GSK3β-JNK3 task against ground truth compounds from ClinTox. While some generations tend to be toxic, we make two key observa￾tions: (i) The average predicted toxicity of 10,000 samples N…
Figure 7
Figure 7. Figure 7: Normalized distributions of the generated molecules across dual-docking scores, QED and SA scores on the (A) EGFR-MET and (B) PIK3CA-mTOR target-pairs. B. Tables and Property Distributions - Ablation Studies
Figure 8
Figure 8. Figure 8: Distributions plots comparing fragmentation methods on the GSK3β-JNK3 task. 15
Figure 9
Figure 9. Figure 9: Distributions plots comparing Tanimoto similarity thresholds on the GSK3β-JNK3 task. A value of 0.3 converges slower in QED. Values of 0.4 and 0.5 perform similarly but 0.4 presents more dual-actives. From 0.6, the search space is too small to find better tradeoffs. C.…
Figure 10
Figure 10. Figure 10: t-SNE distributions on the GSK3β-JNK3 task
Figure 11
Figure 11. Figure 11: t-SNE distributions on the EGFR-MET task. 17
Figure 12
Figure 12. Figure 12: t-SNE distributions on the PIK3CA-mTOR task. 18
Figure 13
Figure 13. Figure 13: Normalized distributions of CombiMOTS against its scalarized version on the GSK3β-JNK3. 20
Figure 14
Figure 14. Figure 14: Normalized distributions of CombiMOTS against a version prioritizing QED/SA over docking score, and its six objectives version on the GSK3β-JNK3 task. 21
Figure 15
Figure 15. Figure 15: Normalized distributions of CombiMOTS against aligned versions of MARS and REINVENT on the GSK3β-JNK3 task. E.3. Impact of Data Availability We investigate the performance of CombiMOTS against AIxFuse (Chen et al., 2024). AIxFuse is a method extracting pharmacophores …
Figure 16
Figure 16. Figure 16: Normalized distributions of the generated molecules across dual-docking scores, QED and SA scores on the DHODH-RORγt target pair. 24
Figure 17
Figure 17. Figure 17: metrics for generated molecules. Mothra returns extremely few valid molecules, among which even less are dual actives: for the run displayed in
Figure 18
Figure 18. Figure 18: Radar chart on the GSK3β-JNK3 task against Mothra and MolSearch. We report average values of originality metrics and median values of quality metrics, normalized to the best score on each axis. 26
Figure 19
Figure 19. Figure 19: The four selected compounds for toxicity prediction. The vector values account for GSK3β-JNK3 predicted activities, and GSK3β-JNK3 predicted docking scores, respectively. I. Generalizability - Broader Applications We discuss in this section how CombiMOTS could be adap…
Figure 20
Figure 20. Figure 20: Example of found molecules active to CDK7 and inactive to off-targets. Vectors are predicted activities towards CDK7, CDK1, CDK2, CDK5, CDK9, CDK12, CDK13 respectively. 29
Figure 21
Figure 21. Figure 21: Chemprop predictions of molecules generated while aiming to minimize toxicity. I.3. Protein-Protein Interaction Modulator Generation In this paper, we mainly tackle the task of identifying dual inhibitors by targeting two proteins related to the same disease. An other…
Figure 22
Figure 22. Figure 22: Plot distributions across synthesizability metrics. For generative models, we evaluate candidates from the GSK3β-JNK3 task. A B
Figure 23
Figure 23. Figure 23: A) Unsolved routed of Aizynthfinder and B) Enamine building blocks found by CombiMOTS leading to the product. (Case 1) 31
Figure 24
Figure 24. Figure 24: C) Unsolved routed of Aizynthfinder and D) Enamine building blocks found by CombiMOTS leading to the product. (Case 2) 32

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

12 extracted references · 12 canonical work pages

  1. [1]

    Jin, W., Barzilay, R., and Jaakkola, T

    PMLR, 2018. Jin, W., Barzilay, R., and Jaakkola, T. Multi-objective molecule generation using interpretable substructures. In International conference on machine learning, pp. 4849–

    work page 2018

  2. [2]

    Kalgutkar, A

    PMLR, 2020. Kalgutkar, A. S. Designing around structural alerts in drug discovery.Journal of Medicinal Chemistry, 63(12):6276– 6302, 2019. Kim, S., Thiessen, P. A., Bolton, E. E., Chen, J., Fu, G., Gin- dulyte, A., Han, L., He, J., He, S., Shoemaker, B. A., et al. Pubchem substance and compound databases.Nucleic acids research, 44(D1):D1202–D1213, 2016. K...

    work page arXiv 2024

  3. [3]

    From the four initial substructures, identify similar building blocks

    work page

  4. [4]

    Note that we only select three objectives to converge faster towards Pareto optimal solutions

    Run a 200k rollout tree search, only guided by CDK7 (maximize), CDK2 and CDK9 (minimize) bioactivity predictors. Note that we only select three objectives to converge faster towards Pareto optimal solutions

    work page

  5. [5]

    We (re)-performpost-hocbioactivity predictions for all kinases, and only retain molecules with predicted values above 0.5for CDK7 and below0.5for off-targets

    work page

  6. [6]

    Optionally performpost-hocdocking simulation for added practical information

    work page

  7. [7]

    Apply industrial and medicinal filters to obtain a final list of candidates. We retain molecules with the following criteria: −0.4≤LogP≤5.6 (Ghose et al., 1999), 250≤M W≤500 , less than 5 Hydrogen Bond Donors (HBD), less than 10 Hydrogen Bond Acceptors (HBA) (Lipinski’s Rule of Five), less than 10 rotatable bonds,50≤T P SA≤140 ˚A (Veber’s rule) and does n...

    work page 1999

  8. [8]

    Notably, the COMPAS-3 dataset exclusively contains (poly)-cyclic compounds

    as representatives of easy-to-synthesize and hard-to-synthesize molecules, respectively. Notably, the COMPAS-3 dataset exclusively contains (poly)-cyclic compounds. We plot and report the molecular distributions in Figure 22. As expected, we observe several inconsistencies: COMPAS-3 is deemed synthesizable by SA and RAscore but not by BR- SAscore. Enamine...

    work page 2006

Show all 12 references
  1. [9]

    Compute the PUCB formula for each child nodev k: P U CB(k, n) = X k,nk +C×Ora(v k) s ln(D) + 4×ln(1 +n) 1 +n k , whereC∈R +∗ is an exploration constant.(10)

  2. [10]

    farthest away

    Randomly select a child node among the Pareto front built upon the PUCB formula. Definition L.1.(Most Dominant Optimal Node) Following Chen & Liu (2021), the demonstration uses the concept of ϵ-dominance of multi-objective optimization (Kollat et al., 2008). Suppose a nodev k ...

    2021

  3. [11]

    Fors k ≥N 0(ξ), we have|δ k,sk,d| ≤ξ ∆k,d 2 for alld∈ {1,2, . . . , D}

  4. [12]

    Fors k ≥ 8 lnt+2 lnD (1−ξ)2 mink,d ∆2 k,d , the confidence termc t,sk becomes sufficiently small: ct,sk = r 4 lnt+ lnD 2sk ≤ vuut 4 lnt+ lnD 2× 8 lnt+2 lnD (1−ξ)2 mink,d ∆2 k,d = (1−ξ) q mink,d ∆2 k,d √ 2 × r 4 lnt+ lnD 8 lnt+ 2 lnD ct,sk ≤ (1−ξ) min k,d ∆k,d 2 .(21) 36 CombiM...

    2021

Pith tools

Reviewed May 8, 2026 · model on record in the stance chip above.