AI-Assisted Peptide Design: From Sequence to Drug Candidate

Over the past several decades, peptide drugs have undergone a major transformation from laboratory tools to blockbuster medicines. With the successful launch of GLP-1 receptor agonists, peptide hormone analogs, and multiple oncology therapeutics, peptides have become one of the fastest-growing areas in modern drug discovery and development. However, compared with small-molecule drugs, the peptide design process has long remained highly dependent on experience and experimental screening. Researchers often need to construct large numbers of candidate sequences and identify molecules with the desired activity through repeated synthesis and testing.

The development of artificial intelligence technologies is changing this situation. In recent years, machine learning, deep learning, and generative artificial intelligence have begun to be widely applied in protein and peptide research, enabling researchers to predict activity, evaluate stability, optimize structures, and even directly generate entirely new candidate sequences before experiments begin. Peptide R&D is gradually shifting from the traditional “experiment-driven” model toward “data-driven” and “model-driven” approaches.

Why Peptide Design Is Difficult

On the surface, peptides are simply formed by linking amino acids in a specific order, but what truly determines their function is not the sequence alone. In theory, a peptide composed of twenty amino acids can generate more than 10²⁶ different combinations. If non-natural amino acids, cyclization strategies, terminal modifications, and side-chain modifications are taken into account, the possible design space expands further to an almost unimaginable scale. For researchers, the real challenge is not whether enough sequences can be generated, but how to quickly identify the most promising candidate molecules from an enormous sequence space.

At the same time, the biological activity of peptides is jointly influenced by three-dimensional structure, conformational flexibility, stability, solubility, immunogenicity, and pharmacokinetic properties. Even two sequences that differ by only a single amino acid may exhibit completely different activities and in vivo behaviors. Therefore, peptide design is essentially a complex multiparameter optimization problem.

How AI Understands Peptide Sequences

Traditional computational methods often rely on predefined rules. For example, they perform analyses based on hydrophobicity, charge distribution, or known structures. Although these methods are effective, they have difficulty handling complex nonlinear relationships. Modern artificial intelligence models use a completely different approach. By learning information from massive protein and peptide databases, AI can establish associations between sequence and structure, and between structure and function. For the model, an amino acid sequence is similar to a special language. Each amino acid is like a word, while the entire peptide forms a biologically meaningful “sentence.”

The Large Language Models that have emerged in recent years have already demonstrated that similar methods can understand not only human language but also biological sequences. By learning hundreds of millions of natural protein and peptide sequences, models gradually learn which amino acid combinations are more likely to form stable structures, which regions are more likely to participate in target binding, and which mutations may improve activity or stability. This capability means that AI is no longer merely a data analysis tool, but is beginning to become a design tool in the true sense.

From Prediction to Generation

Early AI was mainly used for predictive tasks. Researchers entered a peptide sequence, and the model predicted its activity, toxicity, or stability. Although this approach could reduce some of the experimental workload, in essence it was still looking for answers within existing sequences. The greatest breakthroughs in recent years have come from generative artificial intelligence. Generative models can not only evaluate existing sequences, but also actively create new ones. After researchers provide a target protein, binding interface, or functional requirement, the model can directly propose entirely new candidate molecules. This model is similar to a shift from “screening” to “design.” In the past, it might have been necessary to build tens of thousands of candidate molecules to find one active molecule, whereas in the future, generating only a few dozen high-quality candidate sequences may be sufficient to enter the experimental validation stage.

This change is significantly shortening the drug discovery and development cycle.

Structure Prediction Has Changed the Rules of the Game

One of the most important breakthroughs of artificial intelligence in the life sciences is the leap in protein structure prediction capability.

A new generation of structure prediction models represented by AlphaFold enables researchers to predict three-dimensional protein structures with unprecedented accuracy.

The improvement in structure prediction capability has profound implications for peptide design. In the past, the binding mode between a peptide and its target often had to be obtained through crystallography or cryo-electron microscopy experiments. Now, researchers can use computational models to predict the spatial conformation formed after a peptide complexes with a target protein, thereby evaluating binding modes and key interaction sites in advance. This means that most of the screening process in design work can be completed on a computer, greatly reducing the number of candidate molecules that truly need to enter the experimental stage.

Cyclopeptide and Non-Natural Amino Acid Design Enters a New Era

Modern peptide drug discovery and development is no longer limited to natural linear peptides. An increasing number of drug candidates use cyclopeptide structures, non-natural amino acids, and complex chemical modification strategies. These structures can significantly improve stability, enhance pharmacokinetic properties, and strengthen target-binding ability. However, the design complexity of such molecules is far higher than that of traditional peptides. Different cyclization strategies change the overall conformation, and different non-natural amino acids affect hydrophobicity, charge distribution, and protease sensitivity. Traditional experience-based methods have difficulty simultaneously optimizing so many variables. AI models, by contrast, can consider multiple parameters at the same time and search for optimal solutions within a vast design space. Therefore, the combination of artificial intelligence and cyclopeptide engineering is becoming one of the most active directions in the current development of innovative peptide drugs.

From Candidate Molecules to Drug Development

It is worth noting that AI will not replace experiments. A truly successful R&D workflow is usually a combination of artificial intelligence and experimental validation. Researchers first use AI to generate and optimize candidate sequences, and then validate them through structure prediction, molecular simulation, and in vitro experiments. The experimental data are then fed back into the model, helping it continuously improve its predictive capability. This closed-loop model is forming a new R&D paradigm. In future peptide drug development, computers will no longer be merely auxiliary tools, but will become core components of the R&D process.

The Future of AI-Assisted Peptide Design

With the development of generative AI, structure prediction technologies, and high-performance computing, peptide design is undergoing a profound transformation. Researchers in the future may no longer need to search for active molecules among millions of random sequences, but instead directly define the target function, have artificial intelligence propose candidate structures, and then validate and optimize them through experiments. From antimicrobial peptides and antitumor peptides to protein-protein interaction inhibitors, and from linear peptides to cyclopeptides and non-natural amino acid systems, AI is continuously expanding the boundaries of peptide drug design. For the life sciences industry, this means not only improved R&D efficiency, but also that many targets once considered “difficult to drug” are gradually becoming designable, optimizable, and ultimately translatable into real drug candidates.

Artificial intelligence will not replace scientists, but it is becoming one of the most powerful design tools available to them. From sequence to structure, from structure to function, and then from function to drug candidate, peptide R&D is entering a new stage jointly driven by data, algorithms, and experiments.