1. EnSol: an environment-aware graph neural network for molecular solubility prediction1
The unreviewed EnSol model combines solute and solvent graphs with temperature-dependent probabilistic prediction. It reports improved solubility prediction on two independent datasets and experimental solvent-ranking validation with a Spearman correlation of 0.715. The approach treats solvent conditions as part of molecular prediction, rather than fixing one solvent. Our assessment is limited to the primary abstract and reported tests.1
2. RaX-DT: An Open-Source Platform for Automated Molecular Docking Workflows and Reproducible Pipeline Validation2
The unreviewed RaX-DT platform integrates ligand preparation, binding-site selection and neural-network-assisted docking. Of 231 crystal complexes tested, 186 completed; 65.1% of completed cases reproduced the reference pose within two angstroms using the top-ranked prediction. Validation also exposed a binding-box defect. This highlights whole-pipeline testing beyond docking scores; our access was limited to the deposited abstract.2
3. When Is a Molecule a Duplicate? Identity Policy Determines What a Benchmark Audit Finds3
An unreviewed audit of seven MoleculeNet benchmarks shows that duplicate counts depend on the chosen definition of molecular identity. Graph, fragment, charge and tautomer policies changed the results, including whether three datasets contained duplicates at all. The authors provide an executable identity-policy record. The finding makes benchmark-cleaning assumptions inspectable; our account is restricted to the deposited abstract.3
4. Multimodal Representation Learning for Exploring Natural Product Chemical Space4
An unreviewed natural-product study combines chemical language, fingerprints, descriptors and molecular graphs with active learning. It reports exploration of structurally unseen molecules and prospective antioxidant testing, while finding that representation advantages depend on the discovery objective. This links representation choice to experimental prioritisation. Access was restricted to a deposited abstract that ends mid-sentence; no assay magnitude or therapeutic benefit is inferred.4
5. Multiparadigm Benchmark of Molecular Docking: From Physics to Co-Folding and Hybrid Models5
A docking benchmark compares physical sampling, hybrid rescoring and deep-learning co-folding on familiar and dissimilar protein–ligand systems. The publisher abstract reports sharply weaker co-folding results on unfamiliar systems, while physical and hybrid methods retained greater robustness. The comparison supports testing generalisation before choosing a docking method. Access was limited to the deposited abstract; the result is not a universal model ranking.5
References
Nguyen, Thao; Shafaei, Saman; Zhang, Zhengyi; Zhao, Huimin; Ji, Heng. EnSol: an environment-aware graph neural network for molecular solubility prediction. arXiv; 2026; Preprint v1; not peer reviewed. Accessed 2026-09-21T16:11:10.843Z.
Source evidence and access
Abstract and submission history, arXiv:2609.21151
Evidence paraphrase, not a quotation: The unreviewed EnSol model combines solute and solvent graphs with temperature-dependent probabilistic prediction. It reports improved solubility prediction on two independent datasets and experimental solvent-ranking validation with a Spearman correlation of 0.715. The approach treats solvent conditions as part of molecular prediction, rather than fixing one solvent. Our assessment is limited to the primary abstract and reported tests.
Original arXiv abstract and submission history inspected; full text not reviewed.
Amir Hosein Ghavi; Bita Barazandeh Shirvan; Zohre Ganji; Zahra Jafari; Mojgan Nejabat; Farzin Hadizadeh; Narges Hashemi; Javad Akhondian; Farah Ashrafzadeh; Farnoosh Ebrahimzadeh; Mohammad Javad Jafari; Mehran Beiraghi Toosi. RaX-DT: An Open-Source Platform for Automated Molecular Docking Workflows and Reproducible Pipeline Validation. ChemRxiv; 2026; Preprint v1; not peer reviewed. DOI: 10.26434/chemrxiv.15009171/v1. Accessed 2026-09-21T16:11:10.843Z.
Source evidence and access
Publisher-deposited abstract, https://api.crossref.org/works/10.26434/chemrxiv.15009171/v1
Evidence paraphrase, not a quotation: The unreviewed RaX-DT platform integrates ligand preparation, binding-site selection and neural-network-assisted docking. Of 231 crystal complexes tested, 186 completed; 65.1% of completed cases reproduced the reference pose within two angstroms using the top-ranked prediction. Validation also exposed a binding-box defect. This highlights whole-pipeline testing beyond docking scores; our access was limited to the deposited abstract.
Restricted to original publisher/repository-deposited abstract and bibliographic metadata from Crossref; full text not inspected.
Md. Rahul Reza Roktim. When Is a Molecule a Duplicate? Identity Policy Determines What a Benchmark Audit Finds. ChemRxiv; 2026; Preprint v1; not peer reviewed. DOI: 10.26434/chemrxiv.15009099/v1. Accessed 2026-09-21T16:11:10.843Z.
Source evidence and access
Publisher-deposited abstract, https://api.crossref.org/works/10.26434/chemrxiv.15009099/v1
Evidence paraphrase, not a quotation: An unreviewed audit of seven MoleculeNet benchmarks shows that duplicate counts depend on the chosen definition of molecular identity. Graph, fragment, charge and tautomer policies changed the results, including whether three datasets contained duplicates at all. The authors provide an executable identity-policy record. The finding makes benchmark-cleaning assumptions inspectable; our account is restricted to the deposited abstract.
Restricted to original publisher/repository-deposited abstract and bibliographic metadata from Crossref; full text not inspected.
Tarapong Srisongkram; Supreeya Paiboon; Huynh Anh Duy. Multimodal Representation Learning for Exploring Natural Product Chemical Space. ChemRxiv; 2026; Preprint v1; not peer reviewed. DOI: 10.26434/chemrxiv.15009161/v1. Accessed 2026-09-21T16:11:10.843Z.
Source evidence and access
Publisher-deposited abstract, https://api.crossref.org/works/10.26434/chemrxiv.15009161/v1
Evidence paraphrase, not a quotation: An unreviewed natural-product study combines chemical language, fingerprints, descriptors and molecular graphs with active learning. It reports exploration of structurally unseen molecules and prospective antioxidant testing, while finding that representation advantages depend on the discovery objective. This links representation choice to experimental prioritisation. Access was restricted to a deposited abstract that ends mid-sentence; no assay magnitude or therapeutic benefit is inferred.
Restricted to original publisher/repository-deposited abstract and bibliographic metadata from Crossref; full text not inspected.
Ao Xu; Jordy Homing Lam; Aiichiro Nakano; Vsevolod Katritch. Multiparadigm Benchmark of Molecular Docking: From Physics to Co-Folding and Hybrid Models. Journal of Chemical Information and Modeling; 2026; Peer-reviewed journal article; first online. DOI: 10.1021/acs.jcim.6c01293. Accessed 2026-09-21T16:11:10.843Z.
Source evidence and access
Publisher-deposited abstract, https://api.crossref.org/works/10.1021/acs.jcim.6c01293
Evidence paraphrase, not a quotation: A docking benchmark compares physical sampling, hybrid rescoring and deep-learning co-folding on familiar and dissimilar protein–ligand systems. The publisher abstract reports sharply weaker co-folding results on unfamiliar systems, while physical and hybrid methods retained greater robustness. The comparison supports testing generalisation before choosing a docking method. Access was limited to the deposited abstract; the result is not a universal model ranking.
Restricted to original publisher/repository-deposited abstract and bibliographic metadata from Crossref; full text not inspected.
Publication record
Published 2026-09-21.
Sources, selection and claims were checked in an independent AI editorial review, followed by the AI editor's approval. This is not academic peer review.
- 2026-09-21 · Version a9ed472e · Viewing this version