Next-generation peptide therapeutics: An AI-driven end-to-end design system for hypernatural amino acids

【Answer 1】 Training objective: minimize the negative log-likelihood (NLL) of target sequences. 3.3.2 Reinforcement Learning Optimization Framework Built upon REINVENT’s policy gradient framework, this approach couples a generator (Agent) with a multi-objective scoring function (Scoring Function) to enable property-directed iterative optimization: ・• Actor (Agent): A Transformer-based generator that samples peptide SMILES strings via a categorical distribution; ・• Prior: A static copy of the pre-trained model, used to compute the NLL regularization term to prevent the generator from excessively deviating from learned chemical syntax; ・• Scoring Function (Scoring Function): Aggregates multiple scoring components, combined via weighted arithmetic or geometric mean: ・• Diversity Filter (Diversity Filter): Penalizes repeated generation of highly similar scaffolds to prevent mode collapse. 3.4 Performance Evaluation Results 3.4.1 Generation Quality (400 test peptides × 1000 sampling iterations) ・• Near-zero task failure rate: Only one test case exhibited a 0.3% positional count error; ・• Consistent performance across four topological types (linear/head-to-tail cyclized/sidechain-to-tail cyclized/disulfide-bridged), with no topological bias; ・• For monocyclic structures, the model correctly perceives and reproduces ring-closing atoms within generated amino acids—demonstrating its ability to capture long-range chemical dependencies in peptides. 3.4.2 Chemical Space Exploration Capability Under diverse sampling modes (exploration mode), each test peptide on average yields: ・• ~1600 distinct SMILES strings → ~1400 canonicalized molecular formulas representing unique amino acids; ・• Coverage includes: training-set canonical amino acids (recalling known ones) + training-set ncAAs (recalling learned non-canonical amino acids) + novel ncAAs (side-chain innovations beyond the training set); ・• Newly generated ncAAs are distributed tightly adjacent to known ncAA regions in chemical space—indicating interpolation rather than extrapolation, aligning with expectations for controlled, chemically plausible exploration. 3.4.3 Multi-Parameter Optimization (MPO) Use Case Using a 9-mer pharmacophore sequence binding to an HIV replication pathway target protein as the query starting point, six scoring components were defined (maximum ring size, CAMSOL-PTM solubility, membrane permeability classifier, substructure alert filter, diversity filter, and ring topology constraint). A 1000-step reinforcement learning optimization was then executed: ・• Solubility and substructure alert scores converged to high-performance ranges within 400 steps; ・• Predicted membrane permeability probability rose steadily from an initial value of 0.13 to batch-averaged values of 0.40–0.45; some candidate molecules exceeded 0.75; ・• Chemical validity rate remained above 90% throughout the exploration phase and rose to over 95% during the exploitation phase. 3.5 Limitations ・• Generalizability to longer peptides remains insufficiently validated; current focus is limited to peptides ≤10 residues; ・• Complex topologies—including bicyclic structures and stapled peptides—are absent from the training set; ・• Synthetic feasibility constraints are not embedded during generation and require post-hoc evaluation (addressed in Paper IV); ・• Reinforcement learning gains are limited when very few modifiable sites are masked; substantial property leaps require numerous editable positions.

NCFlow: A novel strategy for designing peptides incorporating non-canonical amino acids

【Answer 2】 With the rapid advancement of protein engineering and peptide therapeutics, an increasing number of researchers recognize that the 'chemical space' constructed solely from the 20 canonical amino acids faces inherent limitations. In contrast, non-canonical amino acids (ncAAs) offer unprecedented conformational flexibility and chemical diversity, enabling functional exploration in higher-dimensional chemical space—including enhanced stability, optimized target affinity, incorporation of non-natural catalytic mechanisms, and even construction of entirely novel protein structural topologies. However, successfully integrating ncAAs into peptide or protein structures—and achieving functional design at the structural level—remains an unresolved challenge. Most mainstream protein structure prediction and design tools—such as AlphaFold, RosettaFold, or score-based generative models—are built upon a closed system of 'standard amino acids' and lack expressive power, training data, or computational support for ncAAs. Against this backdrop, Lee, Kim, and colleagues proposed a representative work in their paper, 'Design of peptides with non-canonical amino acids using flow matching.' They developed NCFlow—a structure-generation model centered on flow matching—and integrated molecular modeling, deep-learning-based scoring, and physical energy validation to build a systematic ncAA peptide design platform. This article provides a step-by-step interpretation of their work, clarifying its methodological framework, structural accuracy, application workflow, and case studies—thereby helping readers appreciate its significance in protein engineering and drug design. Why study non-canonical amino acids? Expanding the chemical space of protein design Proteins are composed of 20 standard canonical amino acids—the 'genetic alphabet' honed by natural evolution to confer extraordinary structural and functional diversity to living organisms. Yet, in applications such as drug design, enzyme engineering, and protein-based materials, the structural and chemical diversity offered by canonical amino acids is increasingly revealing its limitations. Non-canonical amino acids (ncAAs)—amino acids not encoded by the standard genetic code—offer a promising solution to this bottleneck. ncAAs can be incorporated into proteins through multiple routes, including synthetic biology approaches (e.g., genetic code expansion, translational system reprogramming) or chemical modifications (e.g., post-translational modifications, semi-synthetic strategies). Their potential includes: ・ Introducing non-natural functional groups to catalyze reactions unseen in nature; ・ Enhancing protein thermal stability, enzymatic activity, or solubility; ・ Optimizing binding affinity and selectivity—especially critical in peptide drug development; ・ Conferring novel conformational capabilities to proteins, such as irregular backbones or macrocyclic architectures.

Aralez Bio Creates Noncanonical Amino Acids

23 hours ago

Aralez Bio launched a startup leveraging Nobel laureate Frances Arnold's directed evolution technology to synthesize noncanonical amino acids for pharmaceuticals. The San Leandro-based company secured over $20 million in funding to commercialize its enzyme-based synthesis process.

cen

2026-10-12

Aralez Bio Creates Noncanonical Amino Acids

Aralez Bio Creates Noncanonical Amino Acids

FAQ