In the development of targeted drugs, it is very important to predict the Interaction between the Drug and the target. This process can be described as the prediction of drug-target interaction (DTI). The process demonstration is shown in Figure 10^[73]. At present, in the task of affinity prediction, neural network models dominated by deep learning can usually do well. Because neural networks can usually deal with more complex drug molecular data, through layer-by-layer learning and transmission, the model can automatically learn and extract relevant data features, thereby improving the accuracy of the model in predicting activity.

Based on the drug target affinity prediction task of deep learning, Lee et al. Proposed a DTI prediction model based on deep learning to capture the local residual patterns of proteins involved in DTI^[61]. When they used a convolutional neural network (CNN) on the raw protein sequence, amino acid subsequences of different lengths were convolved to capture local residual patterns of the generalized protein class. The long short-term memory network (LSTM), a variant of RNN, is also widely used for DTI prediction. Wang et al. Developed a DTIs prediction model based on LSTM architecture^[64]. Evolutionary features of proteins were extracted by position-specific scoring matrix (PSSM) and Legendre moments (LM) and correlated with drug molecular substructure fingerprints to form feature vectors of drug-target pairs. Sparse Principal Component Analysis (SPCA) is then used to compress the features of drugs and proteins into a uniform vector space. The experimental results show that the prediction performance of DTIs is significantly improved, and the AUCs on four important drug target datasets are 0. 9951, 0. 9705, 0. 9951 and 0. 9206, respectively. Yuan et al. proposed a FusionDTA method based on the LSTM architecture^[74]. For the loss of implicit information, a novel multi-head linear attention mechanism is used to replace the coarse pooling method. This allows FusionDTA to aggregate global information based on the attention weights instead of selecting the largest one like max pooling. The experimental results show that the consistency index (CI) of the model on Davis and KIBA datasets is 0.913 and 0.906, respectively. Shao et al. Proposed a method based on the GNN model^[75]. This is an end-to-end model based on the Attention Mechanism Heterogeneous Graph (DTI-HETA). In this model, the heterogeneity graph is first constructed based on the drug-drug, target-target similarity matrix and DTI matrix. Then, the graph convolutional neural network is used to get the embedding representation of drug and target. In order to highlight the contribution of different neighborhood nodes to the central node when aggregating the graph convolution information, the graph attention mechanism is introduced in the node embedding process. Finally, the inner product decoder is used to predict DTIs. Li et al. Proposed a network architecture MINN-DTI based on an improved Transformer^[76]. MINN-DTI combines the interaction converter module (Interformer) with a modified communication message passing neural network (CMPNN), namely Inter-CMPNN, to better capture the bidirectional influence between drug and target, which is represented by molecular graph and distance graph, respectively. This method provides a good interpretation of the results, and the interpretability of the results is shown in Figure 11.

Crystal form is an important property of drugs, which significantly affects the physical and chemical properties of drugs, such as drug stability, solubility and dissolution rate, and then affects the storage, formulation, product design and delivery mechanism of drugs to patients^[77]. At the same time, crystal morphology also affects the downstream processing of drugs, such as processes that require specific particles (morphology to maximize filtration, particle flowability, and tabletability)^[78,79]. While current crystal engineering has been focused on targeting specific interactions by solvent selection or controlling experimental conditions to produce ideal crystal morphology^[80]. The success of the experiment depends on the experience of scientists and frequent trial and error, which requires a lot of manpower, time and resources, so the development of AI technology has brought new opportunities for crystal form prediction. AI Pharmaceutical Company Jingtai Technology and AstraZeneca jointly released the API AZD1305 experiment for cardiovascular diseases, and found two crystal forms with very close physical stability at room temperature. With the help of Jingtai Technology's crystal form prediction and stability evaluation technology, the polymorphic phenomenon was systematically evaluated and the experimental results were verified. The results are shown in Figure 12^[81].

In addition, Bhardwaj et al. Used the random forest model to predict the crystallinity of organic molecules for the first time, with an accuracy of about 70%^[82]. The predictive model is based on calculated molecular descriptors and published crystallization propensity experiments for a library of substituted acylanilides. Wicker et al. Used Python and the RDKit cheminformatics toolkit to build models on the basis of chemical descriptors and unsupervised machine learning methods^[83]. They predicted whether molecules would crystallize without considering crystal growth mechanisms or conditions, a method that focuses on the properties and interactions of individual molecules. By comparing the SVM and RF models, they found that the SVM model had the highest prediction accuracy of 90.3%. On this basis, Pillong et al. Established the RF model to evaluate the solubility and crystallization tendency of 319 small molecules in 18 different solvents^[84]. The model can guide the selection of appropriate crystallization solvent, and effectively reduce the workload to 1/3 of the initial plan while ensuring the success rate of crystallization to exceed 92%. In the latest research, Yang et al. Developed an efficient molecular mechanics protocol based on the Einstein crystal method^[85]. Using this approach, they performed finite temperature free energy corrections to the predicted crystal structure and the experimentally known five polymorphisms of molecule XXIII, a pharmaceutically relevant compound. Wilkinson et al. Proposed a method based on transfer learning and open source robotics, which improved the level of AI in predicting crystal morphology^[86]. Experiments show that the method can predict the crystal morphology with an accuracy of 87. 9% by using the data-driven model. This approach will shorten the time to drug development and create more sustainable and efficient manufacturing methods.

Molecular property prediction is a fundamental but challenging task in drug discovery, which has gained increasing attention in recent decades with the development of machine learning and deep learning^[87]. As shown in Table 1, the properties of compounds can be divided into three categories: physicochemical properties, biophysical properties, and physiological properties, and each category corresponds to several benchmark test data sets. In these test data sets, AI techniques usually divide the prediction task into classification (AUC-ROC criterion) and regression (RMSE criterion) to test the performance of the model in molecular property prediction according to two sets of values^[88].

Based on SMILES sequences, Jiang et al. Proposed the NoiseMol method, which for the first time systematically proposed to inject perturbation noise into SMILES strings labeled at the atomic level or substring level to expand the size of the data set and alleviate the limitation of labeled data on the prediction of molecular properties^[89]. From the experimental results, this method enhances the prediction ability and robustness of Transformer, and the results on small data sets such as BACE, BBBP, FDA and ecoli are improved more significantly. Li et al. Proposed the Multiple SMILES method, which shows excellent performance in molecular property prediction tasks by encoding Multiple SMILES for each molecule as automatic data augmentation, and alleviates the overfitting problem caused by the small amount of data in the molecular property prediction data set^[90].

Based on the graph structure, Maziarka et al. Proposed a MAT method to make the Transformer model more suitable for chemical molecules by increasing the distance between atoms and the self-attention of the molecular graph^[91]. Besides that, the good results of downstream task prediction and the interpretability of attention weights are also the highlights of their work. Liu et al. Proposed a self-supervised learning method ATMOL for molecular representation learning and property prediction and a new molecular graph enhancement strategy, called attention intelligent graph masking, to generate challenging positive samples for contrastive learning^[92]. Hu et al. Proposed an ABT-MPNN method based on a variant of graph convolutional neural network (GCN) to improve the molecular characterization embedding process for molecular property prediction^[93]. Their approach provides a novel architecture that integrates molecular representations at the bond, atom, and molecule levels in an end-to-end manner. Experimental results on nine benchmark datasets show that the proposed ABT-MPNN outperforms or is comparable to state-of-the-art baseline models in quantitative structure-attribute relationship tasks.

Molecular conformation generation is an important task in drug discovery, where the goal is to generate a spatial arrangement of atoms in a molecule at low energy. The traditional way of molecular conformation generation is to first generate a stack of molecular conformations using conformational search methods, and after energy minimization of each conformation, a batch of low-energy molecular conformations is screened out according to the potential energy evaluated by density functional theory (DFT), semi-empirical DFT, or molecular mechanism (MM)^[94]. In recent years, with the development of deep learning of AI technology, data-driven molecular conformation generation has become possible, and its basic process is shown in Figure 13.

Mansimov et al. Used VAE to directly generate atomic coordinates^[96]. However, this model does not preserve translational and rotational equivariance of the generated conformations and performs poorly in benchmarking. To guarantee this geometric property of the conformation, later studies used intermediate structures such as interatomic distances or torsion angles to generate the conformation. For example, Xu et al. Used the flow model to generate the distance matrix, and then used the classical distance geometry technique to iteratively generate the conformation^[97]. After that, the team designed an end-to-end framework to enhance the performance of the model by using two-tier optimization^[98]. Zhu et al. Proposed a method to directly predict the atomic coordinates without predicting the interatomic distance, the gradient of the interatomic distance, or the local structure of the molecule in advance, and achieved good results on two types of data sets^[99]. Recently, Xu et al. Proposed a GeoDiff model, which is a geometric diffusion model that can effectively train the entire framework in an end-to-end manner by optimizing the weighted variational lower bound of the (conditional) likelihood for molecular conformation generation^[100].

Molecular generation is a challenging open problem in cheminformatics. Presumably, the total number of potential drug-like candidate molecules is between 10²³~10⁶⁰ molecules, of which only 10⁸ molecules have been synthesized^[101,102]. Since it is difficult to sift the practically infinite chemical space and there is a huge difference between synthesized molecules and potential molecules, modeling the distribution of molecules with a generative model in order to sample molecules with desirable properties is the main purpose of this task, and the process of molecule generation is shown in Figure 14.

Influenced by natural language processing (NLP), based on SMILES sequences. Segler et al. trained a long short-term memory RNN model based on a set of molecules in ChEMBL, represented by normalized SMILES, which can generate diverse and reasonably regular molecules^[103]. Yang et al. Used an LSTM-based neural network model to train 200,000 compounds in the ChEMBL database, and then fine-tuned the model using a dataset containing 135 published P300 inhibitors and 576 macrocyclic molecules to generate new P300/CBP inhibitors^[104]. As a result, the model generated a centralized library of 672 chemical structures, and a series of highly effective inhibitors were obtained through synthetic optimization. Bagal et al. proposed the LigGPT model for molecule generation, which can learn long-term dependencies in SMILES, such as loop closure^[105]. Since only the attention in front of the token needs to be utilized when predicting the next token in the sequence, a masking self-attention mechanism is applied in the model. Later, the team proposed a MolGPT model that performed as well as other previously proposed modern machine learning frameworks for molecule generation in generating effective, unique, and novel molecules^[106]. Furthermore, they demonstrate that the model can be conditionally trained to control multiple properties of the generative molecule. The model can be used to generate molecules with desired scaffolds and desired properties by tuning the SMILES strings and attribute values of the desired scaffolds. Wang et al. Proposed a Transformer model for target-specific molecular design^[107]. Their proposed method enables the generation of drug-like compounds (without assigned targets) and target-specific compounds, where target-specific compounds are generated by applying different multi-head attention keys and values to each target. The experimental results show that the method can not only generate effective drug-like compounds, but also generate target-specific compounds.

In addition, with the development of GNN, molecular generation model based on molecular graph has become a hot research topic. You et al. Proposed a graph convolution policy network (GCPN) model to generate molecular graphs^[108]. The model combines graph representation, reinforcement learning, and adversarial training in a unified framework, which enables the generation of effective molecular graphs. In addition, the method can directly optimize the properties of the molecular graph to generate target oriented molecules. Samanta et al. proposed a graph-based VAE model^[109]. In this model, the encoder is used to aggregate information and then map this aggregated information into a continuous latent space, which can encode graphs with variable number of atoms; The decoder collectively represents all edges as a non-normalized log-probability vector, which is then fed as a single edge distribution, enabling efficient inference algorithms and decoding. The results show that the model can efficiently generate new molecules. Li et al. Proposed a multi-physical graph neural network (MP-GNN) model based on the developed multi-physical graph molecular graph representation and features^[110]. The various molecular interactions at different atomic types and different scales are systematically represented by a series of scale-specific and element-specific graphs featuring distance-dependent nodes. From these graphs, a graph convolutional network (GCN) model is built with a specially designed weight-sharing architecture. A base learner is constructed from GCN models of different elements at different scales and further integrated together using single-scale and multi-scale ensemble learning schemes. The experimental results show that the model has high accuracy.