Chem ML/AI/Datasets

OPRD-100: An Expert-Curated Dataset of Organic Reactions for Benchmarking Reaction Data Mining Tools 🔥 https://chemrxiv.org/doi/full/10.26434/chemrxiv.15011178/v1

We present OPRD-100, an expert-curated and cross-validated dataset of 3,878 reactions extracted from 100 Organic Process Research & Development articles published between 1999 and 2024. Each entry captures detailed reactivity information, including yield, stereochemistry, temperature, time, solvents, and reagents. Analysis of the dataset reveals the relative importance of reaction data reported in tables versus text—particularly for reactions with low or no yields—and highlights the challenges of consistently capturing information from complex studies. Comparison with Reaxys showed systematic omissions and a strong bias toward high-yield reactions, underscoring the risks of relying on uncurated data. OPRD-100 thus provides a robust benchmark for training and evaluating AI tools in industrially relevant organic synthesis. Reaction data, data mining, benchmarking dataset.

Dataset: https://github.com/GHodg1/OPRD-100

Вчера у нас с замечательными коллегами из ФТИ Иоффе🏛 вышла новая работа. Dataset of magnetic properties, structure and synthesis of doped barium and strontium M-type hexaferrites https://doi.org/10.26434/chemrxiv.15011486/v1 M-гексаферриты бария и стронция (Ba/SrFe₁₂₋ₓMₓO₁₉) — самые массовые материалы для постоянных магнитов. Их структуру и магнитные свойства обычно подстраивают замещением катионов. Экспериментальных данных по допированным гексаферритам много, но они разбросаны по сотням статей и записаны в разных обозначениях и единицах. Открытой базы, на которой можно обучать модели, до сих пор не было. Мы собрали HexaMagDB. Для нас это по сути первый опыт создания баз по материалам :) Объём базы: 4388 записей из 662 статей. 48 допантов в 251 комбинации. Одна запись — один охарактеризованный образец. В записи есть: — формула, допанты, их количество и замещаемая подрешётка; — метод синтеза с описанием процедуры, условия прокаливания и спекания; — параметры решётки, размеры кристаллитов и частиц, фазовая чистота и вторичные фазы; — Ms, Mr, Hc, температура Кюри, поле и константа анизотропии; — температура измерения и максимальное приложенное поле. Все данные в базе проверены по физическим критериям и допустимым диапазонам. Для работы с базой сделан сайт: https://hexamagdb.streamlit.app/. На нём можно выбирать допанты на периодической таблице и смотреть записи по ним. Датасет как обычно доступен на Zenodo: https://doi.org/10.5281/zenodo.23807416

Democratizing Drug Discovery: A Scalable Open-Source Pipeline for Massive Specific Protein Target QSAR Modeling Using ChEMBL 🔥 https://doi.org/10.1021/acs.jmedchem.6c00791

We present an open-source, automated pipeline for large-scale Quantitative Structure−Activity Relationship (QSAR) modeling utilizing ChEMBL. Our batch processing system constructed 2263 predictive models across multiple protein targets, applying both regression and classification strategies with strict quality control thresholds. The pipeline integrates ErG and RDKit descriptor families, employs Random Forest and XGBoost algorithms with Optuna-based hyperparameter optimization, and handles class imbalances using SMOTE. Models are saved only when achieving R2 ≥ 0.7 for regression or MCC ≥ 0.55 for classification.

📕Journal of Medicinal Chemistry (IF=7.3)

Automating MD simulations for proteins using large language models: NAMD-agent 🔥 https://doi.org/10.1039/d6dd00535g

Molecular dynamics simulations are an essential tool in understanding protein structure, dynamics, and function at the atomic level. However, preparing high-quality input files, running, and analyzing MD simulations can be time-consuming and error-prone. In this work, we introduce an automated pipeline that leverages Large Language Models (LLMs) in conjunction with Python scripting to streamline the generation of MD input files. The pipeline exploits HTMD's comprehensive MCP tools interface for preparing simulation-ready inputs for NAMD. We extend the agentic paradigm in two directions that matter for practical structural biology: (i) we cover five distinct biomolecular protein system classes rather than a single workflow, and (ii) we close the loop through automated analysis and benchmark comparison, including self-diagnosis and recovery from build- and runtime failures. Operating an LLM agent over a NAMD3/CHARMM tool server, we built, equilibrated, simulated, and analyzed a protein–DNA complex (p53 DNA-binding domain on its response element), a protein–membrane system (the M2 muscarinic receptor with the agonist iperoxo in a POPC/cholesterol bilayer), a protein–ligand series (five congeneric TYK2 inhibitors), a protein–water reference (ubiquitin), and a protein–protein complex (barnase–barstar). For instance, we compared the protein–DNA complex against an independent published benchmark and successfully reproduced the reference metrics. To separate the framework from the model that drives it, we further repeated the complete protein-DNA study under three different LLM orchestrators Claude Opus-4.8, GPT-5.6 Sol, and the open-source Nemotron 3 Ultra and found that all three completed the workflow but differed in how faithfully they reproduced the benchmark ranking and by up to two orders of magnitude in token consumption and cost. The agent recovered the experimentally and computationally established behavior in every class. Post-processing was performed using additional software to further refine the simulation outputs, thereby enabling a complete and largely hands-free workflow. Our results demonstrate that this approach reduces setup time, minimizes manual errors, and offers a scalable solution for handling multiple protein systems in parallel. This automated framework paves the way for broader application of LLMs in computational structural biology, offering a robust and adaptable platform for future developments in simulation automation.

📕Digital Discovery (IF=7.1)

Defining Applicability Domains for Metal–Ligand Selectivity Prediction: Predictive Uncertainty vs. Molecular Similarity🔥 https://chemrxiv.org/doi/full/10.26434/chemrxiv.15010241/v1

Applicability domains are widely used to determine whether a machine-learning prediction can be trusted, yet they are commonly defined by similarity to the training data rather than by evidence that similarity predicts error. Here, we treat an applicability criterion as a testable claim and evaluate whether it identifies predictions with lower error than an equally sized random subset on a common held-out population. We test 23 criteria on a context-dependent metal–ligand selectivity problem, comparing molecular fingerprints, descriptor-space distances, learned embeddings, training support and predictive uncertainty. At 50% retention, predictive uncertainty reduces mean-of-pairs RMSE by 65%, compared with 32% for the best molecular distance and 18% for the best ligand-only distance. Fingerprint nearest-neighbour and learned-embedding distances provide no useful error ordering, whereas uncertainty remains informative after accounting for training support and predicted magnitude. In a prospective-style selection task, ranking ligands by their standardized predicted margin above a selectivity threshold increases single-ligand success from 0.639±0.035 under random selection to 0.831± 0.059. These results show that, at least for the context-dependent metal-ligand selectivity task studied here, structural similarity to the training data is not, by itself, evidence of predictive reliability in context-dependent molecular prediction. Applicability criteria can instead be treated as empirical hypotheses and tested directly against held-out error before being used to define reliable prediction regions or benchmark difficulty.

Physics-based molecular fingerprints from spectral graph theory provide efficient geometry-aware measures of chemical similarity 🔥 https://doi.org/10.1039/d6sc06309h

Here, we introduce novel physics-inspired molecular fingerprints based on principles from spectral graph theory. We represent a molecule as a complete graph in 3D space, with edge weights encoding heuristic physical interactions. Eigenvalue decomposition of the resulting graph Laplacian matrix results in a computationally efficient fixed-length chemical fingerprint that encodes 3D structure while obeying necessary physical symmetries of permutation and E(3) invariance. Spectral fingerprints differentiate between unique molecular structures with identical 2D connectivity, overcoming a limitation of 2D descriptors, while maintaining the low computational cost needed for efficient screening of vast chemical spaces. We evaluate our fingerprints with community detection algorithms and observe strong performance against representative baselines across datasets from organic, inorganic, biological, reticular, and reaction chemistry.

📕Chemical Science (IF=8.1)

Major Revision, Minor Accountability: LLMs as Reviewers of Published Articles in Top-Tier Journals🔥 https://doi.org/10.48550/arXiv.2610.31414 Даже если статья уже была принята к публикации в топовом журнале, ЛЛМ-рецензент с большой вероятностью направит ее на переработку. При этом далеко не факт, что модели заметят категоричную методологическую ошибку, а вместо этого могут накинуть ничего глобально не меняющих замечаний. Задумайтесь, когда в следующий раз отправите статью, которую согласились рецензировать, клоду с просьбой написать рецензию.

With the growth of scientific output, the peer review system is approaching a breaking point: reviewers are increasingly hard to secure, and turnaround times lengthen accordingly. LLM-assisted or LLM-generated peer review has been proposed as a remedy, but how such models would treat manuscripts already published in top-tier journals, the closest thing the current system has to a gold standard of scientific validity, is not known. Here we assembled a corpus of 200 articles, each released under a CC BY 4.0 license and published in 2026, from 20 general-science, life-science, and clinical-medicine journals including Nature, Science, and Cell, and asked four current LLMs (two proprietary, two open-weight) to produce independent peer-review reports of each one under a blinded protocol. Their decisions were overwhelmingly harsh and starkly inconsistent with each other: 605 of 800 reviews (75.6%) recommended major revision or rejection, but the rate at which any single model did so ranged from 17.5% to 100% (p < 0.001), disagreement on manuscripts that had, by construction, already survived real editorial review at the journals in question. When the same manuscripts were doctored with one deliberately planted, field-specific methodological or statistical flaw, detection rates varied from 17.0% to 60.5% across models (p < 0.001). Recommending revision for papers that had already passed editorial review (low specificity) while missing flaws deliberately placed in front of them (low sensitivity), these models produce output that should be read for what it is: an uneven, framing- and model-dependent signal.

Causal-Chemprop: Causal Machine Learning for Molecular Property Prediction and Design 🔥 https://doi.org/10.1021/acs.jcim.6c01429

Herein, we introduce a causal machine learning framework built on the Chemprop and DAGMA architectures for molecular property prediction called Causal-Chemprop. Its capacity for counterfactual reasoning supports human-in-the-loop optimization of molecular structure, which we demonstrate by improving the rank-ordering of out-of-distribution kinase inhibitor pIC50 values─most where standard encoders extrapolate poorly─both within and across scaffold series, in a lead-optimization setting in which the test scaffold is known. We further integrate Causal-Chemprop with the molecular optimization algorithm EvoMol as a scoring function, where it tends to yield more synthesizable molecules that more closely resemble the sought-after active compound than Chemprop. To our knowledge, this is the first application bridging neural network-learned molecular representations with structural causal models for molecular property prediction and optimization via counterfactual reasoning, suggesting several promising new avenues for exploration in molecular property prediction and design.

📕Journal of Chemical Information and Modeling (IF=6.4)

Новый OdanChem Коллеги, мы обновили OdanChem: новый интерфейс и переработанные поиски. Как оно выглядит, показали в коротком видео. А вот, что оно умеет: 1) Поиск по соединению. В строку можно вставить IUPAC-название, CAS, SMILES или картинку-скриншот. По найденной молекуле сразу видны ЯМР, HRMS, температура плавления и публикации, где она упоминалась. Можно искать как точную структуру, так и по фрагменту структуры. 2) Поиск по реакциям (BETA) — по реагентам или продуктам, с фильтром по выходу и другим параметрам. 3) Спектральный поиск по ЯМР. Вставляете описание спектра из статьи или просто загружаете папку с прибора. Система разбирает его на пики и показывает, сколько из них совпало с найденным спектром. 4) ИИ-расшифровка ЯМР! Модель BLIND предлагает структуру по спектрам ¹H и ¹³C, даже если вы первый, кто эту молекулу получил. 5) Удобная история поиска, где результаты можно посмотреть, объединить и пересечь. В базе больше 60 млн молекул и реакций, больше 20 млн ЯМР-спектров по 37 ядрам и куча других свойств. И ещё кое-что) на странице входа теперь живёт Одан, и его можно увлекательно гонять по странице) Пробуйте: https://odanchem.org Баги, вопросы и пожелания пишите в комментарии или на support@odanchem.org.

Molecular Property Prediction with Well-Calibrated Uncertainties via Graph Neural Networks and Conformal Quantile Regression🔥 https://chemrxiv.org/doi/full/10.26434/chemrxiv.15010324/v1

Accurate predictive uncertainty quantification (pUQ) is critical for assessing chemical property predictions. Incorrect quantification could lead to costly synthesis and characterization of drug candidates destined to fail, or result in misguided environmental regulations. This work integrates Conformal Quantile Regression (CQR) with graph neural networks to achieve accurate molecular property prediction and well-calibrated predictive uncertainty quantification. By combining conformal prediction with quantile regression, CQR provides model-agnostic, distribution-free prediction intervals with valid coverage for finite samples. We benchmark it against established pUQ approaches, including ensembling, Monte Carlo dropout sampling, mean-variance estimation, and evidential regression. Our experiments on datasets spanning quantum mechanics, physical chemistry, and pharmacokinetics show that CQR generally provides strong calibration while maintaining competitive predictive accuracy and uncertainty-ranking performance. It also effectively captures heteroskedastic noise across the data, which is a key requirement for robust pUQ. Furthermore, we demonstrate its efficacy with an uncertainty-guided retrospective virtual screening experiment for antibiotic discovery, achieving hit rates above 90%.

Advancing chemical reaction prediction in data-scarce drug discovery with active and geometric deep learning🔥 https://www.nature.com/articles/s43588-026-01063-0

Here, to address this, we present an integrated closed-loop workflow that combines efficient reaction data acquisition with high-accuracy deep learning. We developed an active learning strategy, driven by a tree-based ensemble, to guide the acquisition of the most informative laboratory experiments. This experimental campaign yielded a diverse benchmark for C–H borylation. The bespoke dataset allowed us to train and evaluate geometric graph neural networks. We demonstrate that augmenting these symmetry-aware models with self-supervised auxiliary tasks consistently improves performance for both reaction outcome prediction and regioselectivity prediction. Prospective tests on unseen substrates, featuring challenging N-heteroaryl motifs, pinpointed the correct borylation positions in all cases, illustrating the practical utility of our workflow for efficient drug optimization.

📕 Nature computational science (IF = 20.3)

Сегодня у нас вышла новая работа: Beyond Organic Molecules: MetalLipoDB, a Lipophilicity Dataset for Metal Complexes and Machine Learning Benchmarks https://doi.org/10.26434/chemrxiv.15010021/v1 Работа продолжает систематизацию данных по комплексам металлов, начатую в MetalCytoToxDB. На этот раз предметом стала липофильность. Для органических соединений существуют большие открытые базы logP, а предсказание липофильности входит в стандартные ML-бенчмарки. Для комплексов металлов экспериментальные значения разбросаны по сотням статей, и существующие подборки покрывают преимущественно платину. В работе представлена MetalLipoDB — крупнейшая на сегодня база экспериментальных значений липофильности комплексов металлов. Объём базы: 5158 значений logP/logD для 4714 комплексов 23 металлов, вручную извлечённых из 1192 статей. Для каждой записи указаны условия измерения: состав водной фазы, pH, метод детекции. В базу включены 546 многоядерных комплексов. Оценена воспроизводимость измерений. Для одного и того же комплекса значения из двух разных статей расходятся в среднем на 0.56 логарифмических единиц. Эта величина задаёт практический предел точности моделей, обученных на литературных данных. На основе базы обучены ML-модели (TabPFN, LightGBM, Chemprop, CheMeleon). Для комплексов из статей, отсутствующих в обучающей выборке, ошибка составляет MAE = 0.72-0.75. Если в серии уже измерены три комплекса, привязка предсказаний к этим измерениям снижает ошибку с 0.73 до 0.51. ❗️В разделе "Limitations and future work" перечислены ограничения подхода: — Значения в базе представляют собой кажущиеся коэффициенты распределения в условиях, указанных авторами — Многие комплексы не инертны в воде: в частности, хлоридные комплексы платины и рутения подвергаются замещению лиганда на воду — Включены только измерения непосредственно распределения между октанолом и водной фазой, хроматографические индексы липофильности не рассматриваются — Модели не учитывают стереохимию — Для металлов, отсутствующих в обучающей выборке, ошибка возрастает до 0.83-0.89 Датасет на Zenodo: https://doi.org/10.5281/zenodo.23084573

Ложился спать, открыл твиттер, почитал и не могу уснуть… OpenAI опубликовала разом 722 манускрипта 🫪 с решениями 722 открытых математических задач, которые выполнила их (пока ещё) закрытая модель, каждое за ~3 часа. Что ж, ждем грозных писем от Теренса Тао и десятков филдсовских/абелевских лауреатов. Но их не жалко, они свою жизнь пожили, жалко того бедного неназванного аспиранта, который, возможно, последние три года сидел и корпел над решением одной из этих задач и вот-вот готовился опубликовать статью, которая уже никому не нужна, как и его занятия фундаментальной математикой «по старинке». Ожидаю резкий спад интереса к поступлению в магистратуры/аспирантуры на матфаки по всему миру, пока светлые головы не переизобретут математический post-graduate и не сформулируют: «что такое занятия математикой в эпоху ИИ?». А что химики? За химиками тоже «придут», обязательно придут, «но потом».

Вчера вышла совместная работа AMD Silo AI и AstraZeneca From Benchmark to Bench: Can Agents Survive Real-World Drug Discovery?🔥 https://arxiv.org/abs/2610.06411

We developed MAGI, an open modular agent that authors objectives, launches and monitors optimization, interprets structure-activity relationships, and revises its strategy accordingly. MAGI generates molecules either directly through the LLM or by delegating to REINVENT 4, with scoring services interchangeable behind a common contract. We tested it across nine retrospective lead-optimization campaigns from three pharmaceutical companies, replayed under fixed temporal cutoffs. Both routes produced valid structures: LLM proposals stayed closer to local chemistry and reached comparable or higher primary activity in fewer operations, whereas REINVENT explored broader chemical space. Whether a campaign met its objective depended on the predictive models, not on the generation route: attainment followed model accuracy on the chemistry proposed, dropping once that chemistry moved outside the model's applicability domain. Separately, a blinded evaluation asked whether the MAGI's output could pass as expert work: chemists were not able to discriminate agentic proposals from held-out compounds, and judged the SAR reasoning broadly plausible yet incomplete. Together, these results position MAGI as a coordination layer pluggable into existing computational chemistry workflows. The ceiling on real projects, however, remains currently set by scorer applicability rather than by tool orchestration.

Atoms to Processes: The Role of Artificial Intelligence and Machine Learning in Chemical Engineering https://doi.org/10.1021/acs.iecr.6c01806

This Perspective provides an account of recent methodological developments and representative applications, emphasizing how AI/ML tools are being integrated with first-principles models to address challenges of predictive accuracy, data scarcity, extrapolation, interpretability, and model life cycle management. Across domains, a unifying trend is the move away from purely black-box approaches toward hybrid and physics-informed frameworks that explicitly respect conservation laws, thermodynamic consistency, and known structural constraints. These approaches not only improve robustness and reliability but also enable meaningful human–AI collaboration by providing information at an appropriate level of abstraction for the task and decision context. We conclude that AI and ML are not replacing the core principles of chemical engineering; rather, they are amplifying them.

📕Industrial & Engineering Chemistry Research (IF = 3.9)

ElemeNet: Multiscale Molecular Machine Learning with Uncertainty Quantification across the Periodic Table https://doi.org/10.1021/acs.jcim.6c02178

This work introduces ElemeNet, a unified, general-purpose software package for molecular machine learning. The ElemeNet software package enables the training of advanced ML models for diverse properties and data sets with an enlarged range of elemental compositions. We define molecular representations compatible with elements 1–100, supporting diverse organometallic and biological systems in addition to organic chemistry already well served by the Chemprop ML toolkit. As well as more common atom-, bond-, and molecule-level predictions, we introduce and allow for moiety-level predictions. We also natively define optional conditioning on charge and spin states. Advanced E(3)-equivariant and transformer architectures are supported in addition to 2D models, with all classes including built-in uncertainty quantification through deterministic and statistical measures. We benchmark our protocols for ML model training against representative data sets from organic, inorganic, coordination, and biological chemistry, achieving competitive and SOTA performance relative to literature baselines and favorable scaling to millions of molecules.

📕Journal of Chemical Information and Modeling (IF=6.4)

Разбирал закладки и понял, что так и не запостил эту статью, хотя она вышла почти год назад. Исправляюсь. Это статья в JCIM о том, как честно сравнивать ML-модели в drug discovery. Practically Significant Method Comparison Protocols for Machine Learning in Small Molecule Drug Discovery🔥 https://doi.org/10.1021/acs.jcim.5c01609

This paper proposes a set of guidelines to incentivize rigorous and domain-appropriate techniques for method comparison tailored to small molecule property modeling. These guidelines, accompanied by annotated examples using open-source software tools, lay a foundation for robust ML benchmarking and thus the development of more impactful methods.

Почти все авторы из индустрии: J&J, Novartis, Bayer, Merck, Pfizer, AstraZeneca, Recursion и т.д. Последний автор Pat Walters теперь Chief Scientist в OpenADMET, так что эти правила, скорее всего, дойдут и до открытых бенчмарков.

Доброе утро начинается с принятия работы на NeurIPS 2026 (Evaluations & Datasets Track), в которой нас пригласили поучаствовать замечательные коллеги из Индии (IIT Delhi) 🎉

We are delighted to inform you that your submission, "SC³: A Multi-Solvent Solubility Challenge and Benchmark", has been accepted at NeurIPS 2026 ED Track. Congratulations! There were 3,757 valid paper submissions with a PDF to the NeurIPS ED Track this year, of which the program committee accepted 971 (25.8%) papers in total.

📄 arxiv.org/abs/2606.07656

Multi-objective optimization in the context of generative chemistry 🔥 https://www.nature.com/articles/s41467-026-77324-3

Designing new medicines requires balancing multiple competing objectives, including biological activity, pharmacokinetics and safety. As generative chemistry expands our ability to explore chemical space, the choice of how objectives are defined, combined and evaluated becomes increasingly important. In this Perspective, we discuss the role of multi-objective optimization in molecular generation, highlighting practical considerations, methodological challenges and emerging opportunities for translating these approaches into effective drug discovery tools.

📕 Nature Communications (IF=18.1)