How Can AI Optimize an Existing Peptide?

Many peptide projects do not start from scratch. A team may already have obtained a parent peptide from the literature, phage display, mRNA display, natural peptides, antimicrobial peptide screening, protein interaction interfaces, or internal experiments. It may already show reproducible activity, or it may simply be a hit worthy of further investigation. At this point, the most common question is not “how many new sequences can still be generated,” but “how can we improve the properties that truly limit project progress while preserving the existing evidence.”

A parent peptide usually satisfies only some requirements: it may have good activity but degrade rapidly in serum, high affinity but poor solubility, increased cellular uptake but also increased hemolysis, improved stability but become difficult to synthesize, or lose its original binding after the introduction of noncanonical amino acids. Existing activity does not mean optimization is complete; peptide optimization is essentially a multi-objective problem that requires evidence-based choices among mutually constraining goals.

The value of AI, therefore, is not to “regenerate a better sequence” without constraints, but to conduct controlled, multi-objective, hypothesis-driven optimization around a validated parent peptide, propose a set of candidates with different trade-offs, and then let synthesis and experiments determine which changes are truly effective.

Define the Problem Before Starting Optimization

“Optimize this peptide” is not a sufficiently well-defined task. The project first needs to determine the current bottleneck and the acceptable range of changes: Is the primary goal to improve activity, extend serum stability, or reduce aggregation and hemolysis? Which residues are known to participate in binding and therefore should not be changed casually? Is there already experimental evidence for the termini, cyclization anchors, or key charge patterns? Are D-amino acids, N-methylation, or other noncanonical amino acids allowed? What delivery timeline, purity, and cost must the candidates meet?

These boundaries determine what the model should search for. If the parent peptide already has reliable activity, local point mutations and small combinatorial libraries are often easier to interpret than completely regenerating the sequence, and they are also more conducive to establishing clear structure–activity relationships. AI can evaluate a larger local combinatorial space, but the outputs should still correspond to specific hypotheses, such as “reducing a protease-sensitive site,” “disrupting a continuous hydrophobic segment,” or “adjusting surface charge without changing hotspot residues.”

Activity and Affinity: Narrow the Space, Rather Than Inventing a Precise Kd

Optimization around activity and affinity can include residue substitution, preservation of activity motifs, protection of binding hotspots, local sequence exploration, and structure-aware mutation when reliable structural information is available. Sequence models, structure models, and scoring methods can help exclude clearly unreasonable variants and focus experimental resources on more informative candidates.

However, model scores cannot be treated as accurate Kd or functional readouts. Affinity is affected by conformation, experimental system, buffer conditions, and kinetic mechanisms, and biological activity may also include cellular entry, degradation, and downstream effects. A more robust use is to compare relative trends, formulate testable mutation hypotheses, and confirm them through binding or functional assays.

Selectivity Is Not a Secondary Metric of Affinity

Some projects do not lack binding to the target, but instead need to improve discrimination between the target and off-targets. In this context, blindly strengthening hydrophobic contacts or electrostatic attraction may simultaneously increase nonspecific binding. Optimization can adjust local charge, spatial complementarity, hydrophobic contacts, and adjacent sequence environment while preserving key binding residues, and, where possible, evaluate both the target and relevant non-target systems in parallel.

Selectivity is highly dependent on the specific target and experimental conditions. AI can help construct paired comparisons and prioritization, but selectivity cannot be proven solely by a single sequence score. Candidate design should be paired with a defined counter-screen or off-target assay; otherwise, a “higher score” may simply indicate stronger general interaction.

Protease and Serum Stability: Every Form of Protection Has a Cost

Common approaches to improving proteolytic stability or serum stability include local sequence redesign, D-amino acid substitution, N-methylation, terminal protection, cyclization, steric shielding, and incorporation of noncanonical amino acids. Models can integrate known cleavage motifs, sequence context, and physicochemical descriptors to help identify the sites and strategies that should be experimentally prioritized.

Stability strategies are clearly position-dependent. D-amino acids may reduce the recognition of the local backbone by certain proteases, but they may also alter side-chain orientation, secondary structure, or binding geometry; N-methylation may reduce backbone hydrogen-bond donors and change protease sensitivity, while also affecting conformation and receptor interactions. Improvements in stability may also come at the expense of affinity, solubility, or synthetic complexity. Therefore, prediction is more suitable for candidate ranking, while the actual half-life still needs to be measured in protease or serum experiments matched to the intended application.

Solubility and Aggregation: You Cannot Rely Solely on Increasing Positive Charge

An active peptide does not necessarily have good formulation and experimental behavior. Hydrophobicity, net charge, charge distribution, aggregation of aromatic residues, continuous hydrophobic surfaces, amphiphilicity, and self-association propensity all affect reconstitution, storage, nonspecific adsorption, and aggregation. AI can link these descriptors with sequence variants to identify changes that reduce risk while preserving the activity motif.

Simply increasing Lys or Arg is not always the best solution. Increased positive charge may improve apparent solubility in the aqueous phase or membrane interactions, but it may also enhance nonspecific binding, cytotoxicity, and hemolysis, and change purification behavior. Similarly, reducing hydrophobicity may improve solubility but weaken an originally important binding interface. Rational optimization should compare multiple strategies with different mechanisms rather than pushing a single descriptor to an extreme.

Permeability and Cellular Entry Must Be Evaluated Together With Safety

Peptide membrane interactions and cellular entry are jointly influenced by charge, hydrophobicity, conformational flexibility, side-chain arrangement, cyclization, N-methylation, and exposure of backbone hydrogen bonds. Appropriate conformational constraints or polarity shielding can sometimes improve permeability, but the specific result depends on the sequence, cell type, and detection method; cyclization or N-methylation does not guarantee effective intracellular delivery.

Stronger membrane binding also does not necessarily mean more valuable cellular entry. Increasing cationic charge or hydrophobicity may simultaneously increase membrane disruption, hemolysis, cytotoxicity, and nonspecific adsorption. Therefore, penetration cannot be optimized in isolation from hemolysis, cytotoxicity, and targeted activity, and prediction results cannot replace PAMPA, Caco-2, cellular uptake, or the corresponding functional experiments.

Toxicity and Hemolysis Prediction Is Suitable for Early Risk Screening

Sequence predictors, known motif analysis, physicochemical descriptors, and model ensembles can help identify high-risk candidates. For example, in antimicrobial peptide or cell-penetrating peptide projects, the relationships among positive charge, hydrophobic moment, aromatic clusters, and known hemolytic tendencies can be compared, and variants in which multiple risk indicators rise simultaneously can be deprioritized first.

These outputs are tools for candidate ranking and early risk filtering, not conclusions about clinical safety. The experimental systems, concentrations, and cell or erythrocyte sources in the training data may differ, and out-of-model sequences and complex modifications further increase uncertainty. A prediction score is not equivalent to experimental toxicity; toxicity, hemolysis, and functional experiments consistent with the intended application are still ultimately required.

Local Mutations of Natural Amino Acids Remain an Important Starting Point

AI optimization does not necessarily require introducing complex modifications, nor does it necessarily require changing the parent peptide beyond recognition. The chemical routes for local point mutations and small combinatorial libraries are more straightforward, the synthetic interpretation is clearer, and site-by-site SAR is easier to establish. For sequences already supported by experimental data, this is often the lowest-risk first-round strategy with a high information density.

During design, known hotspots and conserved sites can be fixed while only a few surrounding positions are explored; alternatively, substitutions with similar chemical properties or clear hypotheses can be defined for each position. This type of local optimization preserves the evidence base of the parent peptide, making both failures and successes easier to interpret, rather than jumping from an active sequence to an entirely new sequence that cannot be traced back.

D-Amino Acids and N-Methylation Require Site-by-Site Evaluation

D-amino acids can be used to explore protease resistance, local conformation, and side-chain orientation, but L→D substitution is not a formula that guarantees improved stability with unchanged activity. Chirality inversion may disrupt secondary structure, change the spatial positions of hotspot side chains, and may also affect chromatographic and analytical behavior. A more reasonable strategy is to avoid known key binding sites, start with suspected cleavage regions or structurally permissible positions, and retain unmodified controls.

N-methylation directly changes the backbone: it removes one hydrogen-bond donor, affects local cis–trans isomerization and conformational preferences, and may alter protease sensitivity, permeability, receptor binding, and synthetic difficulty. It is not an ordinary side-chain modification. Existing structural or structure–activity data can help select positions that do not rely on backbone NH to form key hydrogen bonds, but each strategy still requires actual synthesis and activity confirmation. For more background on this strategy, see The Important Role of N-Methyl Amino Acids in Peptide Drugs.

noncanonical amino acids: Expanding Chemical Space, Rather Than Simply Adding to the Alphabet

Noncanonical amino acids (NCAAs) can modulate hydrophobicity, charge, steric bulk, conformational preference, protease stability, binding interactions, and cyclization options. D-amino acids, N-methyl amino acids, Aib, Nal, Orn, and Dab represent different design purposes, not interchangeable “enhancement modules.”

When using NCAAs, the availability of protected building blocks, steric hindrance during coupling, deprotection compatibility, purification behavior, lead time, and cost must also be checked in parallel. Open-ended generation of a theoretical structure does not mean that the corresponding monomer or route is actually feasible. The Role of noncanonical amino acids in Drug Development introduces their design space, while specific projects still require peptide chemistry to constrain the candidate range.

Cyclization Should Stabilize the Active Conformation, Not Just “Turn the Peptide Into a Ring”

For a linear parent peptide, disulfide, head-to-tail, lactam, side-chain–side-chain, or side-chain–C-terminus cyclization can be explored. Rational cyclization may reduce conformational entropy loss upon binding, improve protease stability, and, in some systems, improve permeability; the prerequisite is that the ring-closing geometry can preserve or enrich the bioactive conformation.

Cyclization is not a universally effective improvement method. Improper anchor selection may misalign key side chains, reduce affinity, or restrict necessary conformational changes, and it also increases the complexity of protection strategies, cyclization yield, and byproduct separation. During design, multiple anchor points and linking modes should be compared, and linear controls should be retained. For structural value and route differences, see the Cyclic Peptide Design Guide.

Terminal and Linker Chemistry Is Also an Optimization Space

N-terminal acetylation and C-terminal amidation can change terminal charge and exopeptidase sensitivity; lipidation, linkers, or PEG-like modifications can be used to explore exposure, binding, or formulation properties. Their effects depend on the attachment position, length, and intended use, and they may also alter solubility, nonspecific binding, analytical methods, and in vivo behavior.

Therefore, terminal modification should address a defined bottleneck and be compared together with the unmodified parent and other sequence variants, rather than being used as a default “upgrade step.” Customer-specific payloads or linker groups also require advance confirmation of chemical compatibility and mass analysis methods.

Multi-Objective Optimization: Why There Should Not Be Only One Overall AI Score

The objectives in peptide optimization often conflict with one another. Increasing hydrophobicity may improve certain binding or membrane interactions, but it can reduce solubility and purification performance, and increase aggregation and hemolysis; increasing positive charge may enhance membrane activity, but may also increase nonspecific binding and cytotoxicity; increasing N-methylation may improve stability or permeability, but can alter conformation, reduce affinity, and increase coupling difficulty.

Compressing all properties into a single “AI overall score” hides these trade-offs. A more appropriate Pareto optimization approach preserves a set of options that are not simultaneously outperformed by other candidates across all key dimensions. For example, Candidate A is biased toward affinity, Candidate B toward stability, and Candidate C has better solubility and early safety characteristics. The project team can then select a differentiated validation set based on the application scenario and experimental budget, rather than assuming there is a universal best sequence that is unaffected by conditions.

Preserving the Knowledge and Lineage of the Parent Peptide

A parent peptide supported by existing experimental data contains valuable information. During optimization, active motifs, conserved residues, known binding sites, key charge arrangements, cyclization anchors, and validated modifications should be annotated, and boundaries should be set for positions where variation is allowed. Unless the project objective explicitly requires a change in mechanism of action, unconstrained generation should not be allowed to turn the parent into a completely different sequence.

Small-step optimization, iterative optimization, and lineage tracking can record which round each candidate came from, which positions were changed, and the corresponding experimental results. Retaining reliable “survivors” in each round and then propagating the next generation of variants around them can link sequence changes with property changes. This traceability is more conducive to establishing SAR and making development decisions than generating a large number of unrelated candidates in a single pass.

AI Optimization Must Not Turn a Good Peptide into an “Unsynthesizable Good Peptide”

Changes that improve predicted activity may simultaneously introduce consecutive hydrophobic segments, difficult couplings, non-procurable NCAAs, multiple N-methylation sites, or inefficient cyclization; even if the target product can be formed, deletion peptides, oxidized forms, or isomers may reduce crude purity and significantly increase purification difficulty. Optimization therefore must simultaneously consider protected building block availability, coupling difficulty, on-resin aggregation, cyclization route, crude purity, and expected isolated yield.

Synthesizability is not an administrative check after candidates are finalized, but part of multi-objective ranking. The previous article, “Why Must AI-Designed Peptides Still Undergo ‘Synthesizability’ Screening?”, further explains how risks related to SPPS, NCAAs, cyclization, and purification enter candidate prioritization. Connecting design with the actual synthesis team can avoid optimizing a valuable parent into a sequence that is too costly or temporarily impossible to validate.

Bring Experimental Data into the Next Round, Rather Than Infinitely Increasing Virtual Candidates

Real-world projects are better suited to an iterative closed loop:

Parent peptide → first round of controlled design → synthesis and QC → activity, stability, and risk experiments → identification of winning candidates → second round of local optimization → synthesis and experiments again → final candidate selection

First-round experiments should cover different design hypotheses, not just several highly similar top-scoring sequences. Actual results can reveal system features that the model does not include, such as the true contribution of a specific hotspot to activity, the effect of a certain class of modification on serum stability, or synthesis problems in a particular sequence segment. Continuous feedback from a small amount of high-quality experimental data usually improves the quality of next-round decision-making more than infinitely increasing virtual candidates.

What AI Can Do, and What It Cannot Replace

AI is suitable for generating controlled local variants, comparing multiple design objectives, identifying known risks, recommending modifications with clear hypotheses, and reducing the space that needs to be searched experimentally. It can also help teams maintain candidate diversity and avoid concentrating all resources on the same combination of properties.

AI cannot replace binding, stability, serum, toxicity, hemolysis, or permeability assays, nor can it replace actual synthesis, purification, and analytical characterization such as HPLC and LC-MS. Predicted affinity, half-life, toxicity, or permeability scores are not experimental results. The value of model output lies in helping decide “which experiments to do first and why,” not in announcing experimental conclusions in advance.

Recommended Workflow for Optimizing Existing Peptides

An executable design-to-synthesis workflow can be summarized as:

Parent peptide → define optimization objectives → annotate protected and key sites → generate controlled variants → evaluate sequence properties → conduct structural evaluation when necessary → synthesizability screening → multi-objective ranking → select diversified candidates → synthesis and QC → experimental testing → next round of redesign

Apollomics can carry out sequence optimization, physicochemical property balancing, D-amino acid substitution, N-methylation, NCAA design, cyclization, structural evaluation, and synthesizability review around existing candidates, and can connect these efforts with custom peptide synthesis, purification, and HPLC and LC-MS quality analysis. The service logic is Existing Peptide → AI Optimization → Chemistry Review → Synthesis → QC → Experimental Iteration; the final solution is jointly determined by project objectives and experimental feedback.

Conclusion

When using AI to optimize an existing peptide, the most important goal is not to make more changes to the sequence, but to ensure that every change corresponds to a clear hypothesis, while preserving interpretable trade-offs among activity, selectivity, stability, solubility, safety, and synthetic feasibility. A small-step, traceable, multi-round design–experiment cycle centered on the parent peptide can maximize the use of existing data and is also closer to real-world R&D decision-making.

A good optimization result is usually not a single sequence that claims to be optimal in all properties, but a set of candidates that can be synthesized, compared, and used to answer the next R&D questions.