Molecular Machine Learning for Chemical Catalysis: Prospects and Challenges

Accounts of Chemical Research, Год журнала: 2023, Номер 56(3), С. 402 - 412

Опубликована: Янв. 30, 2023

ConspectusIn the domain of reaction development, one aims to obtain higher efficacies as measured in terms yield and/or selectivities. During empirical cycles, an admixture outcomes from low high yields/selectivities is expected. While it not easy identify all factors that might impact efficiency, complex and nonlinear dependence on nature reactants, catalysts, solvents, etc. quite likely. Developmental stages newer reactions would typically offer a few hundreds samples with variations participating molecules conditions. These "observations" their "output" can be harnessed valuable labeled data for developing molecular machine learning (ML) models. Once robust ML model built specific under predict outcome any new choice substrates/catalyst seconds/minutes thus expedite identification promising candidates experimental validation. Recent years have witnessed impressive applications world, most them aimed at predicting important chemical or biological properties. We believe integration effective workflows made richly beneficial discovery.As technology, direct adaptation used well-developed domains, such natural language processing (NLP) image recognition, unlikely succeed discovery. Some challenges stem ineffective featurization space, unavailability quality its distribution, making right technically deployment. It shall noted there no universal suitable inherently high-dimensional problem reactions. Given these backgrounds, rendering tools conducive exciting well challenging endeavor same time. With increased availability efficient algorithms, we focused tapping potential small-data discovery (a thousands samples).In this Account, describe both feature engineering approaches applied diverse contemporary interest. Among these, catalytic asymmetric hydrogenation imines/alkenes, β-C(sp3)–H bond functionalization, relay Heck employed approach using quantum-chemically derived physical organic descriptors features─all designed enantioselectivity. The selection features customize interest described, along emphasizing insights could gathered through use features. Feature methods Buchwald–Hartwig cross-coupling, deoxyfluorination alcohols, enantioselectivity N,S-acetal formation are found excellent predictions. propose transfer protocol, wherein trained large number (105–106) fine-tuned library target task reactions, alternative (102–103 reactions). exploitation deep neural network latent space method generative tasks useful substrates demonstrated strategy.

Язык: Английский

---

Guided diffusion for inverse molecular design

Nature Computational Science, Год журнала: 2023, Номер 3(10), С. 873 - 882

Опубликована: Окт. 5, 2023

Язык: Английский

---

De Novo Design of Nurr1 Agonists via Fragment-Augmented Generative Deep Learning in Low-Data Regime

Journal of Medicinal Chemistry, Год журнала: 2023, Номер 66(12), С. 8170 - 8177

Опубликована: Май 31, 2023

Generative neural networks trained on SMILES can design innovative bioactive molecules de novo. These so-called chemical language models (CLMs) have typically been tens of template for fine-tuning. However, it is challenging to apply CLM orphan targets with few known ligands. We fine-tuned a single potent Nurr1 agonist as in fragment-augmented fashion and obtained novel agonists using sampling frequency prioritization. Nanomolar potency binding affinity the top-ranking its structural novelty compared available ligands highlight value an early tool lead development, well applicability very low-data scenarios.

Язык: Английский

---

Data-Driven Elucidation of Flavor Chemistry

Journal of Agricultural and Food Chemistry, Год журнала: 2023, Номер 71(18), С. 6789 - 6802

Опубликована: Апрель 27, 2023

Flavor molecules are commonly used in the food industry to enhance product quality and consumer experiences but associated with potential human health risks, highlighting need for safer alternatives. To address these health-associated challenges promote reasonable application, several databases flavor have been constructed. However, no existing studies comprehensively summarized data resources according quality, focused fields, gaps. Here, we systematically 25 molecule published within last 20 years revealed that inaccessibility, untimely updates, nonstandard descriptions main limitations of current studies. We examined development computational approaches (e.g., machine learning molecular simulation) identification novel discussed their major regarding throughput, model interpretability, lack gold-standard sets equitable evaluation. Additionally, future strategies mining designing based on multi-omics artificial intelligence provide a new foundation science research.

Язык: Английский

---

Current and emerging deep-learning methods for the simulation of fluid dynamics

Proceedings of the Royal Society A Mathematical Physical and Engineering Sciences, Год журнала: 2023, Номер 479(2275)

Опубликована: Июль 1, 2023

Over the last decade, deep learning (DL), a branch of machine learning, has experienced rapid progress. Powerful tools for tasks that have been traditionally complex to automate developed, such as image synthesis and natural language processing. In context simulating fluid dynamics, this led series novel DL methods replacing or augmenting conventional numerical solvers. We broadly classify these into physics- data-driven methods. Physics-driven methods, generally, tune model provide an analytical differentiable solution given dynamics problem by minimizing residuals governing partial differential equations. Data-driven fast approximate any shares some physical properties with observations used when tuning model’s parameters. Meanwhile, symbiosis solvers promising results in turbulence modelling accelerating iterative However, present challenges. Exclusively flow simulators often suffer from poor extrapolation, error accumulation time-dependent simulations, well difficulties training against turbulent flows. Substantial effort is, therefore, being invested approaches may improve current state art.

Язык: Английский

---