The Application of Artificial Intelligence and Machine Learning in Chemistry

1. Introduction to Artificial Intelligence and Machine Learning

Artificial intelligence (AI) and machine learning (ML) are becoming increasingly important in the field of chemistry. AI can be defined as the portrayal of human cognition, such as visual perception, thought, and learning, by computer systems. Machine learning (ML) is a subset of AI that focuses on the development of computer programs that change automatically based on data. In the context of chemistry, the application of AI/ML includes process control, reaction prediction, peak identification, spectral prediction, structure and property prediction, and computational chemistry.

With the growth of new technologies in chemistry, the amount of generated data is increasing rapidly. As a result, the effective usage of big data becomes important, and the need for new technologies arises. AI/ML techniques can examine the data and make predictions by generalizing those data. Similarity-based approaches can be used to search for already known data, which leads to reasonable predictions. AI/ML methods are not new, but due to the development of computer technologies, their scope has expanded rapidly in recent years.

There are five types of AI, ranging from fully reactive to full self-awareness. The most commonly known form of AI is the ability of machines to process data, make predictions, and learn from experience. AI/ML techniques consist of supervised AI/ML and unsupervised AI/ML. Supervised methods require continuous learning and analysis to provide feedback to trained models, while unsupervised methods analyze data without guidance and explore data for patterns. [1][2][3][4][5][6]

2. Fundamentals of Chemistry Relevant to AI and ML Applications

The discipline of chemistry encompasses the study of chemical systems at various scales, from atomic and molecular to macroscopic systems. Chemistry is a vast field, usually categorized into analytical chemistry, biochemistry, inorganic chemistry, materials chemistry, molecular chemistry, and physical chemistry. Analytical chemistry focuses on the detection and quantification of chemical substances. Biochemistry studies the structures and functions of biomolecules and enzymes. Inorganic chemistry deals with the design and properties of inorganic chemical systems. Materials chemistry is the science of design, synthesis, and understanding of novel materials. Molecular chemistry involves the synthesis, characterization, and understanding of small organic molecules. Physical chemistry establishes the relationship between molecular properties and macroscopic properties of the system.

Machine learning senses similarities in the chemical systems based on simple, easier computable features. Developing such a model requires knowledge of the chemistry to identify molecular features that correlate to the chemical system property of interest. It also requires knowledge of the chemical systems and the corresponding property of interest to provide a training set for the model. Such knowledge also helps to develop a rather simple model with good performance instead of a complex model that does not work well.

Electronic structure theory is the key discipline in computational chemistry that computes potential energy surfaces, molecular properties as a function of atomic coordinates, and chemical systems as a function of time. At the moment, density functional theory (DFT) and wave function-based methods in the family of Hartree Fock (HF) methods, perturbation theory (MP2 and higher order), coupled cluster theory (CC), configuration interaction (CI), and many-body theory are the most common electronic structure methods. In addition, there are semi-empirical quantum chemical methods. The methods are coded in computer programs referred to as quantum chemistry packages. There are various accuracy levels for the theoretical calculation of a chemical system, and any property of interest can usually be calculated. [7][8][9][10][11][12]

3. Types of AI and ML Algorithms Used in Chemistry

Artificial intelligence, and particularly machine learning, is a multifaceted field with several different approaches towards achieving intelligent systems, each with its own advantages and pitfalls. This section looks at some of the most important commercial or prevalent algorithms, focusing on techniques that have already been mostly developed and can be applied to a wide range of problems in chemistry.

A core concept of AI is how knowledge is represented. Rule-based (or knowledge-based) systems use explicit representations of knowledge, like “if THRESHOLD > 100 then you have a bad sensor”. These systems are perceived as more interpretable, maintainable, and transferrable between applications, but they are also more brittle, complex, and hard to debug. A reaction can be proposed with a simple rule A + B → C, but a more complex approach is needed to explain how A is made from D and E, C is made from B and C, and so on. In an unstructured environment with high uncertainty and noise, rule-based systems are easily fooled, and there would be the need for many if-then experimental rules, and what produces them is often as important as what they say.

Another point of view for rule-based systems is that they must acquire knowledge, another area in AI where chemical researchers can make significant contributions. A Mixed-Initiative Design System combines the strengths of both machines (computational power, ability to hold and search vast spaces of design alternatives) and humans (capacity to evaluate alternatives in terms of context-specific goals, prior knowledge, insight, and experience). The latter can be incorporated into rule-based systems, and as experience increases, so does their efficiency and effectiveness.

Artificial neural networks (ANNs) are “smart” systems inspired by the architecture and operation of biological neurons. Every neuron has an output that is a function (typically sigmoidal) of its weighted inputs. Large networks have been trained to recognize visual letters, classify text, speak, recognize and produce music, stock market prediction, control fighter jets, etc. There are many variants of ANNs – most widely applied is feedforward multi-layer perceptron trained with backpropagation. ANNs have been applied to many problems in chemometrics, pharmacology, biochemistry, toxicology, biophysics, spectroscopy, etc. They can be used as universal functional approximators and often outperform traditional linear methods, giving rise to the hype about them. They are notably less interpreted than rule-based systems, and there exist many divergences about ANNs (theoretical background, implementation details, use, architecture, cost-function definition, number of hidden layers and neurons, etc.), so a long phase of trial and error is often needed before any result appears. [13][14][15][16][17][18]

4. Applications of AI and ML in Chemical Structure Prediction

Artificial intelligence (AI) and machine learning (ML) have been shown to be effective in reducing the time required for reaction exploration, saving from hours to days compared to traditional screening methods. For instance, the implementation of a random forest algorithm to predict the products of a given reaction offers adequate time savings over automated screening in batch synthesizing of organic reactions.

AI and ML methods have progressively advanced for the prediction of the reaction type, such as the prediction of the product based on reactants and conditions. However, automatic reaction exploration is a different task than designing new chemical structures without any assumption of connectivity or reactivity constraints, while remaining within the scope of chemistry. Generating suitable chemicals often without immediately looking for a targeted compound is challenging in terms of the development of generative methods.

Although most efforts have focused on generative algorithms, a complementary exploration methodology, which combines a generative model and an effectiveness surrogate model (the “catcher”), and challenges arising from this approach were reviewed. Novel target structures are suggested, and machine learners allow designs to propose a first set of alternatives that are chemically meaningful and recover previously observed structures. For sequential design, adaptations to product space sampling or to the discovery of chemical path alternatives were suggested. Thus, the prediction of feasible and accessible chemical space, often including unknown compounds, is an underutilized opportunity where ML could find a key application.

Accessing high-dimensional global chemical coordinate spaces is vastly unexplored. Characterizing chemical structure topologies and properties, several AI/ML architectures, such as unity of differential graphics in three-dimensional spaces, could enable a description of both molecular and periodic solid structures alongside their properties within a common platform. The topological coordination could assist in pruning databases and patterns to minimize the search space for chemistry knowledge networks.

A critical need for international collaborative efforts on backbone databases starting from minimum connectivity, globally integrated and hierarchically (from nodes, facets, edges, and shells to polytopes) characterizing topologies across length and time scales could boost deep chemistry horizons. Novel AI architectures (e.g., qualitative agents capable of model-free formulations) could find new grounds beyond quantum-inspired theory and frameworks that would further generalize chemical realizations. [19][20][21][22][23][24]

5. AI and ML in Drug Discovery and Development

AI and ML are making significant contributions to the traditional field of medicinal chemistry. The combination of AI and ML with high-throughput computing resources and massive data has produced noteworthy progress in recent years. AI can assist in compound design decisions by simulating accurately trained ML and traditional methods. This chapter provides a survey of AI and ML applications to various aspects of drug discovery and development, including the de novo design of small molecules and peptides, binding affinity and ADME prediction, multi-criteria optimization of drug-like compounds, and the prediction of drug toxicity and building trust ML models.

The combination of AI and high-throughput computing resources, as well as massive amounts of data, has stimulated notable progress in the traditional medicinal chemistry field. AI is playing an increasingly important role in various stages of drug discovery and development, particularly in compound design decisions. Assumptions about compound ideas are evaluated by simulation with accurately trained ML and traditional methods. A survey of AI and ML applications to various aspects of drug discovery and development in chemistry has been presented. Novel small molecules and peptides are designed, predicted binding affinity, ADME, and toxicity data using ML, diversified drug-like compounds with given physiochemical features and multi-criteria optimization set, and examples are provided.

Over the past decade, AI and ML have rapidly gained popularity in medicinal chemistry. This chapter summarizes recent applications of AI and ML in drug discovery and development and discusses various challenges in promoting the use of “deep learning” AI methods in this field. AI and ML, particularly ML methods, have played important roles in drug discovery and development. Deep learning methods have some limitations in training and validation set design, but they can provide significant advantages in data-mining literature reports and working with unstructured data. Driven by the availability of large datasets and affordable high-throughput computing resources, even small academic institutions can apply AI and/or ML methodologies to achieve cost-efficient research goals. However, it should be noted that knowledge on proper management of molecules and data normalization and curation is necessary to obtain trustworthy models. This review also includes some best practices that should be considered when developing and validating ML models.

Computer-aided drug design has become a crucial component of the drug discovery and development pipeline. Structure, pharmacophore- and ligand-based virtual screening approaches have been widely employed to enhance the hit-to-lead and lead optimization processes. Fragment-based lead discovery and medicinal chemistry efforts have also profited from these access-modules to help identify new drug-like scaffolds for further development. However, traditional modeling approaches require smart yet labor-intensive parameterization processes. The most popular access-modules for the setup of docking and MM-based scoring functions, AutoDock, DOCK, and GOLD, operate using a pre-defined set of parameters extracted from quantum-mechanical computations of small model systems. These tools cannot fully account for all the subtle electrostatic and reactive aspects that play a role in protein-ligand recognition. Consequently, both docking approaches and ligand/structure-based scoring functions provide limited precision in ranking the compounds from a given library. Important hit compounds might thus be missed as a result of false negatives from docking procedures. Positive hits carefully selected through docking studies might subsequently fail in further profiling due to inaccurate predictions of selectivities or insufficient binding affinities. [25][26][27][28][29][4][30][31][32]

6. AI and ML in Materials Science and Engineering

Artificial intelligence (AI) and machine learning (ML) approaches have attracted researchers and practitioners alike because of their potential to accelerate experimental results in materials science and engineering. This applicability includes but is not limited to topics such as bioinformatics in material design, prediction of mass transport properties in porous materials, corrosion characterization, image classification for effectively determining material properties, event detection in time-series data, and Raman spectroscopy analysis. Similar to the applications with AI and ML, they can also serve as an augmentation to classical computation methods like density functional theory (DFT) in determining thermal, mechanical, and electronic properties of bulk materials and nanostructures.

In conducting computational studies using DFT, the cost burden makes it impractical to perform DFT calculations on many hundreds or thousands of candidate materials. Recently, ML interatomic potentials (MLIPs) based on local or global neural networks have been developed that can provide forces on atomic positions with good accuracy and orders of magnitude speed improvement compared to large-scale DFT calculations. Another method under this application domain is a deep-learning framework based on a neural network model for rapid construction of the band diagrams of one-dimensional photonic crystal structures. This framework is capable of designing photonic crystal structures with desirable band gaps for specific wavelengths and hence has tremendous application potential in the nanophotonics field. Each of the applications involving AI- and ML-based tools in materials science and engineering comes with underlying algorithm ideas that make the overall achievement possible.

These ideas are worthy of familiarity by researchers and practitioners in materials science and engineering who want to jump-start the understanding and implementation of AI and ML tools in tackling their scientific and engineering problems. In addition, a collection of up-to-date ML software codes/libraries for materials modeling/treatment is introduced in conjunction with the presentations of new ML algorithm ideas, so that researchers and practitioners can easily import the codes/libraries into the computational environments or platforms they are familiar with. [33][34][35][36][37][38]

7. AI and ML in Analytical Chemistry

Artificial intelligence (AI) and machine learning (ML) have impacted analytical chemistry. Experts have examined their potential applications in various areas for chemical analysis, including molecular spectroscopy, chromatographic data preprocessing, spectral data analysis, mass spectrometry, and data-driven optimization methods. In the next few years, it is anticipated that AI and ML will become more active in the analytics domain.

One of the research topics of interest is “Identification and characterization of molecules from spectral data,” covering a wide range of influence from AI and ML strategies. This includes various spectroscopic and hyphenated techniques, such as UV-VIS spectroscopy, IR spectroscopy, Raman spectroscopy, GC-MS, and LC-MS. Experts have already synthesized databases that can aid and increase the success rate of the verification and identification of molecules, particularly in drug analysis.

Examples of AI and ML applications within the area of analytical chemistry include the identification of unknown impurities by UV-VIS validation of a GD transfer function, the use of modular ML architectures, and the assessment of batch-to-batch variations in biopharmaceutical production using HPLC data utilizing ML techniques. Potential future applications in the field of “Data-Driven Methods” include automatic chromatograms, mass spectra deconvolution without theoretical knowledge, comprehensive diagnostics algorithms applied to experimental and process data, and the automatic recommendation of instructional experimental designs. [23][16][4][3][39]

8. Challenges and Limitations of AI and ML in Chemistry

Artificial intelligence (AI) and machine learning (ML) have seen increased interest in chemistry due to the field’s vast datasets and complex computational models. The interest has intensified as cloud computing resources and open-sourced datasets have become more widespread. The ongoing chemical and pharmaceutical industries’ growth spurred a need for new drugs and processes, and AI’s help has recently been called for to assist in that direction.

AI and ML have drawn a lot of attention in recent years for application in chemoinformatics, drug discovery, smart batteries, catalysis, etc. Several surveys published give an excellent overview of various AI and ML methods and applications in chemistry. There are also exciting research works about recently developed ML methods along with their applications.

On the contrary, unlike their benefits, there are also concerns regarding the limitations and challenges of AI and ML in chemistry, which are not as well discussed. In a data-driven process, AI or ML (or a combination of both) can discover patterns or predictive chemistry models from datasets with little or no physical or fundamental interpretation. The modeling is popular in chemistry because complex parameters such as chemical structure or active site of a catalyst can be quantitatively represented as simple descriptors.

The most favored type of descriptor is numerical properties calculated by quantum mechanics (QM). Combinations of descriptor properties ‘encoded’ in computationally simple mathematical formulation of model allow fast predictions of the same chemical (or those similar) for larger scale datasets. Meanwhile, the development of programming languages has made implementation of ML methods including deep learning possible and user-friendly for non-expert chemists.

However, it should be noted that even a good ML algorithm cannot model poorly-functioning processes accurately. Since AI and ML do not provide mechanistic understanding, analyzing generated chemical datasets by methods with physical bedrock (the so-called synergy) in addition to pattern recognition is needed. The sharp decline in experimental data quality is also of concern, which can adversely affect the transferability of the model and lose reliable predictions. Moreover, there is limited experimental chemoinformatics that impairs broader implementation and transferability of the generated model. [40][41][42][43][44][45]

9. Future Trends and Innovations in AI and ML for Chemistry

The integration of artificial intelligence (AI) and machine learning (ML) in chemistry continues to evolve, presenting numerous future opportunities. Certain imminent trends are expected to revolutionize chemical research and development. Increasingly available datasets will enhance ML model accuracy and output quality.

Infrastructures featuring high-throughput experimental and computational techniques will be developed to manage and curate vast datasets. Both public and commercial entities will provide these infrastructures for free or subscription-based access. Their use by businesses is expected to grow, intensifying competition and promoting ML-assisted chemical development.

Chemical data repositories will proliferate. The National Institute of Standards and Technology (NIST) and the American Chemical Society (ACS) have initiated programs to develop databases for predictive chemical structures. Such extensive databases will catalyze ML development in the chemical domain, similar to breakthroughs in natural language processing NNs with expansive text databases.

Greater availability of library compounds for high throughput screening is anticipated. Currently, pharmaceutical companies seek libraries containing novel compounds, an area where AI has begun exercising capabilities. Xu and co-workers recently developed a generative chemistry model capable of producing diverse, synthetically accessible compound libraries. Such technologies will facilitate the use of extensive virtual libraries in experimental screenings.

Another anticipated trend is enhanced collaboration across commonly separated chemical sub-fields. Homogeneity across these domains will result from the generative AI revolution, interdisciplinary analysis of chemical data, and increased awareness of similar data structures (e.g., molecular structures and energies). The CNN-based transferability of popular materials informatics models to drug discovery sub-fields highlights the potential for broad impact across chemical disciplines radioactively and predictively.

Finally, new ML model interpretation methods are expected to enable a deeper understanding of molecular structure-property relationships and the backdrop of AI-generated “rationales” for chemical synthesis predictions. Progress towards this goal has been made in the field of materials informatics using TRIP, which aligns chemical synthesis to often more understandable and interpretable solvophobicity reports. During the upcoming ML wave in chemistry, considerable efforts will likely be directed towards the development of new model interpretation methods. [46][47][48][49][50][51][52][53][54]