Periodic Reporting for period 1 - PATTERNCHEM (Shape and Topology as Descriptors of Chemical and Physical Properties in Functional Organic Materials)
Reporting period: 2022-06-01 to 2024-11-30
Functional, topologically complex organic molecules are rising stars in modern materials science due to their biocompatibility and wealth of physico-chemical properties. Their practical applications often involve interactions with small molecular targets (e.g. gases, environmental pollutants, and drugs) via relatively weak non-covalent forces. Moreover, functional organic materials of today typically possess diverse and complex topologies, made possible by the nearly infinite structural variability of organic chemical motifs. This simultaneously presents a great benefit and an enormous challenge for finding an optimal host for a specific molecular guest in conjunction with a targeted practical use. Serendipitous trial-and-error experimentation becomes prohibitively expensive, prompting the need for theoretical insights to guide rational design and in silico testing of candidate materials. These theoretical insights can be obtained by elucidating and rationalising the pertinent structure-property relationships, for which the structures can be encoded by machine-readable representations, and properties – evaluated by high-throughput workflows or predicted by artificial intelligence. However, existing chemical representations generally reduce the dimensionality of atomic composition and connectivity but do not capture the intricacies of shape and topology, while evaluating materials properties comes at a high computational cost and requires careful benchmarking of the methodology.
The overarching aim of PATTERNCHEM is to enable the rational design and facile pre-screening of functional organic materials for applications involving their non-covalent interactions with molecular targets in silico. Several families of functional organic materials – graphene derivatives, covalent-organic frameworks, and hyperbranched polymers – provide a unique foundation for developing application-oriented fingerprints of their topological and non-covalent interaction features. After elucidating diverse structural descriptors of atomistic arrangement, substitution patterns, and two- and three-dimensional shapes of these materials, we will establish a scheme for quantifying the propensity for non-covalent interactions and assessing the host-guest complementarity. Using this scheme, chemical and physical performance indicators relevant to targeted applications (e.g. as sensors, filters, and nanocarriers) can be computed. Finally, structure-property relationships between computed performance indicators and developed descriptors will be established and implemented into predictive frameworks for functional organic materials. The key deliverable of PATTERNCHEM will be a unified, all-encompassing framework for designing new candidate architectures and evaluating host-guest complementarity, which will require only basic structural information as an input and will predict the ultimate performance in the targeted application as its output.
The overarching aim of PATTERNCHEM is to enable the rational design and facile pre-screening of functional organic materials for applications involving their non-covalent interactions with molecular targets in silico. Several families of functional organic materials – graphene derivatives, covalent-organic frameworks, and hyperbranched polymers – provide a unique foundation for developing application-oriented fingerprints of their topological and non-covalent interaction features. After elucidating diverse structural descriptors of atomistic arrangement, substitution patterns, and two- and three-dimensional shapes of these materials, we will establish a scheme for quantifying the propensity for non-covalent interactions and assessing the host-guest complementarity. Using this scheme, chemical and physical performance indicators relevant to targeted applications (e.g. as sensors, filters, and nanocarriers) can be computed. Finally, structure-property relationships between computed performance indicators and developed descriptors will be established and implemented into predictive frameworks for functional organic materials. The key deliverable of PATTERNCHEM will be a unified, all-encompassing framework for designing new candidate architectures and evaluating host-guest complementarity, which will require only basic structural information as an input and will predict the ultimate performance in the targeted application as its output.
Since the beginning of this project in June 2022, we have made initial progress across all key tasks, ranging from the analysis of the published literature to the development of new in silico tools.
In the area of graphene derivatives, we performed an extensive analysis of the published literature and compiled a database of relevant experimentally reported characteristics of graphene-based adsorbents and filters. We have benchmarked a range of computational approaches for predicting the interaction energies of these materials with small molecules and implemented the best-performing approaches together with a structure generator into a high-throughput multiscale computational workflow. We next used this workflow to compute the interaction energies and other related properties for a newly generated dataset of over 5,500 complexes between graphene derivatives and small molecules.
In the area of covalent-organic frameworks (COFs), we have identified reliable computational approaches for simulating and analysing their interactions with drug molecules and applied these to engineer in silico new frameworks for a delayed release of a broader scope of therapeutics. We have developed an in-house tool for identifying the building blocks of diverse COFs and commenced studies on how the properties of the frameworks arise from their constituents and topologies. We also developed a variant of our in-house chemical representation (see below) for machine learning the properties of COFs and are currently testing it in conjunction with several published datasets. We have also performed a critical evaluation of the approaches for assigning the three-dimensional arrangement of the two-dimensional frameworks from experimentally measured X-ray diffraction data, and started investigating how this arrangement impacts the electronic properties of the frameworks.
In the area of chemical representations, we have developed a quantum-inspired representation called MAOC (Matrix of Orthogonalized Atomic Orbital Coefficients), which is uniquely suited to describe charged and open-shell compounds and can be applied to perform kernel-based and deep learning on atoms, molecules, and extended materials. We have also developed a Matrix of Reference Similarity (MRS) technique for simultaneously reducing the dimensionality of the chemical space and representing the entities in it with a drastically reduced fingerprint, leading to tens to hundreds of times speed up in machine learning. To automatically identify the building blocks in topologically complex molecules and materials and to analyse how these building blocks affect the properties of the materials, we developed a Substructural Filter Representation (SFR).
In the area of graphene derivatives, we performed an extensive analysis of the published literature and compiled a database of relevant experimentally reported characteristics of graphene-based adsorbents and filters. We have benchmarked a range of computational approaches for predicting the interaction energies of these materials with small molecules and implemented the best-performing approaches together with a structure generator into a high-throughput multiscale computational workflow. We next used this workflow to compute the interaction energies and other related properties for a newly generated dataset of over 5,500 complexes between graphene derivatives and small molecules.
In the area of covalent-organic frameworks (COFs), we have identified reliable computational approaches for simulating and analysing their interactions with drug molecules and applied these to engineer in silico new frameworks for a delayed release of a broader scope of therapeutics. We have developed an in-house tool for identifying the building blocks of diverse COFs and commenced studies on how the properties of the frameworks arise from their constituents and topologies. We also developed a variant of our in-house chemical representation (see below) for machine learning the properties of COFs and are currently testing it in conjunction with several published datasets. We have also performed a critical evaluation of the approaches for assigning the three-dimensional arrangement of the two-dimensional frameworks from experimentally measured X-ray diffraction data, and started investigating how this arrangement impacts the electronic properties of the frameworks.
In the area of chemical representations, we have developed a quantum-inspired representation called MAOC (Matrix of Orthogonalized Atomic Orbital Coefficients), which is uniquely suited to describe charged and open-shell compounds and can be applied to perform kernel-based and deep learning on atoms, molecules, and extended materials. We have also developed a Matrix of Reference Similarity (MRS) technique for simultaneously reducing the dimensionality of the chemical space and representing the entities in it with a drastically reduced fingerprint, leading to tens to hundreds of times speed up in machine learning. To automatically identify the building blocks in topologically complex molecules and materials and to analyse how these building blocks affect the properties of the materials, we developed a Substructural Filter Representation (SFR).
We have constructed several new datasets of materials structures and properties and developed a range of new tools for identifying how composition and topology of a material affects its properties, and for rapidly predicting these properties with artificial intelligence. Most of these results from the first 18 months of this project are currently being prepared for publication. Future work involves:
• extensive research on hyperbranched polymers,
• customisation of developed fingerprints towards materials of interest,
• development of the non-covalent interactions descriptor,
• testing the utility of the new tools and descriptors across several materials families,
• an in-depth analysis of the structure-property trends across the generated datasets, and
• collaborations with experimentalists to test our predictions.
• extensive research on hyperbranched polymers,
• customisation of developed fingerprints towards materials of interest,
• development of the non-covalent interactions descriptor,
• testing the utility of the new tools and descriptors across several materials families,
• an in-depth analysis of the structure-property trends across the generated datasets, and
• collaborations with experimentalists to test our predictions.