Unsupervised and supervised AI on molecular dynamics simulations reveals complex characteristics of HLA-A2-peptide immunogenicity.

Immunologic recognition of peptide antigens bound to class I major histocompatibility complex (MHC) molecules is essential to both novel immunotherapeutic development and human health at large. Current methods for predicting antigen peptide immunogenicity rely primarily on simple sequence representations, which allow for some understanding of immunogenic features but provide inadequate consideration of the full scale of molecular mechanisms tied to peptide recognition. We here characterize contributions that unsupervised and supervised artificial intelligence (AI) methods can make toward understanding and predicting MHC(HLA-A2)-peptide complex immunogenicity when applied to large ensembles of molecular dynamics simulations. We first show that an unsupervised AI method allows us to identify subtle features that drive immunogenicity differences between a cancer neoantigen and its wild-type peptide counterpart. Next, we demonstrate that a supervised AI method for class I MHC(HLA-A2)-peptide complex classification significantly outperforms a sequence model on small datasets corrected for trivial sequence correlations. Furthermore, we show that both unsupervised and supervised approaches reveal determinants of immunogenicity based on time-dependent molecular fluctuations and anchor position dynamics outside the MHC binding groove. We discuss implications of these structural and dynamic immunogenicity correlates for the induction of T cell responses and therapeutic T cell receptor design.

Structural cavities are critical to balancing stability and activity of a membrane-integral enzyme.

Packing interaction is a critical driving force in the folding of helical membrane proteins. Despite the importance, packing defects (i.e., cavities including voids, pockets, and pores) are prevalent in membrane-integral enzymes, channels, transporters, and receptors, playing essential roles in function. Then, a question arises regarding how the two competing requirements, packing for stability vs. cavities for function, are reconciled in membrane protein structures. Here, using the intramembrane protease GlpG of Escherichiacoli as a model and cavity-filling mutation as a probe, we tested the impacts of native cavities on the thermodynamic stability and function of a membrane protein. We find several stabilizing mutations which induce substantial activity reduction without distorting the active site. Notably, these mutations are all mapped onto the regions of conformational flexibility and functional importance, indicating that the cavities facilitate functional movement of GlpG while compromising the stability. Experiment and molecular dynamics simulation suggest that the stabilization is induced by the coupling between enhanced protein packing and weakly unfavorable lipid desolvation, or solely by favorable lipid solvation on the cavities. Our result suggests that, stabilized by the relatively weak interactions with lipids, cavities are accommodated in membrane proteins without severe energetic cost, which, in turn, serve as a platform to fine-tune the balance between stability and flexibility for optimal activity.

Analysis of Training and Seed Bias in Small Molecules Generated with a Conditional Graph-Based Variational Autoencoder─Insights for Practical AI-Driven Molecule Generation.

The application of deep learning to generative molecule design has shown early promise for accelerating lead series development. However, questions remain concerning how factors like training, data set, and seed bias impact the technology's utility to medicinal and computational chemists. In this work, we analyze the impact of seed and training bias on the output of an activity-conditioned graph-based variational autoencoder (VAE). Leveraging a massive, labeled data set corresponding to the dopamine D2 receptor, our graph-based generative model is shown to excel in producing desired conditioned activities and favorable unconditioned physical properties in generated molecules. We implement an activity-swapping method that allows for the activation, deactivation, or retention of activity of molecular seeds, and we apply independent deep learning classifiers to verify the generative results. Overall, we uncover relationships between noise, molecular seeds, and training set selection across a range of latent-space sampling procedures, providing important insights for practical AI-driven molecule generation.

Simplified, interpretable graph convolutional neural networks for small molecule activity prediction.

We here present a streamlined, explainable graph convolutional neural network (gCNN) architecture for small molecule activity prediction. We first conduct a hyperparameter optimization across nearly 800 protein targets that produces a simplified gCNN QSAR architecture, and we observe that such a model can yield performance improvements over both standard gCNN and RF methods on difficult-to-classify test sets. Additionally, we discuss how reductions in convolutional layer dimensions potentially speak to the "anatomical" needs of gCNNs with respect to radial coarse graining of molecular substructure. We augment this simplified architecture with saliency map technology that highlights molecular substructures relevant to activity, and we perform saliency analysis on nearly 100 data-rich protein targets. We show that resultant substructural clusters are useful visualization tools for understanding substructure-activity relationships. We go on to highlight connections between our models' saliency predictions and observations made in the medicinal chemistry literature, focusing on four case studies of past lead finding and lead optimization campaigns.

CASTELO: clustered atom subtypes aided lead optimization-a combined machine learning and molecular modeling method.

BACKGROUND: Drug discovery is a multi-stage process that comprises two costly major steps: pre-clinical research and clinical trials. Among its stages, lead optimization easily consumes more than half of the pre-clinical budget. We propose a combined machine learning and molecular modeling approach that partially automates lead optimization workflow in silico, providing suggestions for modification hot spots. RESULTS: The initial data collection is achieved with physics-based molecular dynamics simulation. Contact matrices are calculated as the preliminary features extracted from the simulations. To take advantage of the temporal information from the simulations, we enhanced contact matrices data with temporal dynamism representation, which are then modeled with unsupervised convolutional variational autoencoder (CVAE). Finally, conventional and CVAE-based clustering methods are compared with metrics to rank the submolecular structures and propose potential candidates for lead optimization. CONCLUSION: With no need for extensive structure-activity data, our method provides new hints for drug modification hotspots which can be used to improve drug potency and reduce the lead optimization time. It can potentially become a valuable tool for medicinal chemists.

T cell receptors for the HIV KK10 epitope from patients with differential immunologic control are functionally indistinguishable.

HIV controllers (HCs) are individuals who can naturally control HIV infection, partially due to potent HIV-specific CD8+ T cell responses. Here, we examined the hypothesis that superior function of CD8+ T cells from HCs is encoded by their T cell receptors (TCRs). We compared the functional properties of immunodominant HIV-specific TCRs obtained from HLA-B*2705 HCs and chronic progressors (CPs) following expression in primary T cells. T cells transduced with TCRs from HCs and CPs showed equivalent induction of epitope-specific cytotoxicity, cytokine secretion, and antigen-binding properties. Transduced T cells comparably, albeit modestly, also suppressed HIV infection in vitro and in humanized mice. We also performed extensive molecular dynamics simulations that provided a structural basis for similarities in cytotoxicity and epitope cross-reactivity. These results demonstrate that the differential abilities of HIV-specific CD8+ T cells from HCs and CPs are not genetically encoded in the TCRs alone and must depend on additional factors.

Rare Dissipative Transitions Punctuate the Initiation of Chemical Denaturation in Proteins.

Protein unfolding dynamics are bound by their degree of entropy production, a quantity that relates the amount of heat dissipated by a nonequilibrium process to a system's forward and time-reversed trajectories. We here explore the statistics of heat dissipation that emerge in protein molecules subjected to a chemical denaturant. Coupling large molecular dynamics datasets and Markov state models with the theory of entropy production, we demonstrate that dissipative processes can be rigorously characterized over the course of the urea-induced unfolding of the protein chymotrypsin inhibitor 2. By enumerating full entropy production probability distributions as a function of time, we first illustrate that distinct passive and dissipative regimes are present in the denaturation dynamics. Within the dissipative dynamical region, we next find that chymotrypsin inhibitor 2 is strongly driven into unfolded states in which the protein's hydrophobic core has been penetrated by urea molecules and disintegrated. Detailed analyses reveal that urea's interruption of key hydrophobic contacts between core residues causes many of the protein's native structural features to dissolve.

How force unfolding differs from chemical denaturation.

Single-molecule force spectroscopies are remarkable tools for studying protein folding and unfolding, but force unfolding explores protein configurations that are potentially very different from the ones traditionally explored in chemical or thermal denaturation. Understanding these differences is crucial because such configurations serve as starting points of folding studies, and thus can affect both the folding mechanism and the kinetics. Here we provide a detailed comparison of both chemically induced and force-induced unfolded state ensembles of ubiquitin based on extensive, all-atom simulations of the protein either extended by force or denatured by urea. As expected, the respective unfolded states are very different on a macromolecular scale, being fully extended under force with no contacts and partially extended in urea with many nonnative contacts. The amount of residual secondary structure also differs: A significant population of α-helices is found in chemically denatured configurations but such helices are absent under force, except at the lowest applied force of 30 pN where short helices form transiently. We see that typical-size helices are unstable above this force, and ß-sheets cannot form. More surprisingly, we observe striking differences in the backbone dihedral angle distributions for the protein unfolded under force and the one unfolded by denaturant. A simple model based on the dialanine peptide is shown to not only provide an explanation for these striking differences but also illustrates how the force dependence of the protein dihedral angle distributions give rise to the worm-like chain behavior of the chain upon force.

Toward high permeability, selectivity and controllability of water desalination with FePc nanopores.

Nanoporous materials exhibit promising potential in water transportation applications, especially in ocean water desalination. It is highly desired to have great permeability, selectivity and controllability in the desalination performance of these nanopores. However, it is still a challenge to achieve all three features in one material or device. Here, we demonstrate efficient and controllable water desalination with a nanoporous 2D Fe phthalocyanine (FePc) membrane using molecular dynamics simulations. We find the FePc membrane not only conducts fast water flow, but it also suppresses ion permeation. The selectivity is attributed to a mechanism distinct from the traditional steric exclusion: cations are excluded due to electrostatic repulsion, whereas anions can be trapped in the nanopore and induce the reorganization of ions in the vicinity of the nanopore, which in turn creates a tendency for the trapped anions to move back into the saline reservoir. More interestingly, we find such mechanism is largely due to the sufficiently strong electrostatic interaction of the charged nanopore region with ions and is not restricted to the FePc nanopore. In addition, the number of protonated nitrogen atoms in FePc pores can be modulated by adjusting the pH value of the solution. The extent of the anion occupancy can thus be regulated, giving rise to control of the water flow. Taken together, great permeability, selectivity and controllability can be achieved with this nanosheet system. Moreover, our study suggests there is an alternative mechanism of water desalination which may be realized by intrinsically nanoporous materials such as FePc membranes.

Salts drive controllable multilayered upright assembly of amyloid-like peptides at mica/water interface.

Surface-assisted self-assembly of amyloid-like peptides has received considerable interest in both amyloidosis research and nanotechnology in recent years. Despite extensive studies, some controlling factors, such as salts, are still not well understood, even though it is known that some salts can promote peptide self-assemblies through the so-called "salting-out" effect. However, they are usually noncontrollable, disordered, amorphous aggregates. Here, we show via a combined experimental and theoretical approach that a conserved consensus peptide NH2-VGGAVVAGV-CONH2 (GAV-9) (from representative amyloidogenic proteins) can self-assemble into highly ordered, multilayered nanofilaments, with surprising all-upright conformations, under high-salt concentrations. Our atomic force microscopy images also demonstrate that the vertical stacking of multiple layers is highly controllable by tuning the ionic strength, such as from 0 mM (monolayer) to 100 mM (mainly double layer), and to 250 mM MgCl2 (double, triple, quadruple, and quintuple layers). Our atomistic molecular dynamics simulations then reveal that these individual layers have very different internal nanostructures, with parallel ß-sheets in the first monolayer but antiparallel ß-sheets in the subsequent upper layers due to their different microenvironment. Further studies show that the growth of multilayered, all-upright nanostructures is a common phenomenon for GAV-9 at the mica/water interface, under a variety of salt types and a wide range of salt concentrations.

Selection of an HLA-C*03:04-Restricted HIV-1 p24 Gag Sequence Variant Is Associated with Viral Escape from KIR2DL3+ Natural Killer Cells: Data from an Observational Cohort in South Africa.

BACKGROUND: Viruses can evade immune surveillance, but the underlying mechanisms are insufficiently understood. Here, we sought to understand the mechanisms by which natural killer (NK) cells recognize HIV-1-infected cells and how this virus can evade NK-cell-mediated immune pressure. METHODS AND FINDINGS: Two sequence mutations in p24 Gag associated with the presence of specific KIR/HLA combined genotypes were identified in HIV-1 clade C viruses from a large cohort of infected, untreated individuals in South Africa (n = 392), suggesting viral escape from KIR+ NK cells through sequence variations within HLA class I-presented epitopes. One sequence polymorphism at position 303 of p24 Gag (TGag303V), selected for in infected individuals with both KIR2DL3 and HLA-C*03:04, enabled significantly better binding of the inhibitory KIR2DL3 receptor to HLA-C*03:04-expressing cells presenting this variant epitope compared to the wild-type epitope (wild-type mean 18.01 ± 10.45 standard deviation [SD] and variant mean 44.67 ± 14.42 SD, p = 0.002). Furthermore, activation of primary KIR2DL3+ NK cells from healthy donors in response to HLA-C*03:04+ target cells presenting the variant epitope was significantly reduced in comparison to cells presenting the wild-type sequence (wild-type mean 0.78 ± 0.07 standard error of the mean [SEM] and variant mean 0.63 ± 0.07 SEM, p = 0.012). Structural modeling and surface plasmon resonance of KIR/peptide/HLA interactions in the context of the different viral sequence variants studied supported these results. Future studies will be needed to assess processing and antigen presentation of the investigated HIV-1 epitope in natural infection, and the consequences for viral control. CONCLUSIONS: These data provide novel insights into how viruses can evade NK cell immunity through the selection of mutations in HLA-presented epitopes that enhance binding to inhibitory NK cell receptors. Better understanding of the mechanisms by which HIV-1 evades NK-cell-mediated immune pressure and the functional validation of a structural modeling approach will facilitate the development of novel targeted immune interventions to harness the antiviral activities of NK cells.

Molecular mechanism of pancreatic tumor metastasis inhibition by Gd@C82(OH)22 and its implication for de novo design of nanomedicine.

Pancreatic adenocarcinoma is the most lethal of the solid tumors and the fourth-leading cause of cancer-related death in North America. Matrix metalloproteinases (MMPs) have long been targeted as a potential anticancer therapy because of their seminal role in angiogenesis and extracellular matrix (ECM) degradation of tumor survival and invasion. However, the inhibition specificity to MMPs and the molecular-level understanding of the inhibition mechanism remain largely unresolved. Here, we found that endohedral metallofullerenol Gd@C(82)(OH)(22) can successfully inhibit the neoplastic activity with experiments at animal, tissue, and cellular levels. Gd@C(82)(OH)(22) effectively blocks tumor growth in human pancreatic cancer xenografts in a nude mouse model. Enzyme activity assays also show Gd@C(82)(OH)(22) not only suppresses the expression of MMPs but also significantly reduces their activities. We then applied large-scale molecular-dynamics simulations to illustrate the molecular mechanism by studying the Gd@C(82)(OH)(22)-MMP-9 interactions in atomic detail. Our data demonstrated that Gd@C(82)(OH)(22) inhibits MMP-9 mainly via an exocite interaction, whereas the well-known zinc catalytic site only plays a minimal role. Steered by nonspecific electrostatic, hydrophobic, and specific hydrogen-bonding interactions, Gd@C(82)(OH)(22) exhibits specific binding modes near the ligand-specificity loop S1', thereby inhibiting MMP-9 activity. Both the suppression of MMP expression and specific binding mode make Gd@C(82)(OH)(22) a potentially more effective nanomedicine for pancreatic cancer than traditional medicines, which usually target the proteolytic sites directly but fail in selective inhibition. Our findings provide insights for de novo design of nanomedicines for fatal diseases such as pancreatic cancer.

Computational design of a protein crystal.

Protein crystals have catalytic and materials applications and are central to efforts in structural biology and therapeutic development. Designing predetermined crystal structures can be subtle given the complexity of proteins and the noncovalent interactions that govern crystallization. De novo protein design provides an approach to engineer highly complex nanoscale molecular structures, and often the positions of atoms can be programmed with sub-Å precision. Herein, a computational approach is presented for the design of proteins that self-assemble in three dimensions to yield macroscopic crystals. A three-helix coiled-coil protein is designed de novo to form a polar, layered, three-dimensional crystal having the P6 space group, which has a "honeycomb-like" structure and hexameric channels that span the crystal. The approach involves: (i) creating an ensemble of crystalline structures consistent with the targeted symmetry; (ii) characterizing this ensemble to identify "designable" structures from minima in the sequence-structure energy landscape and designing sequences for these structures; (iii) experimentally characterizing candidate proteins. A 2.1 Å resolution X-ray crystal structure of one such designed protein exhibits sub-Å agreement [backbone root mean square deviation (rmsd)] with the computational model of the crystal. This approach to crystal design has potential applications to the de novo design of nanostructured materials and to the modification of natural proteins to facilitate X-ray crystallographic analysis.

Characterization of a novel water pocket inside the human Cx26 hemichannel structure.

Connexins (Cxs) are a family of vertebrate proteins constituents of gap junction channels (GJCs) that connect the cytoplasm of adjacent cells by the end-to-end docking of two Cx hemichannels. The intercellular transfer through GJCs occurs by passive diffusion allowing the exchange of water, ions, and small molecules. Despite the broad interest to understand, at the molecular level, the functional state of Cx-based channels, there are still many unanswered questions regarding structure-function relationships, perm-selectivity, and gating mechanisms. In particular, the ordering, structure, and dynamics of water inside Cx GJCs and hemichannels remains largely unexplored. In this work, we describe the identification and characterization of a believed novel water pocket-termed the IC pocket-located in-between the four transmembrane helices of each human Cx26 (hCx26) monomer at the intracellular (IC) side. Using molecular dynamics (MD) simulations to characterize hCx26 internal water structure and dynamics, six IC pockets were identified per hemichannel. A detailed characterization of the dynamics and ordering of water including conformational variability of residues forming the IC pockets, together with multiple sequence alignments, allowed us to propose a functional role for this cavity. An in vitro assessment of tracer uptake suggests that the IC pocket residue Arg-143 plays an essential role on the modulation of the hCx26 hemichannel permeability.

Hydrophobic interaction drives surface-assisted epitaxial assembly of amyloid-like peptides.

The molecular mechanism of epitaxial fibril formation has been investigated for GAV-9 (NH(3)(+)-VGGAVVAGV-CONH(2)), an amyloid-like peptide extracted from a consensus sequence of amyloidogenic proteins, which assembles with very different morphologies, "upright" on mica and "flat" on the highly oriented pyrolytic graphite (HOPG). Our all-atom molecular dynamics simulations reveal that the strong electrostatic interaction induces the "upright" conformation on mica, whereas the hydrophobic interaction favors the "flat" conformation on HOPG. We also show that the epitaxial pattern on mica is ensured by the lattice matching between the anisotropic binding sites of the basal substrate and the molecular dimension of GAV-9, accompanied with a long-range order of well-defined ß-strands. Furthermore, the binding free energy surfaces indicate that the longitudinal assembly growth is predominantly driven by the hydrophobic interaction along the longer crystallographic unit cell direction of mica. These findings provide a molecular basis for the surface-assisted molecular assembly, which might also be useful for the design of de novo nanodevices.

Interactions between proteins and carbon-based nanoparticles: exploring the origin of nanotoxicity at the molecular level.

The widespread application of nanomaterials has spurred an interest in the study of interactions between nanoparticles and proteins due to the biosafety concerns of these nanomaterials. In this review, a summary is presented of some of the recent studies on this important subject, especially on the interactions of proteins with carbon nanotubes (CNTs) and metallofullerenols. Two potential molecular mechanisms have been proposed for CNTs' inhibition of protein functions. The driving forces of CNTs' adsorption onto proteins are found to be mainly hydrophobic interactions and the so-called π-π stacking between CNTs' carbon rings and proteins' aromatic residues. However, there is also recent evidence showing that endohedral metallofullerenol Gd@C82 (OH)22 can be used to inhibit tumor growth, thus acting as a potential nanomedicine. These recent findings have provided a better understanding of nanotoxicity at the molecular level and also suggested therapeutic potential by using nanoparticles' cytotoxicity against cancer cells.

Lipid Bilayer Strengthens the Cooperative Network of a Membrane-Integral Enzyme.

Lipid bilayer provides a two-dimensional hydrophobic solvent milieu for membrane proteins in cells. Although the native bilayer is widely recognized as an optimal environment for folding and function of membrane proteins, the underlying physical basis remains elusive. Here, employing the intramembrane protease GlpG of Escherichia coli as a model, we elucidate how the bilayer stabilizes a membrane protein and engages the protein's residue interaction network compared to the nonnative hydrophobic medium, micelles. We find that the bilayer enhances GlpG stability by promoting residue burial in the protein interior compared to micelles. Strikingly, while the cooperative residue interactions cluster into multiple distinct regions in micelles, the whole packed regions of the protein act as a single cooperative unit in the bilayer. Molecular dynamics (MD) simulation indicates that lipids less efficiently solvate GlpG than detergents. Thus, the bilayerinduced enhancement of stability and cooperativity likely stems from the dominant intraprotein interactions outcompeting the weak lipid solvation. Our findings reveal a foundational mechanism in the folding, function, and quality control of membrane proteins. The enhanced cooperativity benefits function facilitating propagation of local structural perturbation across the membrane. However, the same phenomenon can render the proteins' conformational integrity vulnerable to missense mutations causing conformational diseases1,2.

Free-energy simulations reveal that both hydrophobic and polar interactions are important for influenza hemagglutinin antibody binding.

Antibodies binding to conserved epitopes can provide a broad range of neutralization to existing influenza subtypes and may also prevent the propagation of potential pandemic viruses by fighting against emerging strands. Here we propose a computational framework to study structural binding patterns and detailed molecular mechanisms of viral surface glycoprotein hemagglutinin (HA) binding with a broad spectrum of neutralizing monoclonal antibody fragments (Fab). We used rigorous free-energy perturbation (FEP) methods to calculate the antigen-antibody binding affinities, with an aggregate underlying molecular-dynamics simulation time of several microseconds (∼2 µs) using all-atom, explicit-solvent models. We achieved a high accuracy in the validation of our FEP protocol against a series of known binding affinities for this complex system, with <0.5 kcal/mol errors on average. We then introduced what to our knowledge are novel mutations into the interfacial region to further study the binding mechanism. We found that the stacking interaction between Trp-21 in HA2 and Phe-55 in the CDR-H2 of Fab is crucial to the antibody-antigen association. A single mutation of either W21A or F55A can cause a binding affinity decrease of ΔΔG > 4.0 kcal/mol (equivalent to an ∼1000-fold increase in the dissociation constant K(d)). Moreover, for group 1 HA subtypes (which include both the H1N1 swine flu and the H5N1 bird flu), the relative binding affinities change only slightly (< ±1 kcal/mol) when nonpolar residues at the αA helix of HA mutate to conservative amino acids of similar size, which explains the broad neutralization capability of antibodies such as F10 and CR6261. Finally, we found that the hydrogen-bonding network between His-38 (in HA1) and Ser-30/Gln-64 (in Fab) is important for preserving the strong binding of Fab against group 1 HAs, whereas the lack of such hydrogen bonds with Asn-38 in most group 2 HAs may be responsible for the escape of antibody neutralization. These large-scale simulations may provide new insight into the antigen-antibody binding mechanism at the atomic level, which could be essential for designing more-effective vaccines for influenza.

Population genetic structure and conservation management of hill pigeons (Columba rupestris) recently endangered in South Korea.

BACKGROUND: Hill pigeons (Columba rupestris) are close to local extinction (ca. less than 100 individuals) in South Korea where a variety of conservation management procedures are urgently required. OBJECTIVE: This study was aimed at determining the conservation direction of captive propagation and reintroduction of hill pigeons using genetic information based on mitochondrial DNA. We also evaluated the extent of hybridization between hill pigeons and cohabiting domestic pigeons. METHODS: We used 51 blood samples of hill pigeons from Goheung (GH), Gurye (GR), and Uiryeong (UR), and domestic pigeons cohabiting with hill pigeon populations. Genetic diversity, pairwise Fst, analysis of molecular variance, and haplotype network analysis were used to examine the genetic structure of hill pigeons. RESULTS: Hill pigeons that inhabited South Korea were not genetically distinct from Mongolian and Russian populations and showed relatively low genetic diversity compared with other endangered species in Columbidae. The GR population that exhibited the largest population size showed lower genetic diversity, compared to the other populations, although the pairwise Fst values of the three populations indicated low genetic differentiation. The GH and GR populations were confirmed to lack hybridization, relatively, whereas the UR population was found to exhibit some degrees of hybridization. CONCLUSION: To conserve hill pigeons with low genetic diversity and differentiation in South Korea, the conservation process of captive propagation and reintroduction may require artificial gene flows among genetically verified populations in captivity and wildness. The introduction of foreign individuals from surrounding countries is also considered an alternative strategy for maintaining genetic diversity.

Deep generative molecular design reshapes drug discovery.

Recent advances and accomplishments of artificial intelligence (AI) and deep generative models have established their usefulness in medicinal applications, especially in drug discovery and development. To correctly apply AI, the developer and user face questions such as which protocols to consider, which factors to scrutinize, and how the deep generative models can integrate the relevant disciplines. This review summarizes classical and newly developed AI approaches, providing an updated and accessible guide to the broad computational drug discovery and development community. We introduce deep generative models from different standpoints and describe the theoretical frameworks for representing chemical and biological structures and their applications. We discuss the data and technical challenges and highlight future directions of multimodal deep generative models for accelerating drug discovery.