arxiv: 2605.13932 · v1 · submitted 2026-05-13 · 💻 cs.LG

Recognition: 2 theorem links

· Lean Theorem

Rethinking Molecular OOD Generalization via Target-Aware Source Selection

Zhuohao Lin , Kun Li , Jiameng Chen , Jiajun Yu , Duanhua Cao , Yizhen Zheng , Wenbin Hu

Authors on Pith no claims yet

Pith reviewed 2026-05-15 05:04 UTC · model grok-4.3

classification 💻 cs.LG

keywords out-of-distribution generalizationmolecular property predictiondomain adaptationreinforcement learningsource selectionscaffold splitting3D molecular models

0 comments

The pith

A reinforcement learning policy selects source subsets to reduce extreme out-of-distribution errors in molecular property prediction by up to 11 percent.

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

The paper contends that common scaffold-splitting methods for testing molecular models leave hidden similarities between training and test data, so models appear to generalize better than they actually do. It demonstrates that state-of-the-art 3D models produce errors up to eight times larger when data is split by clustering in explicit physicochemical space. The proposed approach treats source selection as a reinforcement learning problem that picks the most compatible labeled scaffolds for each unlabeled target, then adapts at both large-scale topology and small-scale pharmacophore levels.

Core claim

Scaffold-splitting protocols permit microscopic semantic overlap that encourages shortcut learning, while conventional domain adaptation injects noise and causes negative transfer under large structural differences. SCOPE-BENCH creates stricter OOD tests through cluster-level partitioning in physicochemical descriptor space and shows mean error increases of 5.9 times. POMA implements a retrieve-compose-adapt pipeline that first retrieves proxy targets, then uses an RL policy to choose an optimal source subset from a large pool, and finally performs dual-scale adaptation to improve accuracy.

What carries the argument

The POMA retrieve-compose-adapt pipeline, in which a reinforcement learning policy adaptively selects an optimal source subset from candidates identified as structurally close to the target, followed by dual-scale domain adaptation at macroscopic topological and microscopic pharmacophore scales.

If this is right

Prediction errors on extreme OOD molecular tasks drop by up to 11.2 percent in mean absolute error when source selection is performed target-aware rather than blindly.
The same selection-plus-adaptation procedure yields an average 6.2 percent relative improvement across multiple 3D molecular backbone architectures.
Negative transfer is reduced because the policy avoids aligning topologically dissimilar source libraries with the target.
Prior evaluations that rely on scaffold splits systematically overestimate true extrapolation capability by a factor of roughly six on average.

Where Pith is reading between the lines

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

The same RL-driven source selection could be tested on protein-ligand or materials property tasks where distribution shifts are also extreme.
Dual-scale adaptation implies that molecular models benefit from explicit alignment at both global graph structure and local functional-group levels.
Benchmarks for scientific machine learning should routinely adopt cluster partitioning to guarantee separation at the semantic level rather than relying on scaffold rules alone.

Load-bearing premise

The reinforcement learning policy can reliably pick source subsets that avoid negative transfer, and cluster partitioning in physicochemical descriptor space fully removes any microscopic semantic overlap between source and target.

What would settle it

Apply the trained RL policy to a fresh collection of target molecules and measure whether the selected sources produce lower mean absolute error than either the full source library or randomly chosen subsets on the same SCOPE-BENCH splits.

Figures

Figures reproduced from arXiv: 2605.13932 by Duanhua Cao, Jiajun Yu, Jiameng Chen, Kun Li, Wenbin Hu, Yizhen Zheng, Zhuohao Lin.

**Figure 2.** Figure 2: Overview of the POMA framework as a retrieve–compose–adapt pipeline. Labeled source scaffolds structurally close to the target are first identified as proxy targets to enable reward estimation. A reinforcement learning policy then selects an optimal source subset from the candidate pool. Finally, a dual-scale adaptation module aligns macroscopic topologies and microscopic pharmacophore features, with polic… view at source ↗

**Figure 3.** Figure 3: t-SNE feature distributions under different splitting protocols via Local Domain Dominance [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

**Figure 4.** Figure 4: Inter-domain scaffold tanimoto similarity heatmaps. Darker blue indicates higher similarity; [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗

**Figure 5.** Figure 5: Hyperparameter sensitivity on the HOMO task for ViSNet using [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗

read the original abstract

Robust prediction of molecular properties under extreme out-of-distribution (OOD) scenarios is a pivotal bottleneck in AI-driven drug discovery. Current scaffold-splitting protocols fail to obstruct microscopic semantic overlap, predisposing models to shortcut learning and overestimating their true extrapolation capability; meanwhile, conventional domain adaptation paradigms suffer under extreme structural shifts, as blindly aligning heterogeneous source libraries injects topological noise and triggers negative transfer. To address these two challenges, scaffold-cluster out-of-distribution performance evaluation benchmark (SCOPE-BENCH), a benchmark built on cluster-level partitioning in an explicit physicochemical descriptor space, is proposed alongside policy optimization for multi-source adaptation (POMA), a framework that formulates knowledge transfer as a retrieve-compose-adapt pipeline: labeled source scaffolds structurally close to the unlabeled target are first identified as proxy targets; a reinforcement learning policy then adaptively selects the optimal source subset from an exponentially large candidate pool; and dual-scale domain adaptation is finally performed at macroscopic topological and microscopic pharmacophore scales. Evaluations show that prediction errors of state-of-the-art 3D molecular models surge by up to 8.0x on SCOPE-BENCH with a mean of 5.9x, while POMA achieves up to an 11.2% reduction in mean absolute error with an average relative improvement of 6.2% across diverse backbone architectures. Code is available at https://anonymous.4open.science/r/Molecular-OOD-Code-73F6.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

SCOPE-BENCH creates harder OOD splits for molecules via physicochemical clustering and POMA adds RL-driven source selection with dual-scale adaptation, but the clustering may still allow local structural overlap that 3D models exploit.

read the letter

The main takeaway is that this paper builds SCOPE-BENCH by clustering molecules in explicit physicochemical descriptor space to produce stricter out-of-distribution tests than standard scaffold splits, then introduces POMA as a retrieve-compose-adapt pipeline that uses reinforcement learning to pick useful source subsets before performing adaptation at both topological and pharmacophore scales. It reports that current 3D molecular models see error jumps of roughly 6x on average under the new splits and that POMA delivers average relative MAE reductions around 6 percent across backbones. Code is released, which helps reproducibility. What is actually new is the specific framing of source selection as an RL policy over an exponential pool combined with the cluster-based benchmark construction; prior work on molecular domain adaptation has mostly used blind alignment or simple retrieval without this adaptive step. The dual-scale adaptation idea is a reasonable way to handle both global and local molecular features. The paper does a service by showing how existing splits can overestimate generalization in drug-discovery settings and by offering a practical alternative pipeline. The soft spots are real but not fatal. The central worry from the stress-test note holds up on the description given: global descriptors such as molecular weight or TPSA often miss local substructures and binding motifs that state-of-the-art 3D models actually use, so cluster separation may leave residual semantic overlap and the reported error surges could partly reflect other factors. The RL policy's success at avoiding negative transfer likewise depends on reward design and held-out validation details that are not visible in the abstract; without strong ablations those gains could be sensitive to post-hoc choices. The quantitative claims are stated clearly but rest on experimental protocols that need full checking. This work is for researchers in molecular machine learning who care about realistic distribution shifts in property prediction. Readers who build or evaluate models for pharmaceutical applications will find the benchmark and the source-selection framing useful even if they end up modifying the clustering or the RL reward. It deserves peer review because the problem is practically important, the approach is concrete and testable, and the claims are falsifiable once the experiments are examined in detail.

Referee Report

3 major / 1 minor

Summary. The manuscript proposes SCOPE-BENCH, a benchmark for molecular OOD generalization constructed via cluster-level partitioning in physicochemical descriptor space to block microscopic semantic overlap and shortcut learning, together with POMA, a retrieve-compose-adapt framework that uses reinforcement learning to select optimal source subsets from an exponentially large pool before performing dual-scale domain adaptation at topological and pharmacophore levels. It reports that state-of-the-art 3D molecular models exhibit prediction-error increases of up to 8.0x (mean 5.9x) on SCOPE-BENCH while POMA delivers up to 11.2% MAE reduction with 6.2% average relative improvement across backbones.

Significance. If the benchmark truly isolates extreme OOD without residual overlap and the RL policy reliably avoids negative transfer, the work would meaningfully advance reliable property prediction in AI-driven drug discovery by supplying both a stricter evaluation protocol and a practical adaptation method. Code availability is a positive factor for reproducibility.

major comments (3)

[SCOPE-BENCH] SCOPE-BENCH construction: the load-bearing claim that cluster-level partitioning in explicit physicochemical descriptor space fully eliminates microscopic semantic overlap is insufficiently supported. Global descriptors (MW, logP, TPSA) do not encode local 3D substructure or pharmacophore similarity exploited by the evaluated 3D models; any residual overlap would invalidate attribution of the reported 5.9x mean error surge solely to OOD.
[POMA] POMA framework: the reinforcement-learning policy is asserted to identify source subsets that avoid negative transfer, yet the reward design and any explicit penalization mechanism (e.g., held-out proxy-target validation) are not detailed. Without such safeguards the adaptive selection may still include harmful sources, directly affecting the claimed 6.2% average improvement.
[Evaluations] Experimental evaluation: the quantitative claims (8.0x surge, 11.2% MAE reduction, 6.2% relative gain) are presented without protocol details, baseline definitions, statistical significance tests, or ablation studies, preventing verification that the gains are not due to post-hoc choices.

minor comments (1)

[Abstract] The abstract would benefit from explicit definitions of 'microscopic semantic overlap' and 'dual-scale domain adaptation' to improve immediate clarity.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback on our manuscript. We have carefully reviewed each major comment and will revise the paper to address the concerns by providing additional clarifications, analyses, and experimental details. Below we respond point-by-point to the major comments.

read point-by-point responses

Referee: [SCOPE-BENCH] SCOPE-BENCH construction: the load-bearing claim that cluster-level partitioning in explicit physicochemical descriptor space fully eliminates microscopic semantic overlap is insufficiently supported. Global descriptors (MW, logP, TPSA) do not encode local 3D substructure or pharmacophore similarity exploited by the evaluated 3D models; any residual overlap would invalidate attribution of the reported 5.9x mean error surge solely to OOD.

Authors: We thank the referee for this important observation. While our cluster-level partitioning in physicochemical descriptor space is designed to enforce a stricter separation than standard scaffold splits by operating at a higher semantic granularity, we acknowledge that global descriptors alone may leave some residual local 3D or pharmacophore similarities unaddressed. In the revised manuscript we will add a quantitative analysis of residual overlap using 3D similarity metrics (e.g., pharmacophore fingerprint Tanimoto scores and 3D shape overlap) computed across cluster boundaries, together with a discussion of the remaining limitations. This will strengthen the attribution of the observed error increases to genuine OOD effects. revision: yes
Referee: [POMA] POMA framework: the reinforcement-learning policy is asserted to identify source subsets that avoid negative transfer, yet the reward design and any explicit penalization mechanism (e.g., held-out proxy-target validation) are not detailed. Without such safeguards the adaptive selection may still include harmful sources, directly affecting the claimed 6.2% average improvement.

Authors: We appreciate the referee's request for greater transparency on the RL component. The reward function in POMA is a weighted sum of (i) negative MAE on a small held-out proxy-target validation set drawn from the target domain and (ii) a diversity penalty that discourages selection of sources whose topological or pharmacophore distributions deviate strongly from the target. In the revision we will provide the exact mathematical formulation of the reward, the policy gradient objective, training hyperparameters, and pseudocode for the selection procedure, thereby clarifying the safeguards against negative transfer. revision: yes
Referee: [Evaluations] Experimental evaluation: the quantitative claims (8.0x surge, 11.2% MAE reduction, 6.2% relative gain) are presented without protocol details, baseline definitions, statistical significance tests, or ablation studies, preventing verification that the gains are not due to post-hoc choices.

Authors: We agree that the current experimental section lacks sufficient detail for full reproducibility and verification. In the revised manuscript we will expand the evaluation section to include: complete data-preprocessing and splitting protocols, precise definitions and implementations of all baselines, results of statistical significance tests (paired t-tests and Wilcoxon signed-rank tests with reported p-values), and comprehensive ablation studies isolating the contributions of the RL source selection and the dual-scale adaptation modules. These additions will substantiate the reported performance numbers. revision: yes

Circularity Check

0 steps flagged

No significant circularity; framework uses external standard components

full rationale

The paper introduces SCOPE-BENCH via explicit cluster-level partitioning in physicochemical descriptor space and POMA as a retrieve-compose-adapt pipeline using a reinforcement learning policy for source subset selection followed by dual-scale domain adaptation. No equations, derivations, or first-principles predictions are presented that reduce to fitted parameters or inputs by construction. The central claims consist of empirical evaluations showing error surges and MAE reductions, which rely on standard RL and domain-adaptation techniques whose grounding is external. No self-citation chains, uniqueness theorems, or smuggled ansatzes serve as load-bearing steps. The derivation chain remains self-contained through explicit methodological definitions and benchmark construction rather than circular reductions.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

Abstract-only review; no explicit free parameters, axioms, or invented entities are stated. The work rests on the domain assumption that physicochemical clustering creates true extreme OOD and that RL policy optimization can avoid negative transfer.

axioms (2)

domain assumption Scaffold-splitting protocols fail to obstruct microscopic semantic overlap
Directly asserted in the abstract as the reason current protocols overestimate extrapolation capability.
domain assumption Blind alignment of heterogeneous source libraries injects topological noise and triggers negative transfer
Stated as the limitation of conventional domain adaptation under extreme structural shifts.

pith-pipeline@v0.9.0 · 5576 in / 1464 out tokens · 51038 ms · 2026-05-15T05:04:32.992385+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

SCOPE-BENCH constructs domains via scaffold extraction, four-dimensional physicochemical feature vectors fk = Vmacro ⊕ Velement ⊕ Vconn ⊕ Vflex, hierarchical clustering with K-Means++ yielding 12 globally disjoint clusters, and asymmetric partitioning.
IndisputableMonolith/Foundation/BranchSelection.lean branch_selection unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

POMA formulates knowledge transfer as retrieve–compose–adapt with GRPO policy optimization over candidate pool C, dual-scale alignment LDA using squared Frobenius norms of covariance matrices at mol and sub scales.

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

58 extracted references · 58 canonical work pages · 5 internal anchors

[1]

Invariant Risk Minimization

Martin Arjovsky, Léon Bottou, Ishaan Gulrajani, and David Lopez-Paz. Invariant risk minimization.arXiv preprint arXiv:1907.02893, 2019

work page internal anchor Pith review Pith/arXiv arXiv 1907
[2]

k-means++: The advantages of careful seeding

David Arthur, Sergei Vassilvitskii, et al. k-means++: The advantages of careful seeding. InSoda, volume 7, pages 1027–1035, 2007

work page 2007
[3]

Gotennet: Rethinking efficient 3d equivariant graph neural networks

Sarp Aykent and Tian Xia. Gotennet: Rethinking efficient 3d equivariant graph neural networks. InThe Thirteenth International Conference on Learning Representations, 2025

work page 2025
[4]

Why is tanimoto index an appropriate choice for fingerprint-based similarity calculations?Journal of cheminformatics, 7(1):20, 2015

Dávid Bajusz, Anita Rácz, and Károly Héberger. Why is tanimoto index an appropriate choice for fingerprint-based similarity calculations?Journal of cheminformatics, 7(1):20, 2015

work page 2015
[5]

E (n) equivariant topological neural networks.arXiv preprint arXiv:2405.15429, 2024

Claudio Battiloro, Mauricio Tec, George Dasoulas, Michelle Audirac, Francesca Dominici, et al. E (n) equivariant topological neural networks.arXiv preprint arXiv:2405.15429, 2024

work page arXiv 2024
[6]

Neural Combinatorial Optimization with Reinforcement Learning

Irwan Bello, Hieu Pham, Quoc V Le, Mohammad Norouzi, and Samy Bengio. Neural combinatorial optimization with reinforcement learning.arXiv preprint arXiv:1611.09940, 2016

work page internal anchor Pith review Pith/arXiv arXiv 2016
[7]

The properties of known drugs

Guy W Bemis and Mark A Murcko. The properties of known drugs. 1. molecular frameworks.Journal of medicinal chemistry, 39(15):2887–2893, 1996

work page 1996
[8]

A theory of learning from different domains.Machine learning, 79(1):151–175, 2010

Shai Ben-David, John Blitzer, Koby Crammer, Alex Kulesza, Fernando Pereira, and Jennifer Wortman Vaughan. A theory of learning from different domains.Machine learning, 79(1):151–175, 2010

work page 2010
[9]

Molevolve: Llm-guided evolutionary search for interpretable molecular optimization.arXiv preprint arXiv:2603.24382, 2026

Xiangsen Chen, Ruilong Wu, Yanyan Lan, Ting Ma, and Yang Liu. Molevolve: Llm-guided evolutionary search for interpretable molecular optimization.arXiv preprint arXiv:2603.24382, 2026

work page arXiv 2026
[10]

On the art of compiling and using’drug-like’chemical fragment spaces.ChemMedChem, 3(10):1503, 2008

Jorg Degen, Christof Wegscheid-Gerlach, Andrea Zaliani, and Matthias Rarey. On the art of compiling and using’drug-like’chemical fragment spaces.ChemMedChem, 3(10):1503, 2008

work page 2008
[11]

Tackling dimensional collapse toward comprehensive universal domain adaptation.arXiv preprint arXiv:2410.11271, 2024

Hung-Chieh Fang, Po-Yi Lu, and Hsuan-Tien Lin. Tackling dimensional collapse toward comprehensive universal domain adaptation.arXiv preprint arXiv:2410.11271, 2024

work page arXiv 2024
[12]

Chembl: a large-scale bioactivity database for drug discovery.Nucleic acids research, 40(D1):D1100–D1107, 2012

Anna Gaulton, Louisa J Bellis, A Patricia Bento, Jon Chambers, Mark Davies, Anne Hersey, Yvonne Light, Shaun McGlinchey, David Michalovich, Bissan Al-Lazikani, et al. Chembl: a large-scale bioactivity database for drug discovery.Nucleic acids research, 40(D1):D1100–D1107, 2012

work page 2012
[13]

Shortcut learning in deep neural networks.Nature Machine Intelligence, 2(11):665–673, 2020

Robert Geirhos, Jörn-Henrik Jacobsen, Claudio Michaelis, Richard Zemel, Wieland Brendel, Matthias Bethge, and Felix A Wichmann. Shortcut learning in deep neural networks.Nature Machine Intelligence, 2(11):665–673, 2020

work page 2020
[14]

Neural message passing for quantum chemistry

Justin Gilmer, Samuel S Schoenholz, Patrick F Riley, Oriol Vinyals, and George E Dahl. Neural message passing for quantum chemistry. InInternational conference on machine learning, pages 1263–1272. Pmlr, 2017

work page 2017
[15]

Good: A graph out-of-distribution benchmark

Shurui Gui, Xiner Li, Limei Wang, and Shuiwang Ji. Good: A graph out-of-distribution benchmark. Advances in Neural Information Processing Systems, 35:2059–2073, 2022

work page 2059
[16]

Inhomogeneous electron gas.Physical review, 136(3B):B864, 1964

Pierre Hohenberg and Walter Kohn. Inhomogeneous electron gas.Physical review, 136(3B):B864, 1964. 10

work page 1964
[17]

Strategies for pre-training graph neural networks.arXiv preprint arXiv:1905.12265, 2019

Weihua Hu, Bowen Liu, Joseph Gomes, Marinka Zitnik, Percy Liang, Vijay Pande, and Jure Leskovec. Strategies for pre-training graph neural networks.arXiv preprint arXiv:1905.12265, 2019

work page arXiv 1905
[18]

Semi-Supervised Classification with Graph Convolutional Networks

Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907, 2016

work page internal anchor Pith review Pith/arXiv arXiv 2016
[19]

RDKit: Open-source chemoinformatics, 2024

Greg Landrum et al. RDKit: Open-source chemoinformatics, 2024. URL https://doi.org/10.5281/ zenodo.12782092

work page 2024
[20]

Out-of-distribution generalization on graphs: A survey.IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025

Haoyang Li, Xin Wang, Ziwei Zhang, and Wenwu Zhu. Out-of-distribution generalization on graphs: A survey.IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025

work page 2025
[21]

Drugpilot: Llm-based parameterized reasoning agent for drug discovery.arXiv preprint arXiv:2505.13940, 2025

Kun Li, Zhennan Wu, Shoupeng Wang, Jia Wu, Shirui Pan, and Wenbin Hu. Drugpilot: Llm-based parameterized reasoning agent for drug discovery.arXiv preprint arXiv:2505.13940, 2025

work page arXiv 2025
[22]

Bsl: A unified and generalizable multitask learning platform for virtual drug discovery from design to synthesis.arXiv preprint arXiv:2508.01195, 2025

Kun Li, Zhennan Wu, Yida Xiong, Hongzhi Zhang, Longtao Hu, Zhonglie Liu, Junqi Zeng, Wenjie Wu, Mukun Chen, Jiameng Chen, et al. Bsl: A unified and generalizable multitask learning platform for virtual drug discovery from design to synthesis.arXiv preprint arXiv:2508.01195, 2025

work page arXiv 2025
[23]

Graph-structured small molecule drug discovery through deep learning: Progress, challenges, and opportunities

Kun Li, Yida Xiong, Hongzhi Zhang, Xiantao Cai, Jia Wu, Bo Du, and Wenbin Hu. Graph-structured small molecule drug discovery through deep learning: Progress, challenges, and opportunities. In2025 IEEE International Conference on Web Services (ICWS), pages 1033–1042, 2025. doi: 10.1109/ICWS67624. 2025.00135

work page doi:10.1109/icws67624 2025
[24]

Contrastive learning-based drug screening model for glun1/glun3a inhibitors.Acta Pharmacologica Sinica, pages 1–13, 2025

Kun Li, Yue Zeng, Yi-da Xiong, Hao-chen Wu, Sui Fang, Zhi-yan Qu, Yan Zhu, Bo Du, Zhao-bing Gao, and Wen-bin Hu. Contrastive learning-based drug screening model for glun1/glun3a inhibitors.Acta Pharmacologica Sinica, pages 1–13, 2025

work page 2025
[25]

Can molecular evolution mechanism enhance molecular representation? InProceedings of the AAAI Conference on Artificial Intelligence, volume 40, pages 15108–15116, 2026

Kun Li, Longtao Hu, Jiameng Chen, Hongzhi Zhang, Yida Xiong, Xiantao Cai, Wenbin Hu, and Jia Wu. Can molecular evolution mechanism enhance molecular representation? InProceedings of the AAAI Conference on Artificial Intelligence, volume 40, pages 15108–15116, 2026

work page 2026
[26]

Pcevo: Path-consistent molecular representation via virtual evolutionary

Kun Li, Longtao Hu, Yida Xiong, Jiajun Yu, Hongzhi Zhang, Jiameng Chen, Xiantao Cai, Jia Wu, and Wenbin Hu. Pcevo: Path-consistent molecular representation via virtual evolutionary. InProceedings of the Thirty-Fourth International Joint Conference on Artificial Intelligence, IJCAI-26. International Joint Conferences on Artificial Intelligence Organization...

work page 2026
[27]

Towards out-of-distribution generalization: A survey.arXiv preprint arXiv:2108.13624, 2021

Jiashuo Liu, Zheyan Shen, Yue He, Xingxuan Zhang, Renzhe Xu, Han Yu, and Peng Cui. Towards out-of-distribution generalization: A survey.arXiv preprint arXiv:2108.13624, 2021

work page arXiv 2021
[28]

The generation of a unique machine description for chemical structures-a technique developed at chemical abstracts service.Journal of chemical documentation, 5(2):107–113, 1965

Harry L Morgan. The generation of a unique machine description for chemical structures-a technique developed at chemical abstracts service.Journal of chemical documentation, 5(2):107–113, 1965

work page 1965
[29]

Can you trust your model’s uncertainty? evaluating predictive uncertainty under dataset shift.Advances in neural information processing systems, 32, 2019

Yaniv Ovadia, Emily Fertig, Jie Ren, Zachary Nado, David Sculley, Sebastian Nowozin, Joshua Dillon, Balaji Lakshminarayanan, and Jasper Snoek. Can you trust your model’s uncertainty? evaluating predictive uncertainty under dataset shift.Advances in neural information processing systems, 32, 2019

work page 2019
[30]

A survey on transfer learning.IEEE Transactions on knowledge and data engineering, 22(10):1345–1359, 2009

Sinno Jialin Pan and Qiang Yang. A survey on transfer learning.IEEE Transactions on knowledge and data engineering, 22(10):1345–1359, 2009

work page 2009
[31]

Moment matching for multi-source domain adaptation

Xingchao Peng, Qinxun Bai, Xide Xia, Zijun Huang, Kate Saenko, and Bo Wang. Moment matching for multi-source domain adaptation. InProceedings of the IEEE/CVF international conference on computer vision, pages 1406–1415, 2019

work page 2019
[32]

Deep reinforcement learning for de novo drug design.Science advances, 4(7):eaap7885, 2018

Mariya Popova, Olexandr Isayev, and Alexander Tropsha. Deep reinforcement learning for de novo drug design.Science advances, 4(7):eaap7885, 2018

work page 2018
[33]

Msanchor: De novo molecular generation from mass spectrometry data with anchor-extended molecular scaffolds

Xiaohan Qin, Chao Wang, Zhengyang Zhou, Linjiang Chen, Wenjie Du, and Yang Wang. Msanchor: De novo molecular generation from mass spectrometry data with anchor-extended molecular scaffolds. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 40, pages 953–961, 2026

work page 2026
[34]

Quantum chemistry structures and properties of 134 kilo molecules.Scientific data, 1(1):1–7, 2014

Raghunathan Ramakrishnan, Pavlo O Dral, Matthias Rupp, and O Anatole V on Lilienfeld. Quantum chemistry structures and properties of 134 kilo molecules.Scientific data, 1(1):1–7, 2014

work page 2014
[35]

Extended-connectivity fingerprints.Journal of chemical information and modeling, 50(5):742–754, 2010

David Rogers and Mathew Hahn. Extended-connectivity fingerprints.Journal of chemical information and modeling, 50(5):742–754, 2010

work page 2010
[36]

Rosenstein, Zvika Marx, Leslie Pack Kaelbling, and Thomas G

Michael T. Rosenstein, Zvika Marx, Leslie Pack Kaelbling, and Thomas G. Dietterich. To transfer or not to transfer. InNIPS Workshop on Transfer Learning, 2005. 11

work page 2005
[37]

E (n) equivariant graph neural networks

Vıctor Garcia Satorras, Emiel Hoogeboom, and Max Welling. E (n) equivariant graph neural networks. In International conference on machine learning, pages 9323–9332. PMLR, 2021

work page 2021
[38]

Evaluating scalable uncertainty estimation methods for deep learning-based molecular property prediction.Journal of chemical information and modeling, 60(6):2697–2717, 2020

Gabriele Scalia, Colin A Grambow, Barbara Pernici, Yi-Pei Li, and William H Green. Evaluating scalable uncertainty estimation methods for deep learning-based molecular property prediction.Journal of chemical information and modeling, 60(6):2697–2717, 2020

work page 2020
[39]

Proximal Policy Optimization Algorithms

John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms.arXiv preprint arXiv:1707.06347, 2017

work page internal anchor Pith review Pith/arXiv arXiv 2017
[40]

Schnet: A continuous-filter convolutional neural network for modeling quantum interactions.Advances in neural information processing systems, 30, 2017

Kristof Schütt, Pieter-Jan Kindermans, Huziel Enoc Sauceda Felix, Stefan Chmiela, Alexandre Tkatchenko, and Klaus-Robert Müller. Schnet: A continuous-filter convolutional neural network for modeling quantum interactions.Advances in neural information processing systems, 30, 2017

work page 2017
[41]

Equivariant message passing for the prediction of tensorial properties and molecular spectra

Kristof Schütt, Oliver Unke, and Michael Gastegger. Equivariant message passing for the prediction of tensorial properties and molecular spectra. InInternational conference on machine learning, pages 9377–9388. PMLR, 2021

work page 2021
[42]

DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models

Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Yang Wu, et al. Deepseekmath: Pushing the limits of mathematical reasoning in open language models.arXiv preprint arXiv:2402.03300, 2024

work page internal anchor Pith review Pith/arXiv arXiv 2024
[43]

Weisfeiler-lehman graph kernels.Journal of Machine Learning Research, 12(9), 2011

Nino Shervashidze, Pascal Schweitzer, Erik Jan Van Leeuwen, Kurt Mehlhorn, and Karsten M Borgwardt. Weisfeiler-lehman graph kernels.Journal of Machine Learning Research, 12(9), 2011

work page 2011
[44]

Deep coral: Correlation alignment for deep domain adaptation

Baochen Sun and Kate Saenko. Deep coral: Correlation alignment for deep domain adaptation. In European conference on computer vision, pages 443–450. Springer, 2016

work page 2016
[45]

Wasserstein weisfeiler-lehman graph kernels.Advances in neural information processing systems, 32, 2019

Matteo Togninalli, Elisabetta Ghisu, Felipe Llinares-López, Bastian Rieck, and Karsten Borgwardt. Wasserstein weisfeiler-lehman graph kernels.Advances in neural information processing systems, 32, 2019

work page 2019
[46]

Visualizing data using t-sne.Journal of machine learning research, 9(11), 2008

Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne.Journal of machine learning research, 9(11), 2008

work page 2008
[47]

Assessing the impact of generative ai on medicinal chemistry.Nature biotechnology, 38(2):143–145, 2020

W Patrick Walters and Mark Murcko. Assessing the impact of generative ai on medicinal chemistry.Nature biotechnology, 38(2):143–145, 2020

work page 2020
[48]

Scientific discovery in the age of artificial intelligence.Nature, 620(7972):47–60, 2023

Hanchen Wang, Tianfan Fu, Yuanqi Du, Wenhao Gao, Kexin Huang, Ziming Liu, Payal Chandak, Shengchao Liu, Peter Van Katwyk, Andreea Deac, et al. Scientific discovery in the age of artificial intelligence.Nature, 620(7972):47–60, 2023

work page 2023
[49]

Enhancing geometric representations for molecules with equivariant vector-scalar interactive message passing.Nature Communications, 15(1):313, 2024

Yusong Wang, Tong Wang, Shaoning Li, Xinheng He, Mingyu Li, Zun Wang, Nanning Zheng, Bin Shao, and Tie-Yan Liu. Enhancing geometric representations for molecules with equivariant vector-scalar interactive message passing.Nature Communications, 15(1):313, 2024

work page 2024
[50]

Characterizing and avoiding negative transfer

Zirui Wang, Zihang Dai, Barnabás Póczos, and Jaime Carbonell. Characterizing and avoiding negative transfer. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11293–11302, 2019

work page 2019
[51]

Moleculenet: a benchmark for molecular machine learning.Chemical science, 9(2):513–530, 2018

Zhenqin Wu, Bharath Ramsundar, Evan N Feinberg, Joseph Gomes, Caleb Geniesse, Aneesh S Pappu, Karl Leswing, and Vijay Pande. Moleculenet: a benchmark for molecular machine learning.Chemical science, 9(2):513–530, 2018

work page 2018
[52]

Electron density-enhanced molecular geometry learning

Hongxin Xiang, Jun Xia, Xin Jin, Wenjie Du, Li Zeng, and Xiangxiang Zeng. Electron density-enhanced molecular geometry learning. InProceedings of the Thirty-Fourth International Joint Conference on Artificial Intelligence, pages 7840–7848, 2025

work page 2025
[53]

Learning substructure invariance for out-of-distribution molecular representations.Advances in Neural Information Processing Systems, 35: 12964–12978, 2022

Nianzu Yang, Kaipeng Zeng, Qitian Wu, Xiaosong Jia, and Junchi Yan. Learning substructure invariance for out-of-distribution molecular representations.Advances in Neural Information Processing Systems, 35: 12964–12978, 2022

work page 2022
[54]

Kernel readout for graph neural networks

Jiajun Yu, Zhihao Wu, Jinyu Cai, Adele Lu Jia, and Jicong Fan. Kernel readout for graph neural networks. InIJCAI, pages 2505–2514, 2024

work page 2024
[55]

A centrality-based graph learning framework

Jiajun Yu, Zhihao Wu, Jielong Lu, Tianyue Wang, and Haishuai Wang. A centrality-based graph learning framework. InProceedings of the Thirty-Fourth International Joint Conference on Artificial Intelligence, pages 3588–3596, 2025. 12

work page 2025
[56]

Adversarial multiple source domain adaptation.Advances in neural information processing systems, 31, 2018

Han Zhao, Shanghang Zhang, Guanhang Wu, José MF Moura, Joao P Costeira, and Geoffrey J Gordon. Adversarial multiple source domain adaptation.Advances in neural information processing systems, 31, 2018

work page 2018
[57]

Optimization of molecules via deep reinforcement learning.Scientific reports, 9(1):10752, 2019

Zhenpeng Zhou, Steven Kearnes, Li Li, Richard N Zare, and Patrick Riley. Optimization of molecules via deep reinforcement learning.Scientific reports, 9(1):10752, 2019. 13 A Training Algorithm and Implementation Details All experiments are conducted on NVIDIA RTX A6000 (48 GB) GPUs. The computational overhead of POMA can be decoupled into two phases: (1) ...

work page 2019
[58]

By contrast, the target-aware selection mechanism in POMA isolates a sparse but highly relevant subset of source knowledge

In these cases, conventional fine-tuning often leads to catastrophic forgetting or negative transfer due to the injection of irrelevant source noise. By contrast, the target-aware selection mechanism in POMA isolates a sparse but highly relevant subset of source knowledge. Consequently, the model maintains high predictive accuracy without being compromise...

work page arXiv 1941