Artificial Intelligence-Driven Protein Design:
A Paradigm Revolution from Structure Prediction to De Novo Design

Introduction

Figure 2 | Complete AI protein design workflow. The five core stages proceed as follows: ① Target Definition → ② Backbone Generation → ③ Inverse Folding → ④ Structure Validation → ⑤ Experimental Testing. The dashed arrow represents the iterative optimization feedback loop based on experimental results.

Proteins are three-dimensional functional molecules folded from amino acid sequences, responsible for nearly all core biological activities including catalysis, signal transduction, structural support, and immune defense. Understanding and designing proteins is one of the "holy grails" of modern biotechnology. However, the protein sequence space is immense — a protein of 100 amino acids theoretically admits 20¹⁰⁰ possible sequences — and traditional experimental screening and physics-based computational methods (such as Rosetta) face fundamental bottlenecks in efficiency and coverage.

In the early 2020s, deep learning methods exemplified by AlphaFold2 achieved a landmark breakthrough in protein structure prediction, elevating prediction accuracy to near-experimental levels. This achievement not only resolved the "protein folding problem" that had challenged structural biology for half a century, but also catalyzed a far-reaching revolution: AI is no longer used merely to predict the structures of known proteins, but is now being deployed to design entirely novel proteins that have never existed in nature.

This review systematically traces the technical trajectory of AI protein design, proceeding from the foundations of structure prediction to the frontiers of generative design, and looks ahead to its broad application prospects in biomedicine and industrial biotechnology.

Protein Structure Prediction: The Starting Point of the AI Revolution

2.1 From CASP to AlphaFold: A Leap in Prediction Accuracy

Computer-aided protein structure prediction has a history spanning several decades, yet accuracy remained limited for much of that time. The biennial Critical Assessment of Protein Structure Prediction (CASP) competition serves as the authoritative benchmark for progress in the field. In 2020, DeepMind's AlphaFold2 won CASP14 by an overwhelming margin, achieving prediction accuracy (median GDT_TS > 92) that for the first time matched experimental structures — hailed as the solution to a "50-year-old problem."

The core innovation of AlphaFold2 lies in its Evoformer architecture: through iterative attention processing of multiple sequence alignments (MSA) and residue-pair representations, the model simultaneously captures evolutionary covariation information and spatial geometric constraints, ultimately outputting high-confidence atomic coordinates. DeepMind subsequently applied AlphaFold2 to the human proteome, predicting the structures of nearly all human proteins (~20,000 entries) and publicly releasing the results in the AlphaFold Protein Structure Database, covering more than 200 million protein sequences.

2.2 AlphaFold3: Extension to All-Atom Biomolecular Complexes

In 2024, DeepMind released AlphaFold3, extending prediction capability from individual proteins to complex systems containing proteins, nucleic acids, small molecules, ions, and modified residues. AlphaFold3 employs a diffusion-based architecture (Diffusion Module) in place of the structure module used in AlphaFold2, achieving significant improvements in the structural prediction of protein–ligand, protein–DNA/RNA, and other complexes. This advance has direct implications for structure-based drug design.

2.3 Single-Sequence Structure Prediction Driven by Protein Language Models

AlphaFold2 relies on multiple sequence alignments (MSA), which limits its performance for orphan proteins or rapidly evolving proteins. ESMFold, developed by Meta AI, combines a large-scale protein language model (ESM-2, with up to 15 billion parameters) with a structure prediction module, enabling atomic-level structure prediction from a single sequence alone — approximately 60 times faster than AlphaFold2. This property makes it particularly well-suited for large-scale proteomics analyses and iterative design scenarios.

2.4 Limitations of Structure Prediction

Protein Language Models: Deep Representation of Sequence Space

Figure 3 | Architectural taxonomy of AI protein design methods. Starting from the root node "AI Protein Design," methods are organized into four columns by architecture family: Transformer/Protein Language Models, GNN/MPNN, Diffusion Models, and Hybrid/Multimodal. Nodes marked [★] denote milestone methods.

3.1 From NLP to Proteins: Transfer of the Language Model Paradigm

Protein sequences share a profound structural analogy with natural language: amino acids serve as "letters," protein sequences as "sentences," and evolutionary pressure shapes the "grammatical" rules of sequences. This insight has motivated researchers to transfer the large-scale pre-trained language model paradigm from NLP to protein science.

ProtTrans (Elnaggar et al., 2021) is an early landmark work, training multiple Transformer architectures (BERT, T5, etc.) on the UniRef and BFD databases containing 393 billion amino acids. The ESM series (Meta AI) pushed scale to the extreme: the largest version of ESM-2 contains 15 billion parameters, and atomic-level structural information spontaneously emerges in its embedding representations.

3.2 ESM3: A Multimodal Protein Generative Model

In 2025, Hayes et al. released ESM3, a multimodal generative language model capable of simultaneously reasoning over protein sequence, structure, and function. Trained on approximately 2.7 billion protein sequences, ESM3 can respond to complex cross-modal prompts. Using ESM3, researchers designed a fluorescent protein that differs from the native GFP sequence by 58% — equivalent to "simulating 500 million years of evolution."

3.3 Downstream Applications

Protein language models have become general-purpose foundation models for protein science. Their embedding representations are widely used for: variant effect prediction, protein function annotation, protein–protein interaction prediction, and as initialization weights for downstream design models.

Diffusion Models: A Paradigm Breakthrough in De Novo Backbone Generation

4.1 Principles of Diffusion Models

Diffusion models are among the most important breakthroughs in generative AI in recent years. The core idea is to progressively add noise to data (the forward process), then train a neural network to learn the reverse denoising process, thereby enabling the generation of high-quality samples from pure noise. In protein design, the "data" are the three-dimensional coordinates of protein backbones.

A key challenge is that protein backbones exist in SE(3) space (the group of three-dimensional rotations and translations), requiring equivariant neural network architectures to ensure physical plausibility.

4.2 RFdiffusion: Fusion of Diffusion Models and Structure Prediction Networks

RFdiffusion (Watson et al., 2023) is the most influential protein backbone generation diffusion model to date. Its core innovation lies in fine-tuning the RoseTTAFold structure prediction network as a denoising network, fully leveraging the rich protein geometric knowledge embedded in the pre-trained structure prediction model.

In 2025, RFdiffusion3 further extended design capability to the all-atom level, enabling protein structure generation under constraints from ligands, nucleic acids, and other non-protein molecules.

4.3 Chroma: Programmable Protein Generation

Chroma (Ingraham et al., 2023) proposes a "programmable" protein generation framework. Unlike RFdiffusion, Chroma is trained from scratch, employing a random graph neural network to process protein structures and achieving fine-grained control through composable conditioning modules — including symmetry constraints, shape constraints, and sequence constraints.

4.4 Sequence-Space Diffusion: ProteinGenerator and DPLM

ProteinGenerator (Lisanza et al., 2024) performs diffusion in sequence space based on RoseTTAFold, enabling simultaneous generation of protein sequences and structures. The Diffusion Protein Language Model (DPLM) combines a discrete diffusion framework with protein language model pre-training, demonstrating strong capability on both generative and predictive tasks.

Inverse Folding: Designing Sequences for Backbones

5.1 Definition and Importance

Protein design is typically divided into two stages: first generating a target backbone structure, then designing an amino acid sequence capable of folding into that structure. The latter is the "inverse folding" problem, and it serves as the critical bridge connecting backbone generation to experimental validation.

5.2 ProteinMPNN: A Milestone in Deep Learning Inverse Folding

ProteinMPNN (Dauparas et al., 2022) is currently the most widely used deep learning inverse folding method. It employs a message passing neural network (MPNN) architecture, taking protein backbone coordinates as input and outputting the amino acid probability distribution at each position.

5.3 LigandMPNN: Extension to Non-Protein Environments

Native proteins often function through interactions with small molecules, nucleic acids, or metal ions, yet standard ProteinMPNN cannot handle these non-protein components. LigandMPNN extends inverse folding capability to enzyme active site design and small-molecule binding protein design by explicitly modeling all non-protein atoms, significantly outperforming Rosetta and ProteinMPNN on relevant benchmarks.

Application Prospects: From the Laboratory to the Clinic

6.1 Enzyme Engineering: Transcending the Boundaries of Natural Evolution

Enzymes are the core tools of industrial biotechnology and green chemistry, yet natural enzymes often fail to meet industrial requirements. AI-driven enzyme design is fundamentally transforming this landscape.

In the area of de novo enzyme design, Yeh et al. (2023) used a deep learning "hallucination" approach to design a series of de novo luciferases with diverse active site geometries, with the best designs achieving catalytic efficiencies approaching those of natural luciferases. In the area of enzyme optimization, machine learning-guided directed evolution has become the dominant strategy.

6.2 Antibody and Protein Therapeutic Design

Antibodies are currently the most important class of protein therapeutics, with more than 165 antibody drugs approved worldwide. AI is accelerating every stage of antibody discovery and optimization — from sequence design (AbDPP) to antibody–drug conjugate (ADC) design and antimicrobial peptide (AMP) discovery.

6.3 Structure-Based Drug Design

The intersection of AI protein design and small-molecule drug discovery is catalyzing new research paradigms. Diffusion model-based methods such as DiffSBDD can directly generate small-molecule ligands under the constraints of protein binding pockets. AlphaFold3's high-accuracy prediction of protein–ligand complex structures provides powerful support for virtual screening and lead compound optimization.

Challenges and Limitations

Future Perspectives

Sequence–Structure Co-generation: The current mainstream design pipeline separates backbone generation and sequence design into two independent steps. Methods like Chroma, ProteinGenerator, and DPLM attempt to jointly model sequence and structure, promising more consistent and functional protein designs.

Multimodal and All-Atom Design: Future models will need to simultaneously handle proteins, nucleic acids, small molecules, and metal ions. AlphaFold3 and RFdiffusion3 have already taken important steps in this direction.

Reinforcement Learning and Self-Improvement: Incorporating RL into protein design — continuously optimizing design strategies through experimental feedback — is an important pathway toward "closed-loop" protein engineering. ProteinZero demonstrates a self-improving inverse folding framework based on online reinforcement learning.

Automated High-Throughput Experimental Validation: The deep integration of automated experimental platforms (robotic synthesis, cell-free expression systems) with AI design will establish a true "design–build–test–learn" closed loop, dramatically accelerating protein engineering iteration speed.

Conclusion

Over the past five years, AI-driven protein design has undergone a transformation from quantitative accumulation to qualitative breakthrough. AlphaFold2 solved the protein structure prediction problem that had challenged the field for half a century; RFdiffusion brought the generative power of diffusion models to protein backbone design; ProteinMPNN redefined the accuracy frontier of inverse folding with deep learning; and ESM3 demonstrated the enormous potential of protein language models in multimodal generation.

The synergistic application of these methods is advancing protein design from "fine-tuning natural proteins" to a new stage of "creating functional proteins from scratch."