A Survey of Generative AI for de novo Drug Design: New Frontiers in Molecule and Protein Generation

De novo drug design

• De novo design generates biological entities from scratch beyond existing chemical libraries, expanding exploration into unknown chemical space.Source: Introduction · p. 1

Scope

• The survey covers ML-driven molecule and protein generation, with antibody and peptide generation as active, relevant protein subfields.Source: Introduction · p. 1

Molecule design

• More than 150 AI-focused biotechnology small-molecule drugs entered discovery, 15 entered clinical trials, and AI-fueled process use grew almost 40% annually.Source: Introduction · p. 1

Protein design

• Protein design serves immune signaling and targeted therapeutics [36], but complex biological constraints complicate generation; de novo methods extend beyond directed evolution’s existing-protein trajectories [38] [39] [40].Source: Introduction · p. 2

Survey structure

• The survey pairs a technical overview of architectures with task-level reviews of datasets, metrics, ML approaches, and SOTA comparisons, then synthesizes trends and future directions.Source: Introduction · p. 2

Related Surveys

• The survey takes a macro-level view of small molecule and protein generation to give newcomers a high-level introduction and reveal relationships across fields.It highlights parallel shifts in input representation, the emergence of equivariant graph neural networks (EGNNs), and similar challenges in molecule and protein design.Source: Related Surveys · p. 2
• Existing surveys focus on molecule generation [24, 43, 44], protein generation [36], antibody generation, or diffusion models [49] [50].Zhang et al. [49] and Guo et al. [50] review diffusion models in particular.Source: Related Surveys · p. 2
• While specialized surveys offer in-depth analysis, their narrow focus may limit scope compared with this survey’s cross-field perspective.The survey organizes its analysis around applications rather than model architecture alone.Source: Related Surveys · p. 2

Preliminary: Generative AI Models

• The survey identifies GANs, VAEs [52], flow-based models [53], and newer diffusion models [54] as prominent generative approaches.It provides a concise mathematical and computational overview of these architectures.Source: Preliminary: Generative AI Models · p. 2

Variational Autoencoders

• VAEs [52] represent latent attributes as distributions, enabling sampling from scratch; their loss balances reconstruction with KL regularization of the latent space.The adjustable β balances reconstruction and KL losses.Source: Variational Autoencoders · pp. 2, 4

Generative Adversarial Networks

• GANs train a generator to create samples from random noise and a competing discriminator to distinguish synthetic data from training samples.The discriminator maximizes correct classification likelihood while the generator minimizes it.Source: Generative Adversarial Networks · p. 4

Flow-Based Models

• Flow-based models [53] transform a simple latent distribution through invertible functions, enabling likelihood computation and modeling of complex data distributions.Composing invertible transformations preserves the dimensionality of the data and latent variable.Source: Flow-Based Models · p. 4

Diffusion Models

• Diffusion models [54] gradually add Gaussian noise in a forward process, then learn to denoise it in reverse to generate data from noise.The survey notes that Austin et al. [56] adapted diffusion to discrete forms such as molecular graphs.Source: Diffusion Models · p. 4

Other Models

• Other task-specific methods include transformers, EBMs, and BERT [57] [63]; GNNs are often paired with generative methods to capture molecular graph structure.The survey names EGNNs, MPNNs, GCNs [67], GINs [68], and CNNs [71] among network variants.Source: Other Models · p. 5

Overview

• Target-agnostic molecule design must learn validity and stability constraints across a vast compound space, while protein structure prediction remains challenging and needs improvement in speed and scale.Protein representation learning creates embeddings that can provide richer data spaces for generative models, though it is not itself generative.Source: Overview · pp. 5, 9

Task

• Target-agnostic generation creates novel, valid, stable molecules without input, whereas target-aware design uses protein sequence in LBDD or 3D structure in SBDD to generate molecules with target affinity.LBDD is useful when a target structure is unavailable but is limited in novelty; SBDD is generally preferred when the structure is available.Source: Task · pp. 5, 7, Overview · p. 7

Datasets

• QM9 contains small stable molecules and GEOM-Drug more complex, drug-like molecules; CrossDocked2020 is heavily used for target-aware design, with 22.5M versus 40K in Binding MOAD.CASP tests models biennially on solved protein structures not released to PDB, while CAMEO runs weekly blind tests using around 20 pre-released PDB targets.Source: Datasets · pp. 5, 7, 9

Metrics

• Molecule metrics include stability, uniqueness, novelty, and QED; novelty may be uninformative on exhaustive QM9, and QED is often unreported, while target-aware metrics additionally assess affinity, drug-likeness, synthetic accessibility, and diversity.Protein structure evaluation compares generated structures with ground truth, but the supplied passage does not specify the similarity measures.Source: Metrics · pp. 5, 7, 8

Models

• Diffusion models all exceed 98.5% atom stability on QM9; on GEOM-Drugs, MiDi generates more stable complex molecules at the expense of validity, while MDM leads four of six conditional tasks and GCDM leads the other two.Performance remains high on QM9, but the larger GEOM-Drugs molecules remain challenging.Source: Table 1 · p. 6, Table 2 · p. 6, Table 3 · p. 6
• Molecule-generation methods progressed from SMILES VAEs to 2D graphs, 3D geometric models, and joint 2D–3D diffusion, addressing representation limits and incorporating information relevant to binding.JTVAE builds and checks a tree-structured scaffold; later diffusion models address scalability and diversity, while JODO and MiDi jointly model connectivity and spatial information.Source: Models · pp. 6–7
• For target-aware design, DiffSBDD leads in Vina score and diversity, TargetDiff leads in high affinity, and Pocket2Mol’s MLP outperforms diffusion models on QED and SA.TargetDiff adjustments such as fragment-based generation or atom-bond prediction could improve QED and SA [99].Source: Models · p. 8

Task Background

• Protein-generation coverage includes representation learning, structure prediction, sequence generation, backbone design, antibodies, and peptides, whose shorter length and greater flexibility call for distinct computational models.Antibody methods must account for the absence of MSA input in some cases and the diversity of CDR regions, especially CDR-H3.Source: Task Background · pp. 8–9

Models

• Structure-prediction approaches range from MSA-based AlphaFold2 and trRosetta to RoseTTAFold, ESMFold, and diffusion-based EigenFold, with tradeoffs in speed, input representation, and accuracy.RoseTTAFold performs comparably to AlphaFold2 on CASP14 data and samples around 100 times faster; EigenFold is less computationally intensive but underperforms in accuracy and range.Source: Models · pp. 9–10

Antibody Structure Prediction

• Antibody structure predictors address the inefficiency of general models on antibody inputs: IgFold uses AntiBERTy embeddings and invariant point attention, while tFold-Ab reduces reliance on external tools such as Rosetta.IgFold achieves state-of-the-art generation speed with accuracy comparable to other models in the field; tFold-Ab performs comparably to IgFold while generating full-atom structures more efficiently.Source: Antibody Structure Prediction · p. 10

Overview; Task

• Fixed-backbone sequence design generates amino acid sequences intended to fold into a specified structure, a task motivated by an estimated sequence space of 10^65–10^130 and the complexity of protein folding.The survey prioritizes structure-conditioned methods over preliminary structure-agnostic approaches.Source: Overview · p. 10, Task · p. 10

Datasets; Metrics

• Sequence-design evaluation uses AAR, diversity, RMSD, and PPL, with Yu et al.’s independent benchmark supporting fair comparisons; CATH is commonly used for training and CATH and TS500 for evaluation.Yu et al. introduced 14 de novo structures absent from CATH to reduce contamination; the survey also includes PPL despite its omission from that benchmark.Source: Datasets · p. 10, Metrics · pp. 10–11

Models

• Structure-conditioned sequence models encode residue constraints with graph neural networks and geometric features, alongside autoregressive CNNs, iterative transformer refinement, and ProteinMPNN’s order-agnostic decoding.ESM-IF1 expands training pairs using MSA Transformer rankings and AlphaFold2-predicted structures; ProRefiner masks high-entropy predictions to limit error accumulation.Source: Models · p. 11
• ProteinMPNN leads sequence recovery, RMSD, and nonpolar loss, while GPD generates sequences about three times faster; diversity can be tuned, and current top models recover fewer than half of target residues.ProRefiner is an add-on that reports 2–7 percentage points of AAR improvement when refining GVP-GNN, ESM-IF1, or ProteinMPNN sequences.Source: Models · p. 11, Table 6 · p. 12

Tasks; Models

• Backbone design generates backbone-atom coordinates from scratch or context; context-free generation targets diverse structures, while context-given inpainting fills missing residues around a supplied motif.Methods progress from coordinate- and angle-based diffusion to frame-based approaches; RFDiffusion self-conditions and outperforms other frame-based models, while GPDL generates backbones 10–20 times faster than RFDiffusion. Co-design methods generate sequence and structure together, and the field faces complexity-scaling and latent-space-regularization challenges.Source: Tasks · p. 12, Models · pp. 12–13

Antibody CDR-H3 Generation; Peptide Generation

• Antibody CDR-H3 generation has moved from LSTM sequence generation to iterative sequence–structure co-design, with dyMEAN extending to structure prediction and docking; peptide methods include MMCD, which co-designs peptide sequences and structures and outperforms baselines on antimicrobial and anticancer peptide datasets.Peptide generation is treated as a specialized application because peptide structure is intricate and context-dependent, with diverse downstream uses.Source: Antibody CDR-H3 Generation · p. 13, Overview · p. 13, Peptide Generation · p. 13

Peptide-Protein Interaction

• For novel protein, peptide, and pair settings, PepGB raises AUCprecision by at least 9%, 9%, and 27%, and AUC-recall by 19%, 6%, and 4%, respectively.PepGB uses graph attention neural networks to learn peptide–protein interactions; it was trained on interaction benchmarks, validated with MDM2-targeting peptide analogs, and pretrained on UniProt peptide sequences.Source: Peptide-Protein Interaction · pp. 13–14

Peptide Representation Learning

• PepHarmony combines sequence (ESM) and structure (GearNet) encoders with contrastive or generative learning and outperforms ESM2 and GearNet on most reported downstream metrics.Its downstream tasks include cell-penetrating peptide, solubility, peptide–protein affinity, and self-contact map prediction.Source: Peptide Representation Learning · p. 14

Peptide Sequencing

• 0.373 peptide-level precision and 0.618 amino acid-level precision on human data make AdaNovo more precise than DeepNovo, PointNovo, and Casanovo.The respective peptide-level and amino acid-level precision pairs are DeepNovo (0.293, 0.610), PointNovo (0.351, 0.606), and Casanovo (0.343, 0.585). AdaNovo uses conditional mutual information and adaptive training to identify post-translational modifications and handle noisy mass-spectrometry data.Source: Peptide Sequencing · p. 14

• Generative AI methods are associated with faster, lower-cost drug development, alongside a shift from sequence-based to structure-based and integrated sequence-structure generation.Source: Current Trends · p. 14
• Graph-based diffusion models dominate molecular generation: GeoLDM and MiDi lead target-agnostic design, TargetDiff, Pocket2Mol, and DiffSBDD lead target-aware design, and Torsional Diffusion leads conformation generation.Source: Current Trends · p. 14
• For target-aware molecule design, structure-based approaches demonstrate clear advantages over sequence-based ligand-based approaches.Source: Current Trends · p. 14
• Protein generation is shifting from sequence to structure, with GearNET leveraging sequence-based ESM-1B and UniRep models, AlphaFold2 remaining SOTA for structure prediction, and RFDiffusion leading scaffolding.Source: Current Trends · p. 14

• Molecule-generation models produce valid, stable molecules on simple QM9 but struggle on the more complex GEOM-Drugs dataset.Source: Challenges and Future Directions · p. 15
• Models still struggle to generate molecules with high binding affinity for targets, making protein–molecule binding a challenging applicable task.Source: Challenges and Future Directions · p. 15
• Molecule-generation methods are largely black-box, and explainable AI remains undeveloped for identifying important atoms or structures.Source: Challenges and Future Directions · p. 15
• Protein design lacks standardized evaluation procedures, with differing model metrics and testing conditions impeding objective assessment of designed-protein quality.Source: Challenges and Future Directions · p. 15
• State-of-the-art protein-generation models still struggle with fold classification, gene ontology, and antibody CDR H3 generation.Source: Challenges and Future Directions · p. 15
• Some molecule and antibody-generation tasks rely on classifiers to predict properties or binding affinity, so improving those classifiers could better align predictions with real-world biology.Source: Challenges and Future Directions · p. 15

Conclusion

• The survey maps advances in molecule and protein generation and the datasets, model architectures, and evaluation metrics used to assess them.It presents these topics as an overview of the current landscape of generative AI in de novo drug design.Source: Conclusion · p. 15
• It identifies better benchmarking, improved explainability, and alignment with real-world tasks as priorities for increasing the applicability of generative AI in drug design.Source: Conclusion · p. 15
• The survey concludes that generative AI shows promise in drug design and that continued research may lead to future advances.Source: Conclusion · p. 15

• The survey covers novel small-molecule and protein generation, including task definitions, datasets, benchmarks, and model types for designing biologically plausible, pharmaceutically promising compounds.Source: Key Points · p. 15
• Equivariant graph neural networks and diffusion models have emerged as key drivers in the progression of drug-design architectures.Source: Key Points · p. 15
• The survey identifies challenges in applicability, performance, and scalability and delineates future research directions.Source: Key Points · p. 15
• An organized repository of covered sources is intended to facilitate collaboration at the intersection of computational biology and artificial intelligence.Source: Key Points · p. 15

Funding

• Xiangru Tang and Mark Gerstein are supported by Schmidt Futures.Source: Funding · p. 15

Graph Neural Networks

• GNNs are not generative models on their own, but serve as important components of larger generative methods.Source: Graph Neural Networks · p. 23
• For proteins, graphs represent amino acids as nodes and their spatial or sequential proximity as edges.Source: Graph Neural Networks · p. 23
• After data are transformed into a graph G = (V, E), a GNN maps nodes to embeddings through message passing and aggregation.Source: Graph Neural Networks · p. 23
• For each node pair, the GNN computes a message based on existing node features and the corresponding adjacency-matrix entry.Source: Graph Neural Networks · p. 23
• Each node aggregates messages received from its neighbors.Source: Graph Neural Networks · p. 23
• At the next layer, the model combines prior embeddings and positions with messages to produce updated embeddings and positions.Source: Graph Neural Networks · p. 23
• Positions and features are often concatenated into a single feature vector because general GNNs treat them equivalently.Source: Graph Neural Networks · p. 23

Equivariant Graph Neural Networks

• For conditional distributions, equivariance transforms inputs and outputs together, while invariance leaves the distribution unchanged; the transformations commonly use the Euclidean E(3) group.For 3D structures, rotations, reflections, and translations should not change how related structures are treated.Source: Equivariant Graph Neural Networks · p. 23
• The EGNN adapts a traditional GNN with separate point and feature embeddings per node and relative positions or distances in message passing, preserving equivariance under global rotations, translations, and reflections.Its final message aggregation and embedding updates are equivalent to those of a GNN.Source: Equivariant Graph Neural Networks · p. 23

Overview

• Conformers are distinct 3D arrangements of the same molecular connectivity, and 3D information is important for understanding chemical properties such as binding affinity.Protein representation learning instead maps sequence or structure tokens to embeddings intended to capture chemical traits for use by other models.Source: Overview · pp. 23–24

Datasets

• GEOM-QM9 provides representative conformers where QM9 supplies only one 3D structure per molecule; protein pretraining uses sequence-focused and structure-focused datasets, while antibody models pretrain on OAS.GEOM-Drugs covers more complex drug-like molecules; Pfam is used for both sequence- and structure-based protein learning.Source: Datasets · pp. 23, 24, 26

Task

• Conformation generation maps a 2D connectivity graph to stable 3D conformations; COV measures coverage of reference conformers, while MAT measures average closest-reference RMSD.Coverage thresholds are typically 0.5 Å for QM9 and 1.25 Å for Drugs.Source: Task · p. 23, Metrics · p. 23

Models

• Torsional Diffusion [177] appears to outperform other conformation models, while GeoMol [172] outperforms distance-graph methods GraphDG [171] and CGCF [169].Table 8 notes that Torsional Diffusion uses a 0.75 Å rather than 1.25 Å coverage threshold on GEOM-Drugs, which deflates its score; it still outperforms GeoDiff and GeoMol at that threshold.Source: Models · p. 24, Table 8 · p. 24

Tasks

• Protein representation benchmarks cover contact prediction, fold classification, stability prediction, and protein–protein interaction, but the field lacks uniform evaluation standards.TAPE and PEER compile metrics into standardized benchmarks; these tasks use precision, accuracy, or Spearman’s ρ as appropriate.Source: Tasks · pp. 24–25, Metrics · p. 25

Models

• Protein representation models learn from sequences, structures, or both, using transformers, evolutionary information, knowledge injection, or residue-graph architectures to capture protein features.ESM-1B used a 33-layer, 650-million-parameter transformer but was expensive to train; RSA [40] retrieves evolutionary information without costly MSA alignment, while OntoProtein [38] and KeAP [39] inject biological knowledge.Source: Models · pp. 25–26

Antibody Task Background

• Antibody generation centers on the highly variable CDR-H3, and antibody-specific structure prediction avoids reliance on MSA because relevant evolutionary histories may be unavailable.The broader pipeline can include sequence representation, structure prediction, paratope–epitope prediction, docking, CDR generation, and evaluation; some recent methods jointly generate sequence and structure or address the end-to-end process.Source: Antibody Task Background · p. 26, Overview · p. 27

Models

• PARA outperforms previous antibody representation models on CDR-H3 amino acid recovery, while other models report gains on paratope prediction or B-cell clustering.There is no standardized antibody representation benchmark, so models use differing evaluations; Table 10 reports each model’s findings.Source: Models · p. 27, Table 10 · p. 27

Metrics

• Evaluation metrics include RMSD and OCD for structural accuracy, AAR for sequence recovery, TM-score for local and global structural similarity, and ΔΔG for binding-energy change.OCD measures similarity between light-heavy orientational coordinates, including the relative position of antibody heavy and light chains; ΔΔG is predicted using Shan et al.’s geometric network.Source: Metrics · pp. 27–29

Models

• Antibody structure models range from AlphaFold-like approaches to neural networks that predict restraints or individual CDR structures, with DeepAb adding interpretable attention layers.tFold-Ab, xTrimoABFold [201], and ABodyBuilder2 omit MSA searching; AbLooper uses five EGNNS to predict CDR structures and estimates confidence from their deviation. DeepH3 uses Rosetta to turn predicted restraints into atomic structures, while SimpleDH3 [203] adds ELMo embeddings.Source: Models · p. 27
• IgFold is state-of-the-art for speed despite not leading RMSD or OCD, while tFold-Ab produces full-atom structures faster than the IgFold–PyRosetta pipeline.IgFold takes about 0.46 seconds to produce backbone atoms, with PyRosetta adding an average 21.86 seconds; tFold-Ab takes 2.23 seconds versus 22.32 seconds for the full IgFold–PyRosetta pipeline. The reported times use different benchmarks, so they may not be directly comparable.Source: Models · pp. 27–28, Table 11 · p. 28

Overview

• CDR-H3 is the most variable and difficult antibody region to generate, making it the primary focus of CDR generation.CDR generation concerns antibody variable regions that directly determine binding affinity.Source: Overview · p. 28

Datasets

• SAbDab is the primary training dataset, while RAbD and SKEMPI support testing of specific downstream tasks.SAbDab contains annotated antibody structures from PDB; RAbD is a hand-selected set of 60 diverse antibody–antigen complexes, and SKEMPI records energy changes for protein-interaction mutations.Source: Datasets · p. 28

Tasks

• Benchmarks cover sequence–structure modeling on SAbDab, CDR-H3 generation on RAbD, and affinity optimization on SKEMPI v2.0.AAR evaluates sequences, RMSD evaluates structures, TM-Score evaluates CDR-H3 structures, and ΔΔG measures affinity-optimization performance.Source: Tasks · p. 28

Models

• CDR design has progressed from sequence-based methods to structure–sequence co-design, while dyMEAN [158] combines structure prediction, docking, and CDR generation end to end.RefineGNN [157] introduced iterative co-design; MEAN [189] predicts all CDR amino acids at once, AntiDesigner [211] designs sequences and structures in one shot, and DiffAb diffuses over sequences and backbone coordinates. DockGPT jointly docks and designs CDRs, while HERN [208] requires an identified epitope; other CDR-generation methods require fully docked antibody–antigen structures as input.Source: Models · p. 29

Signal Peptide Prediction

• PEFT-SP fine-tunes ESM2 with Prompt Tuning and LoRA, outperforming SignalP 6.0 and other models in most signal peptide prediction categories.Performance was evaluated separately for Archaea, Eukarya, Gram-positive bacteria, and Gram-negative bacteria.Source: Signal Peptide Prediction · p. 29