Generative Modeling in Protein Design: Neural Representations, Conditional Generation, and Evaluation Standards

Senura Hansaja Wanasekara; Minh-Duong Nguyen; Xiaochen Liu; Nguyen H. Tran; Ken-Tye Yong

cs.LGMar 27, 2026

Generative Modeling in Protein Design: Neural Representations, Conditional Generation, and Evaluation Standards

Authors: Senura Hansaja Wanasekara, Minh-Duong Nguyen, Xiaochen Liu, Nguyen H. Tran, Ken-Tye Yong

Organizations: Biomedical Engineering, The University of Sydney, Darlington, NSW 2006, Australia. · the College of Engineering and Computer Science, VinUniversity, Vietnam. · School of Computer Science and School, The University of Sydney, Darlington, NSW 2006, Australia.

Abstract

Generative modeling has become a central paradigm in protein research, extending machine learning beyond structure prediction toward sequence design, backbone generation, inverse folding, and biomolecular interaction modeling. However, the literature remains fragmented across representations, model classes, and task formulations, making it difficult to compare methods or identify appropriate evaluation standards. This survey provides a systematic synthesis of generative AI in protein research, organized around (i) foundational representations spanning sequence, geometric, and multimodal encodings; (ii) generative architectures including $\mathrm{SE}(3)$ -equivariant diffusion, flow matching, and hybrid predictor-generator systems; and (iii) task settings from structure prediction and de novo design to protein-ligand and protein-protein interactions. Beyond cataloging methods, we compare assumptions, conditioning mechanisms, and controllability, and we synthesize evaluation best practices that emphasize leakage-aware splits, physical validity checks, and function-oriented benchmarks. We conclude with critical open challenges: modeling conformational dynamics and intrinsically disordered regions, scaling to large assemblies while maintaining efficiency, and developing robust safety frameworks for dual-use biosecurity risks. By unifying architectural advances with practical evaluation standards and responsible development considerations, this survey aims to accelerate the transition from predictive modeling to reliable, function-driven protein engineering.

Figures & tables

Symbol	Meaning
$S=(s_{1},\dots,s_{n})$	Amino acid sequence of length $n$
$n$	Sequence / residue length
$\mathcal{A}$	Amino acid alphabet ( $\|\mathcal{A}\|=20$ )
$\mathbf{X}\in\mathbb{R}^{n\times 3}$	Backbone or atomic coordinates
$T_{i}=(\mathbf{R}_{i},\mathbf{t}_{i})$	Residue frame ( $\mathbf{R}_{i}\!\in\!\mathrm{SO}(3)$ , $\mathbf{t}_{i}\!\in\!\mathbb{R}^{3}$ )
$\mathrm{SE}(3)^{n}$	Product of $n$ residue-frame groups

Table I: Summary of principal notation used in this survey.

Figure 1: Illustration of the hierarchical organization of protein structure, progressing from the (a) linear amino acid sequence (primary structure) through (b) local folding patterns (secondary), (c) overall 3D conformation (tertiary), (d) to multi-subunit complexes (quaternary).

Family	Symmetry	Information Content	Typical Tasks
Frame-based	$\mathrm{SE}(3)$ -equivariant	Full backbone pose per residue	Structure prediction, backbone generation
Coordinate GNN	$\mathrm{SE}(3)$ -equivariant	Atomic positions with vector features	Inverse folding, docking, property prediction
Invariant	$\mathrm{SE}(3)$ -invariant	Pairwise distances, dihedrals, orientations	Contact prediction, scoring, classification
Multimodal	Task-dependent	Sequence + structure + function tokens	Joint generation, cross-modal retrieval

Table II: Comparison of protein representation families.

Model Type	Representation	Symmetry	Strengths	Limitations	Typical Tasks
VAE	Sequence or latent backbone embeddings	Invariant	Continuous latent space enables smooth interpolation between folds or sequences; supports property-guided latent optimization	Posterior collapse; difficulty enforcing long-range tertiary contacts; limited stereochemical fidelity	Sequence generation; fold interpolation; low-dimensional backbone sampling
GAN	Sequence, contact map, or distance map	Typically invariant	Implicit density modeling; capable of generating sharp structural patterns when stable	Mode collapse reduces structural diversity; unstable training; weak geometric constraint enforcement	Contact/distance map generation; structural feature augmentation
Autoregressive / Masked Transformer	Sequence tokens; optional structure-aware extensions	Invariant	Captures long-range evolutionary dependencies; scalable pretraining; strong conditional generation and zero-shot variant scoring	Limited explicit geometric awareness unless coupled with structure modules; high data and compute demands; evolutionary bias	Protein language modeling; inverse folding; variant prediction; sequence design
Score-Based Diffusion (DDPM)	3D backbone or atomic coordinates	Often $\mathrm{SE}(3)$ -equivariant	Stable training; models coordinate uncertainty and conformational ensembles; flexible conditioning on motifs or targets	Slow iterative sampling; explicit stereochemical constraints required; scaling to large complexes is costly	De novo backbone generation; motif scaffolding; conformational sampling
Flow Matching	3D backbone or atomic coordinates	Often $\mathrm{SE}(3)$ -equivariant	Continuous-time formulation enables fewer sampling steps; scalable vector-field parameterization	Sensitive to solver stability and vector-field estimation; implementation complexity	Fast backbone generation; peptide and scaffold design
Energy-Based / Score Models	Sequence or structure space	Invariant or equivariant	Integrates physical priors and multi-state objectives; flexible constrained generation	Intractable partition functions; expensive sampling in high dimensional spaces	Multi-state design; docking pose sampling; stability-aware optimization

Table III: Comparison of major generative model families in protein modeling and design. We summarize representation type, symmetry properties, protein-specific strengths and limitations, and typical tasks.

Model	MSA	Complexes	All-Atom	Open Source	Key Features
AlphaFold2 [ 18 ]	Yes	No	Backbone	Weights only	Evoformer + IPA; deterministic refinement
RoseTTAFold [ 19 ]	Yes	Extended	Backbone	Yes	Three-track architecture; single-GPU inference
ESMFold [ 39 ]	No	No	Backbone	Yes	ESM-2 embeddings; millisecond inference
AlphaFold3 [ 21 ]	Yes	Yes	Yes	Restricted	PairFormer + diffusion; unified complex prediction
Boltz-1 [ 31 ]	Yes	Yes	Yes	Yes (MIT)	AF3-level accuracy; full training code released
Chai-1 [ 32 ]	Yes	Yes	Yes	Weights only	Multimodal; commercial API; PoseBusters 77%

Table IV: Comparison of representative structure prediction models.

Method	Backbone Setting	Conditioning	Architecture	Key Feature
ESM-IF1 [ 99 ]	Fixed	Backbone only	GNN encoder + Transformer decoder	Trained on 12M AF2-predicted structures
ProteinMPNN [ 34 ]	Fixed	Backbone only	Message-passing GNN	Standard in de novo pipelines; fast sampling
CarbonDesign [ 126 ]	Fixed	Backbone only	Evoformer + amortized MRF	Global pairwise coupling; multitask side-chain prediction
DynamicMPNN [ 127 ]	Flexible (multi-state)	Structural ensembles	MPNN on aligned ensembles	Multi-state design; conformational heterogeneity
AntiDIF [ 128 ]	Fixed (antibody)	Backbone + CDR context	Discrete diffusion + PiGNN	Antibody-specific; diverse CDR sampling
LigandMPNN [ 36 ]	Fixed	Backbone + ligand atoms	MPNN with ligand featurization	Ligand-conditioned; enzyme/binder design

Table V: Comparison of representative inverse-folding methods.

Figure 2: Overview of the ESM-IF inverse folding architecture. The model transforms a protein’s 3D structure into its corresponding amino acid sequence by processing vector and scalar features through a structure encoder, followed by sequence decoding using a transformer-based architecture.

Figure 3: Overview of the DiffDock model architecture. The system integrates a 3D protein structure and a ligand pose as inputs, which are processed through a score model and a confidence model to predict docking outcomes. The output includes top ligand poses with associated 3D coordinates (translation, rotation, torsion angles) and a confidence score, enabling accurate structure-based drug docking predictions.

Figure 4: DLM-DTI employs a dual-encoder framework for drug-target interaction prediction. The Target Encoder processes protein sequences using a Teacher Model (ProtBERT) and a Student Model, while the Drug Encoder utilizes ChemBERTa to encode chemical structures. The outputs from both encoders are concatenated and passed through a feed-forward neural network to predict binding probability.

Model	Type	Input Modality	Primary Output	Architecture Highlights	Strengths	Limitations
iNGNN-DTI [ 147 ]	Structure + Sequence	Drug graph + Protein graph (AlphaFold2) + embeddings	Binary DTI	NestedGNN; cross-AFT interaction; fusion with Chemformer/ESM	3D context; interpretable; strong generalization	Requires structure; heavy preprocessing
DLM-DTI [ 146 ]	Sequence	Protein sequence + SMILES	Binary DTI	DualLM (ProtBERT+ student); ChemBERTa; concatenation head	Lightweight; competitive performance	Limited interpretability; likelihood only
PLAPT [ 144 ]	Sequence	Protein sequence + SMILES	Binding affinity	Transfer-learning; Branching NN (1792d embeddings)	Fast; compute-efficient; good for screening	No 3D pose; no pocket context
DiffDock [ 35 ]	Structure	Protein 3D + ligand graph	3D binding pose	$\mathrm{SE}(3)$ -equivariant diffusion; confidence model	Blind docking; multiple modes; fewer clashes	Requires 3D; no affinity; runtime scales
SurfDock [ 141 ]	Structure	Surface-mesh + residue-graph + ligand	3D pose + confidence	Surface-informed diffusion; SurfScore MDN scorer	SOTA accuracy; surface-aware; strong generalization	Complex pipeline; requires preprocessing

Table VI: Comparison of representative AI models for protein-drug interaction tasks.

Task	Key Metrics
Structure prediction	RMSD, GDT-TS, pLDDT, pTM, PAE
De novo design	scRMSD, designability rate, novelty (TM-score to PDB), pLDDT
Inverse folding	Sequence recovery (%), pairwise dissimilarity, Shannon entropy
Docking	Ligand RMSD ( $\leq 2\,\text{\AA}$ ), PoseBusters validity, steric clash rate
Affinity prediction	Pearson $r$ , RMSE, virtual screening AUC
PPI	DockQ, interface RMSD, $F_{\text{nat}}$

Table VII: Unified summary of evaluation metrics by task.

Explore similar work

SimpleDesign: A Joint Model for Protein Sequence and Structure Codesign
Sep 3, 2026Jiarui Lu, Yuyang Wang, Yizhe Zhang +4Unified Multimodal Models Generative Modeling
Mila, Université de Montréal, Apple · Apple
Protein Structure Prediction: From Evolutionary Constraints to Generative Modeling
Aug 17, 2026Wengan He, Yongsheng Luo, Lihong Jiang +2Protein Structure Prediction Protein Representation Learning
Zhuhai College of Science and Technology, Zhuhai, China · Zhuhai UNO Technology Co., Ltd., Zhuhai, China
From Snapshots to Symphonies: The Evolution of Protein Prediction from Static Structures to Generative Dynamics and Multimodal Interactions
Mar 19, 2026Jingzhi Chen, Lijian XuProtein Structure Prediction Protein Representation Learning
Shenzhen University of Advanced Technology

Sep 3, 2026J/K move · Enter open · S save

SimpleDesign: A Joint Model for Protein Sequence and Structure Codesign

Jiarui Lu, Yuyang Wang, Yizhe Zhang, Jiatao Gu+3

Mila, Université de Montréal, Apple · Apple

Unified Multimodal Models Generative Modeling Multimodal Generation Molecular Generation Protein Design

Open paper AskPDF arXiv

Proteins are fundamental to biological processes, with their function determined by the complex interplay between the amino acid sequence and the three-dimensional structure. Developing generative models capable of understanding this intrinsically multi-modal relationship is crucial for fields like drug discovery and protein engineering. Existing models often rely on a multi-stage training process where autoencoders that tokenize data into latent representations are trained in a first stage. Secondly, a generative model is trained on the latent representation of the autoencoder(s), i.e., generative modeling in a latent space. We hypothesize that this multi-stage training is not necessary to obtain performant co-design models and thus present SimpleDesign, an effective multi-modal protein design model trained directly in the data space. SimpleDesign leverages a single-stage end-to-end objective that combines discrete cross-entropy for sequences and a regression objective for structures. In order to effectively model the difference in sequence and structure modalities, we develop a Mixture-of-Transformer architecture that allows modality-specific processing while keeping global self-attention over both modalities. We train SimpleDesign on over 2M sequence-structure pairs achieving strong performance across co-design and unconditional sequence/structure generation benchmarks.

Abstract

Explore similar work

SimpleDesign: A Joint Model for Protein Sequence and Structure Codesign

Protein Structure Prediction: From Evolutionary Constraints to Generative Modeling

From Snapshots to Symphonies: The Evolution of Protein Prediction from Static Structures to Generative Dynamics and Multimodal Interactions