Molecular Systems Design & Engineering最新文献

We report the use of a multiagent generative artificial intelligence framework, the X-LoRA-Gemma large language model (LLM), to analyze, design and test molecular design. The X-LoRA-Gemma model, inspired by biological principles and featuring 7 billion parameters, dynamically reconfigures its structure through a dual-pass inference strategy to enhance its problem-solving abilities across diverse scientific domains. The model is used to first identify molecular engineering targets through a systematic human-AI and AI-AI self-driving multi-agent approach to elucidate key targets for molecular optimization to improve interactions between molecules. Next, a multi-agent generative design process is used that includes rational steps, reasoning and autonomous knowledge extraction. Target properties of the molecule are identified either using a principal component analysis (PCA) of key molecular properties or sampling from the distribution of known molecular properties. The model is then used to generate a large set of candidate molecules, which are analyzed via their molecular structure, charge distribution, and other features. We validate that as predicted, increased dipole moment and polarizability is indeed achieved in the designed molecules. We anticipate an increasing integration of these techniques into the molecular engineering workflow, ultimately enabling the development of innovative solutions to address a wide range of societal challenges. We conclude with a critical discussion of challenges and opportunities of the use of multi-agent generative AI for molecular engineering, analysis and design.

Retraction of ‘Heteroatoms chemical tailoring of aluminum nitrite nanotubes as biosensors for 5-hydroxyindole acetic acid (a biomarker for carcinoid tumors): insights from a computational study’ by Chioma B. Ubah et al., Mol. Syst. Des. Eng., 2024, 9, 832–846, https://doi.org/10.1039/D4ME00019F.

We here report rodlike liquid-crystalline (LC) molecules consisting of bicyclohexyl and linear carbonate moieties connected through flexible spacers for the development of nanostructured ion-conductive materials. The molecular assemblies of the linear carbonate-based rodlike compounds mixed with a lithium salt provide 2D ion-conductive pathways in the smectic LC phases. The LC materials containing polar oligooxyethylene spacers coupled with linear carbonate moieties have been shown to function as efficient ion conductors, while those containing nonpolar alkylene spacers form thermally stable and ordered smectic LC structures. Molecular dynamics simulations provide insights into the conformation and packing of the molecules containing oligooxyethylene spacers in the LC phases. The combination of flexible oligooxyethylene chains and linear carbonates may lead to design of new LC electrolytes with highly mobile 2D nanochannels for applications in energy devices.

Adsorption-based processes are showing substantial potential for carbon capture. Due to the vast space of potential solid adsorbents and their influence on the process performance, the choice of the material is not trivial but requires systematic approaches. In particular, the material choice should be based on the performance of the resulting process. In this work, we present a method for the process-based screening of porous materials for pressure and vacuum swing adsorption. The method is based on an equilibrium process model that incorporates one-dimensional classical density functional theory (1D-DFT) and the PC-SAFT equation of state. Thereby, the presented method can efficiently screen databases of potential adsorbents and identify the best-performing materials as well as the corresponding optimized process conditions for a specific carbon capture application. We apply our method to a point-source carbon capture application at a cement plant. The results show that the process model is crucial to evaluating the performance of adsorbents instead of relying solely on material heuristics. Furthermore, we enhance our approach through multi-objective optimization and demonstrate for materials with high performance that our method is able to capture the trade-offs between two process objectives, such as specific work and purity. The presented method thus provides an efficient screening tool for adsorbents to maximize process performance.

Phase separation in multicomponent mixtures is of significant interest in both fundamental research and technology. Although the thermodynamic principles governing phase equilibria are straightforward, practical determination of equilibrium phases and constituent compositions for multicomponent systems is often laborious and computationally intensive. Here, we present a machine-learning workflow that simplifies and accelerates phase-coexistence calculations. We specifically analyze capabilities of neural networks to predict the number, composition, and relative abundance of equilibrium phases of systems described by Flory–Huggins theory. We find that incorporating physics-informed material constraints into the neural network architecture enhances the prediction of equilibrium compositions compared to standard neural networks with minor errors along the boundaries of the stable region. However, introducing additional physics-informed losses does not lead to significant further improvement. These errors can be virtually eliminated by using machine-learning predictions as a warm-start for a subsequent optimization routine. This work provides a promising pathway to efficiently characterize multicomponent phase coexistence.

Soft porous coordination polymers (SPCPs) are flexible porous materials comprised of metal–organic polyhedrons (MOPs) connected by organic linkers, with potential in adsorption applications. We performed molecular simulations of various SPCPs that vary in the length and flexibility of the organic linkers to address how the flexibility can result in various configurations and affects adsorption performance. We examined free energy profiles as a function of volume of different SPCPs while varying methane loading, resulting in different stable configurations. We found significant differences in the volume of the stable configurations and their number for the various structures, with more flexible linkers having more stable configurations in free energy. We also characterized the textural properties and methane adsorption isotherms of the stable configurations for the SPCPs and analyzed density profiles of the adsorption in the various configurations. Altogether, our examination can be used to predict the relevant configurations of the SPCPs at a given loading and provides molecular-level understanding of how the flexibility of the organic linkers affects the structure of the system and adsorption performance.

An efficient screening of azobenzene (AB) derivatives for Molecular Solar Thermal (MOST) applications based on ground state properties (energy stored per molecule and Z isomer stability) could be performed with quasi-CASPT2 accuracy. In this work, we show how wavefunction and electron density based methods can be efficiently combined in a computational protocol that yields accurate potential energy profiles with a significant reduction in computational cost compared to that of a fully-CASPT2 characterization. Our results on prototypical electron donor/withdrawing AB derivatives clearly identify pull–pull substitution as the most promising, allowing to draw guidelines for the chemical design of promising azo-MOST candidates.

Peptides are a powerful class of molecules that can be applied to a range of problems including biomaterials development and drug design. Currently, machine learning-based property prediction models for peptides primarily rely on amino acid sequence, resulting in two key limitations: first, they are not compatible with non-natural peptide features like modified sidechains or staples, and second, they use human-crafted features to describe the relationships between different amino acids, which reduces the model's flexibility and generalizability. To address these challenges, we have developed PepMNet, a deep learning model that integrates atom-level and amino acid-level information through a hierarchical graph approach. The model first learns from an atom-level graph and then generates amino acid representations based on the atomic information captured in the first stage. These amino acid representations are then combined using graph convolutions on an amino acid-level graph to produce a molecular-level representation, which is then passed to a fully connected neural network for property prediction. We evaluated this architecture by predicting two peptide properties: chromatographic retention time (RT) as a regression task and antimicrobial peptide (AMP) activity as a classification task. For the regression task, PepMNet achieved an average R² of 0.980 across eight datasets, which spanned different dataset sizes and three liquid chromatography (LC) methods. For the classification task, we developed an ensemble of five models to reduce overfitting and ensure robust classification performance, achieving an area under the receiver operating curve (AUC-ROC) of 0.978 and an average precision of 0.981. Overall, our model illustrates the potential for hierarchical deep learning models to learn peptide properties without relying on human engineering amino acid features.

Two-photon 3D laser printing (2PLP) is one of the most versatile methods for additive manufacturing of micro- to nano-scale objects with arbitrary geometries and fine features. With advancing technological capability and accessibility, the demand for new and versatile inks is increasing, with a trend toward printing functional or responsive structures. One approach for ink design is the use of a macromolecular ink consisting of a ‘pre-polymer’ functionalized with photocrosslinkable groups to enable printability. However, so far the synthesis of pre-polymer inks for 2PLP often relies on an arbitrary choice rather than systematic design. Additionally, current structure–property relationship studies are limited to commercial or small molecule-based inks. Herein, three macromolecular inks with varied compositions, molecular weights, and glass transition temperatures are synthesized and formulated into inks for 2PLP. 3D microstructures are fabricated and characterized in-depth with scanning electron microscopy as well as infrared spectroscopy and nanoindentation to enable the determination of structure–processability–property relationships. Overall, it is clearly demonstrated that the macromolecular design plays a role in the printability and mechanical properties of the obtained materials.

Machine learning models have been progressively used for predicting materials' properties. These models can be built using pre-existing data and are useful for rapidly screening the physicochemical space of a material, which is astronomically large. However, ML models are inherently interpolative, and their efficacy for searching candidates outside a material's known range of properties is unresolved. Moreover, the performance of an ML model is intricately connected to its learning strategy and the volume of training data. Here, we determine the relationship between the extrapolation ability of an ML model, the size and range of its training dataset, and its learning approach. We focus on a canonical problem of predicting the properties of a copolymer as a function of the sequence of its monomers. Tree search algorithms, which learn the similarities between polymer structures, are found to be inefficient for extrapolation. Conversely, the extrapolation capability of neural networks and XGBoost models, which attempt to learn the underlying functional correlation between the structure and properties of polymers, shows strong correlations with the volume and range of training data. These findings have important implications on ML-based new material development.