1. An AI-Driven Chemogenomics Knowledgebase for Human-Transmissible Pathogens: A Platform for Antimicrobial Drug Discovery1
HPM-CKB combines pathogen chemogenomic data, virtual screening and language-model queries. The publisher abstract reports compound binding and cell-based antiviral tests, plus experimentally checked binding to a proposed bacterial target for deep-learning-prioritised leads. These examples connect the database to laboratory follow-up. Access was restricted to the abstract; the assays do not establish clinical effectiveness or validate every database prediction.1
2. BondNet: Noise-Robust Molecular Bond Perception from 3D Coordinates with Explicit Hydrogen Context2
The unreviewed BondNet preprint reconstructs molecular bonds from three-dimensional coordinates using geometric connectivity prediction and topology-aware bond typing. Explicit hydrogen context and noise augmentation improved robustness; a leakage-controlled comparison also favoured diverse training data over simply adding more molecules. The work addresses converting generated coordinates into chemically usable graphs. Reporting is limited to the deposited abstract and its tested datasets.2
3. Benchmark Label Definition, Not Model Capability, Explains a Machine-Learning Advantage over TD-DFT for λmax Prediction3
An unreviewed absorption-maximum benchmark shows how label definitions can determine an apparent machine-learning advantage over quantum calculations. A fingerprint-based Gaussian-process model transferred better when external datasets used compatible labels and failed on transition-specific photoswitch labels; its uncertainty was poorly calibrated externally. The study makes target definition part of model evaluation. This account is limited to the deposited abstract.3
4. A traceable agentic framework for chemistry research query and analysis with ThermoML database4
An unreviewed ThermoML agent framework combines language-model queries with condition matching, evidence records and numerical checks. Five case studies illustrate retrieval and mixture-property analysis. Disabling numerical grounding in one open-ended case produced severe numerical hallucinations despite coherent reasoning. The work highlights traceable experimental conditions and numbers as requirements for useful chemical database assistants. Our access was limited to the deposited abstract.4
5. SaltyMeta: a curated benchmark and protein language model-informed web tool for salty peptide prediction5
The unreviewed SaltyMeta study curates 580 positive and negative peptide examples and evaluates descriptors, protein-language-model embeddings and their combinations using similarity-grouped splits. Its best reported test ROC-AUC was 0.715, with modest repeated-cross-validation gains. The benchmark makes leakage and uncertain negative labels visible in a small-data discovery task. It is a screening aid, not a replacement for sensory testing.5
References
Jiaxiong Kang; Lei Zhu; Haibo Li; Xiangyu Xie; Cong Liu; Zhiwei Feng; Yi Wang; Ouyang Mo; Chengsu Wang; Xinzi Lin; Ying Xue; Haibin Liu; Qin Ouyang. An AI-Driven Chemogenomics Knowledgebase for Human-Transmissible Pathogens: A Platform for Antimicrobial Drug Discovery. Journal of Chemical Information and Modeling; 2026; Peer-reviewed journal article; first online. DOI: 10.1021/acs.jcim.6c02102. Accessed 2026-09-20T16:23:23.853Z.
Source evidence and access
Publisher-deposited abstract, https://api.crossref.org/works/10.1021/acs.jcim.6c02102
Emerging infectious diseases (EIDs) pose a critical threat to global biosecurity. Integrating bio/chemical information and artificial intelligence would bring new strategies for antimicrobial drug discovery. Herein, the Human Pathogenic Microorganisms Chemogenomics Knowledgebase (HPM-CKB) is presented as the largest domain-specific resource, consolidating chemical, genetic, and proteomic data on human-transmissible pathogens, together with multiple computational functional modules. The current release covers 7876 pathogenic proteins from 267 microorganisms (including 13,914 protein 3D structures) and 234,287 associated with bioactive molecules. HPM-CKB enables large-scale virtual screening, target identification, and drug repurposing, and integrates a large language model (LLM) for interactive queries. The computational prediction performance of HPM-CKB is corroborated by known inhibitors targeting SARS-CoV-2 replicase polyprotein 1ab. In wet-lab validations, four approved drugs (cefixime, ceftazidime, saquinavir, and rilapladib) identified via virtual screening show binding activity to SARS-CoV-2 nucleoprotein in affinity assays and inhibit SARS-CoV-2 replication in Vero E6 cells, demonstrating HPM-CKB’s potential in drug repurposing. Meanwhile, two anti-Staphylococcus aureus lead compounds with novel scaffolds (CYC-HXL-9124 and CYC-HXL-9126) are identified via deep learning, and the potential target protein, cell division protein FtsZ, is subsequently prioritized using HPM-CKB (http://cgai.asia/g/pathogenDB) and experimentally validated by affinity assays. Collectively, these findings establish HPM-CKB as both a chemogenomic knowledgebase and a systematic drug development platform against EIDs.
Abstract only: original publisher/repository-deposited metadata retrieved from official Crossref API. Full text not inspected.
Xiaochen Zhang; Hui Zeng; Xiao-Chen Zhang. BondNet: Noise-Robust Molecular Bond Perception from 3D Coordinates with Explicit Hydrogen Context. ChemRxiv; 2026; Preprint v1; not peer reviewed. DOI: 10.26434/chemrxiv.15008897/v1. Accessed 2026-09-20T16:23:23.853Z.
Source evidence and access
Publisher-deposited abstract, https://api.crossref.org/works/10.26434/chemrxiv.15008897/v1
Recovering molecular graphs from atomic identities and 3D coordinates is a key step in coordinate-first molecular modelling and generation. We present Bond-Net, a two-stage framework in which a PaiNN model first predicts connectivity on a distance-derived candidate graph and a topology-aware network then classifies heavy–heavy bond types. Explicit hydrogens participate in message passing as local valence context but are excluded from bond-type supervision and scoring. On a fixed 26,940-molecule GEOM-DRUGS validation cohort, the near-clean model achieves pipeline macro-F1 = 0.9994 and correctly types every reference heavy–heavy bond in 99.4% of molecules. A historical ablation shows that removing hydrogen context reduces macro-F1 by 6.7 percentage points and the exact-type rate to 19.7%. With Gaussian coordinate-noise augmentation and candidate graphs rebuilt from the perturbed coordinates, BondNet retains pipeline macro-F1 = 0.9859 and a true-bond exact-type rate of 87.7% at σ = 0.1 Å; its stricter HH-graph exact match is 83.0% because extra predicted bonds are also penalized. BondNet’s pipeline macro-F1 exceeds RDKit DetermineBonds at every tested noise level on the common denominator. Residual strict-label disagreements are concentrated in resonance-ambiguous N–O and carboxyl motifs. An exact-molecule-leakage-controlled diversity study further shows that a curated multi-source set of 80,000 molecules transfers better than both a same-size random GEOM subset and a 3.4× larger single-source set. 1 Scientific Contribution. BondNet isolates geometric connectivity from topology-aware bond typing and quantifies how explicit-hydrogen context and noise augmentation affect robust graph recovery. The study also provides atom-pair-matched baselines, explicitly reports tool failures, and introduces an exact-molecule-leakage-controlled comparison of training-set diversity and scale.
Abstract only: original publisher/repository-deposited metadata retrieved from official Crossref API. Full text not inspected.
Md Sahanawaz. Benchmark Label Definition, Not Model Capability, Explains a Machine-Learning Advantage over TD-DFT for λmax Prediction. ChemRxiv; 2026; Preprint v1; not peer reviewed. DOI: 10.26434/chemrxiv.15008972/v1. Accessed 2026-09-20T16:23:23.853Z.
Source evidence and access
Publisher-deposited abstract, https://api.crossref.org/works/10.26434/chemrxiv.15008972/v1
Machine-learning models trained on literature-mined UV-Vis data are frequently reported to outperform time-dependent density functional theory (TD-DFT) for predicting absorption maxima (λ max ). Such comparisons usually evaluate the model on held-out molecules from the same corpus that supplied its training labels, and they set experimental labels of heterogeneous transition type against computed vertical excitations of a single electronic transition. Using a Gaussian process (GP) with 256-bit Morgan fingerprints, trained on 6,878 curated compounds from the Beard literature-mined database, the study shows that two features of the benchmark, rather than model capability, govern the reported margin. On leakage-free in-corpus holdouts the GP outperforms the quantum-chemical references (RMSE 87.3 vs 108.7 nm against TD-DFT, n = 36; 81.6 vs 117.1 nm against sTDA, n = 1,011), and the advantage survives Bemis–Murcko scaffold splitting. Applied unchanged to genuinely external corpora, the same model keeps working where the label convention is compatible with training (R² = +0.209 on 5,811 compounds) and fails where it is not (R² = −0.269 with a +53.5 nm offset on 318 photoswitches labelled specifically to the E -isomer π→π* transition); chemical novelty alone does not break it, reassigning the label does. Against published ωB97X-D3 vertical excitations on 5,501 training-disjoint compounds, the calculations are the more accurate predictor once each method’s removable offset is discounted (≈58 vs 95 nm residual scatter) and correlate with experiment nearly twice as strongly (r = 0.85 vs 0.46). On the transition-specific corpus, CAM-B3LYP/6-31G** reaches 25.4 nm against the GP’s 80.9 nm. The 57–65 nm quantum-chemical underestimation seen in-corpus reappears out of sample at −61.9 nm against independently computed values, supporting a definitional rather than a functional origin. GP predictive uncertainty is poorly calibrated in all external regimes and does not reliably identify high-error predictions; its apparent correlation with error reflects the label variance of the evaluation set rather than structural distance from the training distribution. For this model class, reported margins over quantum chemistry should therefore be regarded as properties of the benchmark—its label definition and its same-corpus provenance,not of the method, unless the benchmark is external and its label convention is stated.
Abstract only: original publisher/repository-deposited metadata retrieved from official Crossref API. Full text not inspected.
Yunkai Sun; Adwaith Ravichandran; Hassan Harb; Rajeev S. Assary; Brian J. Ingram; Zhenzhen Yang. A traceable agentic framework for chemistry research query and analysis with ThermoML database. ChemRxiv; 2026; Preprint v1; not peer reviewed. DOI: 10.26434/chemrxiv.15009028/v1. Accessed 2026-09-20T16:23:23.853Z.
Source evidence and access
Publisher-deposited abstract, https://api.crossref.org/works/10.26434/chemrxiv.15009028/v1
Reliable use of thermophysical databases in exploratory chemistry requires resolving heterogeneous identifiers, reconciling experimental conditions, and preserving numerical provenance across large, nested records. We present a hierarchical large language model (LLM) agent framework, implemented with a custom ReAct reasoning engine, that translates naturallanguage research questions into traceable retrieval and analysis workflows over the NIST ThermoML database. The framework combines hierarchical canonical identifiers, queryconditioned evidence compaction, deterministic inspection ledgers and numerical-grounding gates, and tool-assisted Redlich–Kister fitting. Five case studies are discussed in detail, representing chemistry tasks spanning data-rich and data-sparse retrieval, binary-mixture correlation, and open-ended cosolvent screening. For an ethanol–water equimolar density query, the system reduced ~190k characters of search output to 2k and precisely recovered a measured density of 859.3 kg m-3 at an ethanol mole fraction of 0.498, 298.15 K, and 100 kPa. For methanolethanol viscosity, it selected condition-aligned mixture and pure-component records, examined the fitting results at different orders, and produced a 2nd-order Redlich–Kister fit with 𝑅 = 0.982. Results for an open-ended iron-electrodeposition electrolyte design query synthesized evidence from approximately 12k measurements across 200 data blocks. However, with numerical grounding disabled for this specific case, the answer contained severe numerical hallucinations despite qualitatively coherent reasoning. These cases establish the feasibility of agent-assisted ThermoML research and identify identity management, evidence compaction, and deterministic numerical provenance as central design requirements for reliable chemistry-database research agents with some level of abstraction chemistry reasoning.
Abstract only: original publisher/repository-deposited metadata retrieved from official Crossref API. Full text not inspected.
Wanchao Chen; Wen Li; Yanan He; Yan Yang. SaltyMeta: a curated benchmark and protein language model-informed web tool for salty peptide prediction. arXiv; 2026; Preprint v1; not peer reviewed. Accessed 2026-09-20T16:23:23.853Z.
Source evidence and access
Abstract and submission history, arXiv:2609.16809v1
Excess sodium intake remains a major public health challenge, while salty and saltiness-enhancing peptides offer a potential route to preserve sensory saltiness in reduced-sodium foods. Machine-learning studies of salty peptides, however, are constrained by small datasets, heterogeneous evidence standards, uncertain negative labels, and sequence similarity leakage. Here we present SaltyMeta, a curated benchmark and web-accessible screening framework for salty or saltiness-enhancing short peptides. The benchmark contains 580 peptides, including 280 positive peptides and 300 negative peptides, all standardized to 2-15 residue one-letter amino-acid sequences. Quality control found no non-standard residues, exact duplicates, or positive-negative overlaps. A similarity-grouped split retained 456 peptides for training and 124 for held-out testing. We evaluated 548 interpretable peptide descriptors, frozen ESM2 embeddings at 8M, 35M, and 150M parameter scales, and descriptor-embedding fusion models under grouped cross-validation. The traditional ExtraTrees baseline selected by training-set grouped cross-validation achieved ROC-AUC=0.693 in cross-validation and ROC-AUC=0.704, PR-AUC=0.702, F1=0.626, and MCC=0.304 on the held-out test set. The best initial protein-language-model fusion was traditional descriptors plus ESM2-8M embeddings, with grouped CV ROC-AUC=0.696 and test ROC-AUC=0.700. Advanced optimization using PCA95 dimensionality reduction and ExtraTrees feature-importance filtering yielded a practical ESM2-8M PCA95 top-300 model with test ROC-AUC=0.715 and PR-AUC=0.703, although repeated-CV gains remained modest. SaltyMeta is a transparent prioritization tool, not a sensory validation substitute. We provide benchmark, models, scripts, GitHub, and Streamlit for reproducible screening of food-derived peptides.
Original arXiv abstract and version history inspected; prior full-text assessment retained in September19 recovery evidence.
Publication record
Published 2026-09-20.
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.
- 2026-09-20 · Version 1c2c6b42 · Viewing this version