Keywords artificial intelligence - drug screen - molecular generation - retrosynthesis - deep
learning
Introduction
Traditional drug discovery methods are usually associated with significant challenges
including time-consuming processes, low hit rates, and a narrow focus on known targets.
In comparison to the traditional drug discovery methods, the application of artificial
intelligence (AI) presents a paradigm shift.[1 ] AI is playing a growing role in speeding up and improving drug discovery and has
been widely used in pharmaceuticals and health care including, but not limited to,
target prediction, virtual screening, molecular design, accelerating the identification
of potential drug candidate, and optimizing their pharmacological properties, revolutionizing
medical imaging and disease management,[2 ] enhancing operational efficiency, and minimizing downtime in health care services.
AI-driven generative models, such as deep learning-based approaches, which can be
used to design new molecules with desired properties, have emerged as a new paradigm
in chemical sciences.[3 ] This helps to create diverse chemical libraries for drug screening. Variational
autoencoders (VAEs) and generative adversarial networks (GANs) were two major models
commonly used in molecular generation. VAEs learn the latent space of molecular structures
and generate novel molecules by sampling from this latent space. GANs generate molecular
structures by training on a dataset of known structures. They consist of a generator
that creates molecules and a discriminator that evaluates how well the generated molecules
resemble real ones.
AI is used to design novel drug-like molecules with desired pharmacological properties.
Generative models can propose chemical structures with optimized properties such as
binding affinity, solubility, and bioavailability.[4 ] A systematic literature review found that six AI algorithms are commonly used for
de novo molecule generation, including evolutionary algorithms, adversarial autoencoders,
VAEs, GANs, long short-term memory recurrent neural networks (RNN), and gated recurrent
units.[5 ] Recently, the diffusion model has been widely used in molecule generations, where
random noises are added into three-dimensional (3D) molecule geometries, and the desired
3D geometries are constructed by learning through a reverse process.[6 ]
AI models predict the outcome of chemical reactions, recommend viable reactants, and
design chemical synthesis routes and conditions under which the desired product is
formed,[7 ] which helps to generate reaction pathways of molecular synthesis. By training on
databases of known reactions, these models can learn patterns and relationships between
reactants and products, allowing them to predict possible outcomes of new reactions.
In retrosynthetic analysis, AI suggested synthetic routes for a target molecule by
considering known reactions and associated conditions to provide viable synthesis
pathways for the desired compound.
Many aspects related to generative models for de novo drug design encompassing the categories based on molecular representations in silico ,[8 ] focusing on reinforcement learning (RL),[9 ] incorporation of protein structure,[10 ] and comparing small molecule and protein generation,[11 ] have been documented. However, the current review highlights the transformative
potential of AI across multiple facets of pharmaceuticals and health care, focusing
on two interrelated approaches: AI-driven molecule generation and synthesis planning.
We also systematically review molecular generations with one-dimensional (1D; SMILES
strings), two-dimensional (2D; graph), and 3D ligands generations.
From Drug Screening to Artificial Intelligence-Assisted Molecular Generation
From Drug Screening to Artificial Intelligence-Assisted Molecular Generation
Machine learning (ML) greatly accelerates the drug discovery process in terms of virtual
screening to predict pharmacokinetic properties, toxicity, bioavailability, cellular
localization, and screening molecular targets and bioactivity. ML algorithms fall
into two fundamental types: supervised learning and unsupervised learning. Supervised
learning uses regression analysis (e.g., decision trees, random forests, support vector
machines (SVMs), and artificial neural networks [AAN]) and classifier methods to train
ML models specifically for datasets with active and nonactive compounds. Unsupervised
algorithms classify compound datasets by identifying patterns, with examples such
as Hidden Markov Models, hierarchical clustering, and k-means clustering.[12 ] In drug screening, ANN show great promise by efficiently filtering candidate drugs
from extensive databases, simplifying the processes, and establishing relationships
between multiple targets.[13 ] The chemical spaces in the existing databases are still limited; fortunately, AI
helped molecular design by automatic generation of new drug-like molecules, offering
the promise of exploring the vast chemical space.
Molecular generation methods can be divided into structure-based and ligand-based
methods. The former uses high-precision structural features of the target protein
pocket to provide direct guidance for optimizing the interaction between the ligand
and the target, thereby driving rational compound design. The latter uses datasets
of known active ligands to design effective molecules with optimal properties, which
can be heavily influenced by training data.
Ligand-Based Molecular Generation and SMILES
Multiconstrained molecular generation (MCMG) utilizes knowledge distillation, combined
with a conditional transformer and a QSAR (quantitative structure–activity relationship)-based
RL algorithm, to satisfy multiple constraints and generate new molecules with desired
pharmacological and physicochemical properties. The process involves preconditioning
the generative model without destroying the output diversity. MCMG consists of three
essential submodels: a prior model, a distilled model based on RNN, and an agent model.
A c-transformer is trained and then distilled into RNN for subsequent application
with RL. The MCMG can effectively balance the convergence speed of the molecule generation
model and partially address the challenge of output diversity.[14 ]
de novo drug design based on the SMILES format of ligands is a convenient method because
all organic compounds can be easily represented by SMILES strings. Thus, it converts
ligand information processing into a sequence-processing procedure and allows for
learning of grammatical rules of known compounds using various models (e.g., transformer,
RNN). It is shown that in the case of BRAF inhibitor design, transformer-encoder-based
generative model trained using ChEMBL's 1.6 million data sets can be fine-tuned using
transfer learning and RL to design a new BRAF inhibitor with desirable activity.[15 ] SMILES of the ChEMBL dataset, which can also be combined with protein sequence information,
can be used to generate target-binding drugs.[16 ]
The SMILES generators are compatible with the de novo generation of dual-target ligands by using two discriminators to drive molecules
from the overlap of two bioactive compound distributions.[17 ] Even without specific 3D pocket inputs, protein–ligand interaction can incorporate
the quantitative strength of common interaction types, such as van der Waals force,
electrostatic interactions, and hydrogen bonds. Integrating this energy information
into a VAE framework minimizes SMILES reconstruction error and generates compounds
with the desired interactions.[18 ] In model training, the ligand 3D grid information of atomic physicochemical properties
can be combined with SMILES strings.[19 ]
The combination of BiLSTM (bidirectional long short-term memory) and Mol-CycleGAN
(molecular cycle generative adversarial network) methods can retain molecular input
information with a cycle architecture.[20 ] SMILES-based generative models can be generated starting from a selected core molecule
and then using Monte Carlo Tree Search and a RNN to insert the generated partial SMILES
into the initial core SMILES.[21 ]
REINVENT is a seminal molecular de novo design via deep RL.[22 ]
[23 ]
[24 ] It is interesting to note that REINVENT uses one-hot encoded SMILES as input ([Fig. 1 ]) and uses a language-based generative model RL to maximize a reward provided by
an external scoring function to optimize molecule generation.[22 ] The latest REINVENT 4 extends a number of functions including de novo design, molecule optimization, library design, R-group replacement, linker design,
and scaffold hopping.[24 ] These functions are also included in DrugHIVE, a structure-based drug design hierarchical
generative model.[25 ]
Fig. 1 Illustration of one-hot representation derived from the SMILES of 4-(chloromethyl)-1H -imidazole. Reproduced with permission from Olivecrona et al.[22 ]
Molecules generated as SMILES strings are usually accompanied by invalid molecules.
To address this problem, Krenn proposed self-referencing embedded strings (SELFIES),
where each SELFIES corresponds to a valid molecule, even for entirely random strings.[26 ] TransGEM is a molecule generation model based on a transformer with gene expression
data. Zhou et al used the SELFIES to construct a molecule generation model to incorporate
gene expression data.[27 ] The study found that high attention scores obtained from the transformer model were
associated with the onset of the disease, indicating the potential of these genes
as disease targets.[27 ]
Molecular Generation with Two-Dimensional (Graphs) Molecular Representations
Graph-based molecular generation extends the description of molecular structure regarding
realistic chemicals.[28 ]
[29 ] Fragment-based conditional molecular generation is an effective method to generate
valid molecules, which can be accomplished using 1D SMILES[21 ] or frequently graph-based models.[30 ] Initial seed used for fragment-based generation can be optimized by activity-swapping
methods that allow for the activation, deactivation, or retention of activity of molecular
seeds.[31 ] To keep more global information than random fragment search, scaffold-based deep
generative models are increasingly used, helping in considering stereochemical information
by searching scaffold and pharmacophore constraints,[32 ]
[33 ]
[34 ]
[35 ]
[36 ] or by scaffold hopping to increase diversities.[34 ]
[36 ] Graph-based models can used to generate molecules with increased drug-likeness,
decreased/increased size, and enhanced bioactivity.[37 ] Additional studies have been reported to reduce the gap between graph generative
models and target-based discovery.[38 ]
Examples of fragment-based generators include FAME,[39 ] Modof,[40 ] and NIMO.[41 ] FAME treats molecules as sequences of fragments and can be combined with gene expression
profiles.[39 ] Modof-pipe improves octanol–water partition coefficient to optimize synthetic accessibility.[40 ] NIMO uses two tailor-made motif extraction methods to map a molecular graph into
a semantic motif sequence.[41 ] Drug Design based on graph-fragment molecular representation can perform multiobjective
molecular optimization, including desired physicochemical properties and binding affinity
scores as targets.[42 ]
Many fragment-based molecular generation models use Monte Carlo tree search (MCTS)
to find an optimal attached point for extended fragment growth.[43 ]
[44 ]
[45 ]
[46 ] MCTS combines the standards of Monte Carlo strategies with tree-primarily based
search techniques that sample and explore only promising areas of the targeted area.
VGAE-MCTS is a molecular generative model that combines the variational graph autoencoder
and MCTS.[43 ] Mothra and AlphaDrug also use MCTS as a conditional molecular generation algorithm.[44 ]
[45 ]
Existing graph-based deep generative models can be easily extended to 3D representations
of molecules and target pharmacophores.[47 ] 3D-based models are more efficient in exploring chemical space in comparison to
2D methods.[48 ] The graph information embedded in relative coordinates also helps to encode the
3D structure of a molecular, thus satisfying the requirements of translation and rotation
invariance.[49 ]
Structure-Based Molecular Generation and 3D Information of Protein Binding Sites
Great advancements have been made in de novo drug design using 3D deep generative models.[50 ] Generated molecular properties and protein binding affinity often depend on the
environment in which the protein binds to the ligand. Protein binding pocket can be
represented by atomic density grids,[51 ] electrostatic environments,[52 ] or experimental electron density directly.[53 ] In the RELATION model, geometric features of the desired protein–ligand complexes
were extracted and transferred to a latent space for generation.[54 ]
An important aim of incorporating protein structure into molecule generation is to
maximize the predicted on-target binding affinity of generated molecules.[10 ]
[55 ] Specific physics-based features including the binding mechanism between a receptor
and a ligand,[56 ] or drug–target interaction, can be described in model training.[52 ]
[53 ] For example, four kinds of atomic interactions including π–π interaction, cation–π
interaction, hydrogen bond interaction, and halogen bond interaction, were tested
using learnable vector embeddings with a diffusion model.[57 ] The ligand–protein interaction can also be converted to fingerprint as constraints.[58 ]
Indeed, diffusion-based generative models have proved to be a powerful tool.[59 ] PILOT is a diffusion-based de novo ligand generation that combines pocket conditioning with large-scale pretraining
and property guidance. For a given pocket of proteins, the generated molecules have
higher binding affinity while maintaining high synthetic accessibility.[60 ] PMDM is a dual diffusion model consisting of a conditional equivariant diffusion
model with both local and global molecular dynamics.[61 ]
ResGen is a SE(3)-equivariant conditional generative model that generates 3D molecules
based on the structure of the protein pocket.[62 ] The model employs a parallel multiscale modeling strategy and a two-level autoregression
protocol, which is capable of capturing higher-level interaction between protein targets
and ligands with better computational efficiency. The molecules generated by ResGen
can bind tightly to previously unseen protein pockets of therapeutic relevance, have
potentially enhanced drug-likeness and ease of synthesized properties, and are closely
similar to the known active compounds. Notably, ResGen could be used for conformation
generation and analysis.[62 ]
Since the emergence of 3D molecular generation models, most methods have conditioned
on the target structure, thereby neglecting interaction information related to complex
molecule conceptualization and stability. In SurfGen, inspired by the simple lock–key
mode, protein surface channels are used as protein representation.[63 ] To stimulate complementarity between small molecules and protein pockets, topology
learning was subsequently performed via a Geodesic-Graph Neural Network (Geodesic-GNN).
SurfGen has the highest performance in docking and scoring compared with other methods,
e.g., GraphBP and Pocket2Mol, and can generate molecules with highly similar electron
distribution and shape to the original ligand. SurfGen's high sensitivity to pocket
structures provides an effective solution for drug resistance.
Deep learning-based molecular generation methods produce some biases related to the
ligands in the training sets, which restrict their application to data with limited
biological activity. To achieve structure-based 3D molecular design, a new network
architecture—Ligand neural network (L-Net), is used for end-to-end 3D molecular construction.[46 ] L-Net is based on a graph convolutional neural network and is trained using molecular
structures extracted from the ChEMBL database, which allows for the generation of
drug-like molecules with high-quality 3D conformations. Combining L-Net and MCTS (Monte
Carlo tree search) algorithm, DeepLigBuilder is developed to achieve de novo drug design based on target structures, which allows direct manipulation of 3D molecular
structure while optimizing the topology and 3D structure of the molecules in the binding
pocket.
Molecular design still faces many challenges.[64 ] The performances of these models may be unsatisfactory when generating a large number
of molecules with a lack of diversity.[8 ] How to synthesize strange molecules generated by AI, is an open question, and much
work has been devoted to obtaining synthesizable molecules, e.g., by selecting reactants
from commercially available compounds and constructing a synthesis route as a tree
of reaction template.[65 ] ChemistGA combines deep learning with a genetic algorithm to enhance the accessibility
and success of synthesizing molecules with desired properties.[66 ]
Artificial Intelligence Optimization of the Reaction Product and Reaction Conditions
Artificial Intelligence Optimization of the Reaction Product and Reaction Conditions
Artificial Intelligence Prediction of Reaction Outcome
Accurately predicting the outcome of organic reactions is the core of organic synthesis
in chemical drug development. Usually, this depends on the chemists' experience and
past reaction data, which is largely driven by intuition. In contrast to traditional
methods, AI algorithms can be trained with a large number of reaction precedents in
literature covering a wide range of reaction types. AI can provide possible reaction
products with a high degree of accuracy, and much faster than humans after considering
various reaction conditions and physicochemical parameters. The algorithm can be template-free,
e.g., predictions are automatically inferred from the characteristics of reactant,
reagent, and product present in the data set, i.e., looking for correlations between
the presence and absence of common chemical motifs.[67 ] In another aspect, Chen and Jung proposed a generalized-template-based GNN for accurate
prediction of organic reactivity.[68 ] The network is based on a generalized reaction template that catches organic reactivity
from the net changes in electron configuration between reactants and products.
The chemical structure figure provides a natural way to describe the structure of
molecules; nodes correspond to atoms and edges to bonds. Convolutional neural networks
use graph theory methods to understand chemical reactivity and predict reaction results
through graph editing. In 2019, Coley and colleagues used graph-convolutional neural
networks to predict the probability distribution of a mixture of reaction products.[69 ] As illustrated in [Fig. 2 ], the reactant molecules, including building blocks, catalysts, bases, ligands, and
solvents, are represented as atomic maps before encoding the data. Deep learning methods
evaluate the probability of chemical bond recombination, predict the most likely changes,
and generate a set encompassing all potential products through enumeration. Subsequently,
a new convolutional neural network reallocates the initially predicted products according
to the rules of the chemical valence state to establish a probability distribution.
According to statistical models, the molecule with the highest probability corresponds
to the primary product. The method incorporates solvent information and descriptions
of all relevant species as molecular maps for atomic mapping, which significantly
improves performance and enhances model interpretability. In more than 85% of cases,
the main reaction products can be accurately identified, with each molecule computed
in just 100 ms.
Fig. 2 Weisfeiler–Lehman network model for predicting probability distribution in reaction
product mixtures. Reproduced with permission from Coley et al.[69 ]
In ML, a random forest serves as a classifier consisting of multiple decision trees
whose output categories are determined by the mode of individual tree outputs. Its
versatility extends to handling classification, regression, and dimensionality reduction
problems. Notably, random forests exhibit robustness against outliers and noise, showcasing
superior predictive and classification performance compared with independent decision
trees. The main advantages of random forests include: (1) generating highly accurate
classifiers for various data types; (2) handling a substantial number of input variables;
and (3) assessing the importance of variables in category determination. Despite these
advantages, the application of random forest algorithms in chemical synthesis still
faces historical challenges. The complexity of implementation, particularly for nonprofessionals,
posed an obstacle. In addition, the “curse of dimensions,” where data requirement
grows exponentially with the number of dimensions studied, added to the complexity.
This challenge is particularly pronounced in the multidimensional nature of chemical
structure and reactivity, making it difficult to collect sufficient, complete, and
consistent data from databases to implement algorithms.[70 ] Fortunately, the database established through high-throughput experiments has made
it possible to predict reaction products through random forests.[71 ] In 2018, Ahneman and colleagues trained a random forest algorithm using high-throughput
datasets to predict which specific palladium catalyst is most tolerant to imidazole
during C–N bond formation.[72 ] These predictions also help guide the analysis of catalyst inhibition mechanisms.
Ross et al proposed MoLFormer, an efficient transformer encoder model for predicting
a variety of different molecular properties, which was trained on SMILES sequences
of 1.1 billion unlabeled molecules from the PubChem and ZINC datasets.[73 ] Yoshikawa et al performed CLAIRIFY for automation of experiments in a chemistry
lab using general-purpose robot manipulators and natural language commands. The large
language models (LLMs) make chemical reactions more scientific, reasonable, effective,
and practical, and provide stronger support and guarantee for the development and
application of chemistry.[74 ]
In summary, AI has significant advantages over traditional methods in predicting reaction
products, as it can predict the main products of a chemical reaction in a very short
time and with a high degree of accuracy. The disadvantage, however, is that when it
comes to predicting a reaction, a specific model needs to be established for that
reaction, and the higher the accuracy, the larger the data required to train the model.
Application of Artificial Intelligence in Predicting Reaction Conditions
While AI excels at predicting reaction outcomes, it remains a challenge to experimentally
verify computer-generated predictions, especially to determine the reaction conditions.
The reaction conditions include the chemical environment (catalyst, reagent, solvent)
and operational parameters (temperature, pressure, etc.). Different reaction conditions
often produce different results. Thus, employing AI to predict and optimize these
conditions can help improve the precision and success rate of reaction predictions.[75 ]
Gao et al discussed how ML can be used to predict fitness bars for organic reactions
([Fig. 3 ]).[76 ] The authors pointed out major limitations of existing methods, including the inability
to accurately predict complete reaction conditions, the lack of consideration of chemical
background and temperature compatibility, and the lack of large-scale reaction data
machine-readable data. To overcome these challenges, the authors developed a neural
network-based model that trained approximately 10 million reactions on Reaxys to predict
appropriate reaction conditions for organic conversion. The advantages of the model
include coverage of a wide range of organic reactions, the ability to predict factors
of reaction conditions, and the ability to quantify the similarity of reaction conditions.
However, the model has some limitations, such as a limited number of predictions and
a limited ability to predict unusual situations. The potential applications of the
model are also mentioned in the text, including route screening and prioritization
at the path level.
Fig. 3 Machine learning models can predict the conditions of organic synthesis reactions
and quantify the similarity of solvents and reagents. (Reproduced with permission
from Gao et al.[76 ] This is an unofficial adaptation of an article that appeared in an ACS publication.
ACS has not endorsed the content of this adaptation or the context of its use.)
Amar et al developed a hybrid mechanical-machine learning method for solvent selection
in process development.[77 ] They used a library of 459 solvents and calculated 12 conventional molecular descriptors,
two reaction-specific descriptors, and additional descriptors based on the screening
charge density. The method combines physically meaningful solvent descriptors with
a Gauss process-based algorithm to find solvents that are more favorable for asymmetric
hydrogenation, and better than intuitively selected solvents in terms of conversion
and enantiomer. In addition, automated ML workflow is successful for solvent selection.
However, this approach requires a large amount of data support and needs to be complemented
by proxy models with statistical predictive capabilities. Continuing to develop bridges
between chemical information and data-intensive ML methods makes a lot of sense and
promises to save time and resources for process chemists.
Rinehart et al developed an ML tool to predict substrate-adaptative conditions for
palladium-catalyzed C–N coupling reactions.[78 ] The neural network model actively learns a wide range of C–N coupling reactions
by designing an experimental data set. A challenge model using a neural network model
was used in experimental validation and successfully isolated 10 products from a series
of samples in over 85% yields. In addition, the prediction ability of the model is
gradually improving with the continuous accumulation of data ([Fig. 4 ]).
Fig. 4 Representative scope of nitrogen nucleophiles for the B–H coupling reaction and comparison
to other validated ML studies on B-H couplings. Reproduced with permission from Rinehart
et al.[78 ]
Gong et al introduced DeepReac+, a computational framework designed for predicting
chemical reactions and determining optimal reaction conditions ([Fig. 5 ]).[79 ] DeepReac+ includes the DeepReac model and sampling strategy and offers a robust
solution. The DeepReac Model is a graph-neural network-based model that specializes
in chemical reaction representation learning. It takes 2D molecular structures as
inputs, adapting seamlessly to a variety of prediction tasks, including yield and
stereoselectivity. Graph Attention Network, serving as its core, facilitates modeling
interactions among reaction components. The sampling strategy is a key element of
the DeepReac+ framework, employing two strategies: diversity sampling and adversary
sampling. These strategies select informative experimental data during model training
to improve model performance and cost efficiency. DeepReac+ efficiently predicts chemical
reaction outcomes and identifies optimal reaction conditions by combining the DeepReac
model with an active learning strategy. This integration positions DeepReac+ as a
valuable AI tool in chemical synthesis.
Fig. 5 Schematic workflow of the DeepReac+ framework. Reproduced with permission from Gong
et al.[79 ]
Shields et al presented a Bayesian reaction optimization framework along with an open-source
software tool.[80 ] This tool empowers chemists to seamlessly incorporate state-of-the-art optimization
algorithms into their everyday experiments, facilitating an enhanced and user-friendly
approach to reaction optimization. A large baseline data set of palladium-catalyzed
direct arylation reactions is collected, and Bayes optimization and human decision-making
in reaction optimization are systematically investigated. Bayesian optimization is
applied to two real optimization works (Mitsunobu reaction and defluorination reaction).
Its main advantages are higher average optimization efficiency and better consistency,
highlighting the potential of Bayesian optimization and allowing us to make better-informed,
data-driven decisions about which experiments to run, ultimately leading to more efficient
synthesis of functional chemicals.
Burger et al reported a method for autonomous experimental search using mobile robots.[81 ] A mobile robot took center stage in the quest for enhanced photocatalysts for water-based
hydrogen production. Over 8 days, the mobile robot executed a remarkable 688 experiments,
using a batch Bayesian search theory algorithm within a 10-variable experimental space.
This innovative approach greatly accelerated the exploration of improved photocatalysts,
showcasing the efficiency and potential of robotic systems in accelerating experimentation
processes. The results showed that this autonomous search identified the photocatalyst
mixture that was six times more active than the original formulation, selected the
beneficial component, and eliminated the negative component. Autonomous experiment
search using mobile robots is flexible, efficient, and safe, but the Bayesian optimization
algorithm is still somewhat blind and requires a large amount of initial investment.
Therefore, it is still a good hope that robots will replace humans in experimental
operations.
Currently, the prediction of reaction conditions (RCs) using a DL framework is hindered
by several factors, including (1) the lack of a standardized dataset for benchmarking,
(2) the lack of a general prediction model with powerful representation, and (3) the
lack of interpretability. To address these issues, we first created two standardized
RC datasets covering a broad range of reaction classes and then proposed a powerful
and interpretable Transformer-based RC predictor named Parrot.
There are several factors affecting the prediction of reaction condition,[82 ] including the lack of a general prediction model with powerful representation and
the lack of a standardized dataset for benchmarking. The lack of interpretability
is common for most ML models. Based on a self-attention mechanism, the Transformer
may boost prediction accuracy and provide interpretability, as demonstrated by interpretable
Transformer-based reaction condition predictor Parrot[82 ] and Molecular Transformer.[83 ] Meanwhile, Relational Graph Convolutional Networks may also provide accurate multilabel
classification solutions for prediction of reaction conditions.[84 ]
Application of Artificial Intelligence in Reaction Yield Prediction
With the growing abundance of molecular property datasets and reaction datasets, coupled
with advancements in computing power, the application of ML technology in reactivity
prediction has garnered significant attention.[85 ]
[86 ] Notably, Reymond and colleagues[86 ] showcased the extension of Natural Language Processing architectures, particularly
the Transformer-based bidirectional encoder representations from transformers (BERT),
for predicting reaction yields based on SMILES representations of reactants. Their
work involved fine-tuning a BERT encoder with a regression layer, pretrained using
a masked language to model the loss of chemical reactions, resulting in high-quality
yield predictions. The model was trained on two distinct datasets: one from high-throughput
experimentation (HTE) and another from patent datasets. This trained model demonstrated
its capabilities to predict a variety of reactions, including Buchwald–Hartwig and
Suzuki–Miyaura, including data from the U.S. Patent and Trademark Office (USPTO) dataset.
It is worth noting that the HTE and USPTO datasets differ significantly in content
and quality, with the former covering a specific chemical reaction region and providing
high-quality data, while the latter spans a broader reaction space with noisy and
sparse data.[86 ] Additional studies have highlighted the limitations of this dataset's suitability
for reaction yield prediction. Saebi et al found that using data from electronic laboratory
notebooks (ELNs) to train attributed GNNs does not lead to a predictive model, contrary
to the initial expectation that the ELNs could provide less biased, large datasets.[87 ]
Yield prediction methods vary using techniques such as one-hot encoding of reactants,
tandem molecular fingerprints, or computational chemical descriptors. Probst et al
proposed a differential reaction fingerprint (DRFP) for reaction searching and categorization
as well as yield prediction.[88 ] The DRFP algorithm takes a reaction SMILES as an input and creates a binary fingerprint
based on the symmetric difference of two sets containing circular n-grams.
Glorius'group introduced a structure-based ML platform with diverse applications in
organic chemistry,[89 ] to achieve generality in molecular representation, they developed an input based
on a multifingerprint feature. This approach applies to a variety of problem sets.
Initially, it was able to accurately predict the molecular properties of diverse molecular
arrays. Then, the platform successfully predicted reaction outcomes, including stereoselectivity
and yield, for previously evaluated experimental datasets using problem-specific descriptor
models. In a final application, the platform showed effective correlations when applied
to the systematic analysis of a high-throughput dataset, showcasing its practical
utility in structure-based modalities.
Reymond's group[90 ] used a natural language processing architecture to predict response properties based
on a text-based response representation. Using an encoder-transformer model paired
with a regression layer, they achieved excellent predictive performance on two high-throughput
experimental reaction sets. However, when analyzing yields from the USPTO dataset,
they observed differences in distribution based on mass scale. To obtain a high-quality
generic reaction yield dataset, Yin et al curated a generic reaction yield dataset
containing information on 12 reaction categories and reaction conditions.[91 ] Subsequently, using BERT-based reaction yield predictor, they found that contrastive
learning based on reaction conditions enhances the sensitivity of the model to reaction
conditions.
ML models using quantum chemical calculations were trained to predict the transition
state and yield in copper-catalyzed P–H insertion reactions.[92 ] The transition state was identified by analyzing 120 experimental data points using
density functional theory. Subsequently, an ML algorithm was applied to analyze the
16 descriptors derived from the transition states to predict product yields. Among
the algorithms investigated, SVM had the highest prediction accuracy of 97%, with
a correlation of over 80% in leave-one-out cross-validation. Sensitivity analysis
was performed for each descriptor and the reaction mechanism was thoroughly examined
to enhance the understanding of transition state characteristics. Matsubara's group
used ML methods to build a multiple linear regression model based on batch reaction
data of 29 substrates to predict the Wittig methylene reaction rate diagram of any
aldehyde and diiodomethane.[93 ] The predicted profile allows the simultaneous determination of the highest achievable
yield and the shortest reaction time. This can be interpreted as the residence time
required to reach the maximum yield of the methylation of diiodomethane in a flow
microreactor.
Conclusion
AI offers new opportunities for the design of innovative chemical drugs, and it has
changed the traditional research paradigm of medicinal chemistry by designing and
generating small molecules in a more efficient, smarter, and more precise way, and
increasing their potential to become drugs.
For a given chemical reaction, AI has been developed to predict reaction products,
and reaction yields, and to optimize reaction conditions. However, accurate prediction
of reaction products is often dependent on the amount of data required to train the
model. For the optimization of reaction conditions, there are difficulties such as
the inability to accurately predict complete reaction conditions, the lack of consideration
of chemical background and temperature compatibility, and the lack of machine-readable
data for large-scale reaction data. The neural network-based model developed by the
researchers, as well as the use of mobile robots for autonomous experimental searches,
opens up additional possibilities for more comprehensively determining optimal chemical
reaction conditions. Feature learning methods such as language model (LM) and GNN
show good promise in chemical reaction yield prediction. On this basis, researchers
have proposed structure-based ML platforms or means incorporating quantum chemical
computing to accurately predict the highest yields and shortest reaction times of
chemical reactions.
In small-molecule drug design, AI techniques are utilized to generate molecules with
potential biological activity. Structure- and ligand-based molecular generation models
offer the possibility of efficient molecular discovery. However, many challenges remain
regarding the structural diversity of the generated molecules and the ability of current
molecular generation models to generate large molecules. In terms of drug screening,
AI realizes the validation of drug targets and the optimal design of drug structures
faster than conventional drug screening techniques based on traditional multidisciplinary.
Different algorithms, as well as predictive models, have been used to evaluate the
physicochemical properties as well as in vivo activity and toxicity of small molecule
drugs. These technological innovations can significantly reduce the time required
for new drug discovery.
