Peptide libraries and design: how do screening and prediction work?

A peptide library is a collection of diverse peptide sequences prepared or presented in such a way that they can be systematically compared. Instead of analyzing a single sequence without a broader reference point, researchers can use a library to check which peptides bind to a specific molecular target, act as enzyme substrates, elicit a measurable response in a given assay, contain a specific motif recognized by other molecules, or exhibit another defined property.

Not all peptide libraries are the same. They can exist as physical collections of synthesized molecules, peptides displayed on biological carriers, peptides bound to beads or surfaces, as well as entirely virtual collections of sequences and molecular structures analyzed computationally.

Each of these formats answers slightly different scientific questions.

The peptide for which a computer predicted the possibility of binding to a specific protein is not the same as the peptide actually selected in a screening experiment. Similarly, a peptide enriched during screening does not automatically become a confirmed biological candidate.

Studies utilizing phage display, synthetic bead libraries, overlapping peptide sets, and computationally aided experiments demonstrate one of the fundamental principles of peptide discovery:

Finding an interesting signal starts the research process, it does not end it.

The strength of the final conclusions depends, among other things, on:

• design method and library quality;
• real, and not only theoretical diversity;
• test construction;
• appropriate controls;
• molecular identification;
• repeatability;
• exclusion of alternative explanations;
• independent experimental validation. [1] [2] [3]

Therefore, when analyzing peptide libraries, it is worth distinguishing several consecutive stages:

library construction → screening → hit identification → confirmation → characterization → prediction or optimization → independent validation.

Each of these stages provides a different level of evidence.

What is a peptide library?

Ordered set of distinct molecular variants

A peptide library contains related or diverse peptide sequences organized around a specific research question.

Some libraries are used to study minor variations within a known peptide. Others cover a much larger space of possible sequences.

Individual elements of the library may differ in terms of:

• peptide lengths;
• types of amino acids;
• positions of the individual residues;
• amino acid sequence;
• terminal group;
• cyclization;
• stereochemistry;
• chemical modifications;
• other intentionally modified structural features.

The library must also retain the ability to identify the element of interest.

In the case of a peptide array, the identity of the peptide can be linked to its physical location.

In a phage display library, the displayed peptide sequence is linked to the genetic information contained within the bacteriophage.

In a synthetic bead-based library, the peptide present on a selected bead may require chemical identification or the use of an additional encoding system.

The connection between the signal and the identity of the molecule is of fundamental importance.

A strong experimental signal from an unknown mixture does not yet reveal which molecule caused it.

The value of a peptide library therefore depends not only on the number of candidates it contains, but also on the possibility of linking:

candidate → identity → measured response.

Random, targeted, and systematic libraries

Different library constructs are used to answer different scientific questions.

A random or combinatorial library encompasses combinations of building blocks at selected positions.

The term „random” describes the planned method of construction or sampling. It does not mean that every theoretically possible sequence actually occurs with the same frequency.

A targeted library reduces variability around:

• well-known motif;
• stem peptide;
• of a specific structural skeleton;
• binding sequence;
• active region;
• hypotheses concerning a specific molecular target.

A library of systematic substitutions modifies selected residues in an orderly manner.

For example, an amino acid at one position can be successively replaced by several other amino acids to see if this position affects the studied property.

The library of overlapping peptides uses consecutive fragments of a longer parent sequence, with adjacent fragments sharing some common residues.

This approach is particularly useful for identifying the regions of a protein that are recognized in a specific test.

Every strategy has its advantages and limitations.

A broad combinatorial library can reveal an unexpected sequence motif.

A targeted library allows for a more detailed examination of a sequence of interest.

Systematic substitutions can indicate which items appear particularly important.

The library of overlapping fragments helps locate the active or recognized region of a larger protein.

Therefore:

A larger sequence space does not automatically mean better experimental data.

The appropriate type of library depends primarily on the research question.

Theoretical diversity does not imply actual representation

The space of possible peptide sequences is growing very rapidly.

If the library allows all 20 standard amino acids at each of the six positions, the theoretical number of possible sequences is:

20⁶ = 64,000,000 possible combinations.

This is a theoretical diversity.

This does not mean that the physical library actually contains 64 million different, correctly represented peptides.

Effective diversity can be reduced or distorted by, among other things:

• synthesis failures;
• uneven coupling efficiency;
• sequence-dependent synthetic difficulties;
• biological selection during library construction;
• limitations of the transformation;
• differences in reproduction;
• repeating sequences;
• workloads associated with sequencing;
• sampling limitations;
• The number of individual elements is not the same.

The proportion of the entire library that was actually used in a specific experiment also matters.

Even if a dataset theoretically contains a huge number of sequences, a single screening can cover only a portion of that diversity.

Deep sequencing of a commercial phage library revealed clear deviations from uniform representation, including differences in peptide abundance and sequence composition. The researchers also identified factors that reduce the effective library size. [4]

These results apply to the specific library analyzed and should not be automatically extrapolated to every phage display library.

However, they illustrate an important principle:

Theoretical diversity should not be treated as measured experimental diversity.

What is an overlapping peptide library?

A library of overlapping peptides represents a longer parent sequence as a series of shorter fragments.

Adjacent peptides share a portion of their sequence.

Let us consider hypothetical 15-residue fragments:

• remainder 1–15;
• remainder 6–20;
• the rest, 11–25.

Each subsequent peptide begins five positions further along than the previous one.

This means that the adjacent segments have ten common residues.

This arrangement helps identify regions associated with recognition or other measured responses.

If two adjacent peptides produce a signal, their shared sequence can help narrow down the area requiring further investigation.

However, this does not prove that:

• every remnant in the shared region is necessary;
• The common sequence alone is sufficient to achieve the desired effect;
• a minimal active sequence was identified;
• The fragment behaves exactly the same way as its corresponding region in the entire protein.

Further experiments may be needed to resolve these issues.

In a study on the outer membrane protein of meningococci, 45 overlapping synthetic peptides were used to analyze T-cell recognition in cells derived from human donors. [5]

The study demonstrated both the utility of empirical peptide mapping and the limitations of the predictive algorithms available at the time.

The observed responses remained specific to the applied set of peptides, test conditions, and samples derived from the tested donors.

Physical and virtual libraries answer different questions

The term „peptide library” can refer to very different experimental systems.

This distinction is very important:

A digital candidate is not automatically a synthesized molecule, and a synthesized molecule is not automatically an experimentally validated candidate.

Phage display and synthetic libraries

How does phage display link a peptide to its sequence?

In phage display technology, the peptide is presented as part of a protein located on the surface of a bacteriophage.

The phage contains both the genetic information encoding the presented sequence.

This creates a useful connection between:

phenotype — presented peptide

a

genotype — its corresponding genetic sequence.

Researchers can expose a phage display library to a specific target or test system, recover the selected particles, amplify or analyze them, and determine which sequences have been enriched.

Linking to genetic information significantly facilitates the identification of candidates compared to analyzing an unidentified molecular mixture.

Early studies utilizing random peptides displayed on filamentous phages demonstrated the possibility of enriching high-affinity antibody ligands and compared the behavior of phage-displayed peptides with binding results obtained for chemically synthesized peptides. [1]

This comparison was particularly important because the peptide displayed on the phage particle and its corresponding isolated synthetic peptide do not constitute an identical experimental presentation.

Enrichment is the result of selection, not a full explanation of the mechanism

A sequence whose proportion increases in successive selection stages is often referred to as enriched.

Enrichment is important information.

However, it does not automatically explain why a given sequence became more numerous.

Possible causes include:

• binding with an intended purpose;
• binding to the surface used in the test;
• interaction with another element of the experiment;
• better survival of selection conditions;
• greater infectivity;
• more effective multiplication;
• advantage during propagation;
• another characteristic dependent on the sequence or a specific clone.

These are scientifically different explanations.

One study identified a phage clone whose enrichment was associated with a replication advantage resulting from a genomic deletion. [6]

Another study analyzed the presented peptide, which bound to polystyrene instead of the intended target. [7]

These studies show two different mechanisms by which goal-unrelated enhancement can occur.

This does not mean that every enriched phage sequence is an artifact.

It shows, however, why:

Enrichment identifies a candidate requiring clarification; by itself, it does not prove specific recognition of a molecular target.

The method of presentation can influence the observed binding

The peptide displayed in the phage display system is physically linked to the carrier.

His behavior may therefore differ from that of the same molecule freely moving in solution.

Avidity is an important concept here.

Affinity typically refers to the strength of a single interaction within a given model.

Avidity, on the other hand, can result from multiple interactions occurring simultaneously.

If several copies of the presented peptide interact simultaneously with multiple binding sites, the total interaction may appear stronger than a single interaction of the free peptide.

Attaching to the medium can also:

• restrict the peptide conformation;
• change its availability;
• increase the local concentration;
• influence orientation;
• cause steric effects;
• place multiple copies of the peptide close to each other.

Therefore, a peptide successfully selected in a presentation system may behave differently after independent synthesis.

That is precisely why studying the isolated molecule is an important confirmatory step.

Comparing phage-displayed ligands to their chemically synthesized counterparts in early phage display studies clearly shows the importance of linking selection outcomes to independent molecular behavior. [1]

Synthetic libraries can cover a different chemical space

Synthetic peptide libraries are created by chemical methods rather than exclusively through biological expression.

Depending on the synthesis method used, they may potentially include:

• non-standard amino acids;
• selected D-amino acids;
• end modifications;
• cyclization;
• chemical tags;
• modified side chains;
• other synthetic building blocks.

This makes it possible to expand the chemical space available for research.

However, merely designing a sequence in a spreadsheet or computer program does not prove that the corresponding molecule can be easily and correctly obtained.

Practical problems may include:

• incomplete coupling;
• deletion sequences;
• aggregation;
• cleaning difficulties;
• poor solubility;
• side effects;
• instability of modifications;
• synthetic problems dependent on the sequence.

Therefore:

designed chemical diversity ≠ experimentally confirmed chemical diversity.

Genetically encoded systems of presentation and chemical synthesis therefore involve partially overlapping but non-identical molecular spaces.

Modified presentation technologies can additionally expand the range of chemistry available in biological systems, therefore it should not be simplified to state that biological libraries can only contain unmodified standard peptides.

One-bead–one-compound libraries

In the one-bead–one-compound (OBOC) library, each bead is intended to contain multiple copies of a single specific compound.

One of the frequently used strategies is the split-and-pool type of synthesis.

The balls are divided into groups, each group is reacted with another building block, then all the balls are combined again and divided before the next step of synthesis.

Repeating this process can lead to the creation of a large combinatorial library, while maintaining the planned connection between a single ball and a specific sequence.

The first studies regarding the one-bead–one-peptide system presented a method for creating and screening large synthetic peptide libraries for ligand binding.[2]

The term one bead–one compound, however, describes the principle of the organization of the library.

Does not prove that:

• each ball contains a chemically ideal peptide;
• each bead has an identical loading of the peptide;
• all coupling reactions occurred with the same efficiency;
• The peptide on each bead is equally accessible.

In a later study of individual beads, Raman spectroscopy was used to assess the amount of peptide.

Researchers observed significant variability between library beads, especially in the case of longer sequences, despite a more uniform loading of the initial material.[8]

This is important because the signal intensity during screening may depend on:

forces of interaction,

as well as from:

the amount of available peptide presented on the bead.

Mothers, individually synthesized kits and pooled studies

Peptide libraries can also be organized in the form of master libraries or individually synthesized sets.

In the peptide matrix, the defined peptides are located in known physical positions.

This provides a direct connection:

location → sequence identity → test signal.

However, it should be remembered that immobilization on the surface can affect the conformation and availability of the peptide.

Another solution is a staggered screening.

Several peptides are combined in one sample and tested together.

If the pool gives a positive response, it can then be deconvolutioned by testing smaller groups or individual elements.

Pooling can reduce the number of measurements needed at the initial stage.

Pula does not always behave exactly the same as each of its individual components.

The influence may be:

• differences in concentrations;
• competition;
• solubility;
• aggregation;
• interference;
• antagonistic effects;
• saturation of the test.

The positive result therefore indicates a group that requires further research.

It does not prove that every peptide contained in it has the property being studied.

Screening and selection of candidates

First, define what the test actually measures

The significance of the screening result depends entirely on the measured endpoint.

The binding assay answers the question of whether a peptide interacts with a specific target under given conditions.

Screening of enzymatic substrates examines whether the enzyme converts a specific peptide.

The inhibition test checks whether the peptide reduces a specific activity.

The cell test measures the defined response in the cells.

These results are not equivalent.

The peptide can bind to the target without activating it or inhibiting it.

It can be a substrate for an enzyme, but it does not exhibit proper behavior in the cellular system.

The cellular response, however, may result from an unintended pathway.

Therefore, the term „active peptide” is often too broad if it is not specified what kind of activity was actually measured.

In an iterative study using machine learning, peptide substrates for a specific class of enzymes were optimized through repeated experimental testing of selected candidates. [3]

The applications concerned the biochemical functions measured in this particular system.

They do not prove that the same sequences would work in other unrelated biological applications.

The controls help to separate the recognition of the target from the behavior of the test itself.

Appropriately selected controls help distinguish the intended molecular phenomenon from other possible explanations.

Depending on the test, this may be:

• surfaces without a target molecule;
• checks involving the media itself;
• non-related molecular targets;
• positive controls;
• negative controls;
• background measurements;
• competitive experiments;
• other controls appropriate to the specific system.

The surface without the target molecule may reveal non-specific binding to the test material.

Media checking helps detect responses caused by the presentation system itself.

An unrelated goal can provide information about selectivity.

The positive control shows whether the test is capable of detecting the known response.

A negative control helps determine the background level.

However, the controls should be selected based on the actual possible sources of error in the given experiment.

There is no single universal control peptide that would be biologically inert in every possible system.

The study of polystyrene binding phage shows why surface control is important. Researchers separately analyzed binding and multiplication and attributed the enrichment to the interaction with the test surface, rather than assuming that phage recovery proves the recognition of the intended target.[7]

Necessity, selectivity, and function are different properties

The three concepts that are often confused with each other are:

affinity, selectivity and function.

The causality concerns the strength of the interaction under specific conditions and within the appropriate interaction model.

Selectivity describes the extent to which a candidate preferentially acts on one goal compared to the corresponding alternatives.

The function concerns the biological or biochemical consequences of such an interaction.

The peptide can therefore exhibit:

• high affinity, but weak selectivity;
• high selectivity, but poor binding;
• strong binding without the expected functional effect;
• functional activity through an unexpected mechanism.

These distinctions determine which experiments should be performed next.

For example:

demonstrated binding ≠ demonstrated inhibition.

A peptide found in binding screening should not be automatically called an inhibitor if its inhibitory activity has not been directly measured.

Candidates are worth evaluating based on several types of data

Candidate selection is more reliable if it is not based solely on one ranking parameter.

Among other things, it is possible to consider:

• signal size;
• repetitiveness;
• the preservation of the background;
• the identity of the sequence;
• enrichment compared to the initial value;
• selectivity;
• the family of sequences;
• the possibility of independent synthesis;
• solubility;
• test repeatability;
• the possibility of further validation.

A sequential variety among the selected hits can also be useful.

The study of representatives of several different sequence families reduces the dependence on a single potentially misleading motif or clone.

In the case of phage libraries, it is particularly useful to compare the number of copies before and after selection.

A sequence that already appeared very frequently in the initial library can be represented in large numbers after screening without exceptionally strong goal-dependent enrichment.

A study of deep sequencing of a naive phage library showed an unequal initial representation [4].

The study on the advantage in reproduction further shows why population size alone should not be automatically interpreted as a measure of affinity. [6]

Overall, these results support the analysis:

how the given sequence was enriched,

and not only:

which sequence took first place after the selection process was completed.

The result from the screening is not yet a validated candidate

A hit, or a hit, is a candidate that meets the criteria of the initial screening.

This is a valuable result.

However, it does not complete the validation process.

The confirmation should answer, among other things, the following questions: whether the candidate's behavior:

• it is repeatable;
• it persists after appropriate controls have been applied;
• occurs during an individual examination;
• remains visible after re-synthesis;
• can be assigned to the intended molecule;
• remains significant in the appropriate secondary test.

Further characteristics may then include properties relevant to a specific research objective.

Re-synthesis is particularly useful if the original screening used a carrier, a bead, a presentation system, or a mixture.

Independent synthesis allows the candidate to be separated from the original way in which he or she presented themselves.

Confirmation of the identity and other important parameters of the sample further strengthens the relationship between the observed result and the intended molecule. At this point, proper interpretation becomes particularly important. Purity of peptides and laboratory tests, Because the HPLC result, mass confirmation and actual peptide content answer different analytical questions.

The correct order is therefore as follows:

hit in screening → confirmed hit → characterized candidate,

and not:

Target in screening → proven functional peptide.

A favorable ranking, a visually convincing docking model, or multiple times achieving a similar result in the same error-prone test do not automatically replace independent validation.

Sequence design and predictive tools

Design begins with defining the molecular goal

The design of a peptide can begin with:

• well-known motif;
• stem peptide;
• surface area of the proteins;
• hit obtained experimentally;
• the structural model;
• evolutionary sequence;
• a dataset of data regarding the tested peptides.

Then it is possible to analyze changes such as:

• amino acid substitutions;
• shortening of the sequence;
• lengthening of the sequence;
• cyclization;
• end modifications;
• change in the order of amino acids;
• other structural modifications.

The goal must be clearly defined.

For example:

prediction of binding

it is not the same task as:

prediction of recognition by an enzyme,

and this in turn is not the same as:

prediction of solubility.

A modification that improves one property may not affect another or may even have an adverse effect on it.

Practical peptide design therefore may require balancing several competing goals.

The experimentally assisted study of enzymatic substrates demonstrated a targeted approach in which an algorithm proposed candidates, biochemical experiments evaluated them, and the obtained results influenced subsequent decisions regarding the selection of the sequence.[3]

The calculations supported a specific experimental goal, and did not replace measurements.

The prediction based on the sequence depends on the training data

A computer model based on the sequence learns the relationships between the characteristics of peptides and the labels or results of measurements found in the training dataset.

Depending on the model, the result may be:

• classification;
• a result resembling probability;
• ranking value;
• predicted quantitative property;
• another type of computer estimation.

The interpretation depends on what the labels actually meant during the training session.

For example, the term „active” can mean:

• binding above a specified threshold;
• transformation by an enzyme;
• antimicrobial activity;
• cellular response;
• another criterion specific to a particular test.

These are not equivalent properties.

Data collections may also contain results obtained under different experimental conditions.

An untested peptide should not be automatically considered a negative result just because there is no positive result for it.

The evaluation of the model is equally important.

If almost identical sequences are found in both the training and test set, the predictive problem can be much easier than predicting the properties of truly new peptides.

Therefore:

The model's high overall accuracy does not in itself prove its ability to generalize to new sequences.

Generalization requires conscious evaluation

Generalization describes the model's ability to maintain usefulness beyond the examples used during training.

This is particularly important when the proposed peptide differs significantly from the training data in terms of:

• sequences;
• lengths;
• structural class;
• chemical modifications;
• molecular target;
• the experimental context.

In a study on the prediction of T-cell receptor binding, the effect of differences between training and test peptides on the evaluation of models was analyzed. [9]

The division of collections based on similarity of sequences and structures did not yield identical results.

This shows why a single split of training and test data or a single value describing the model's effectiveness may not fully reflect its behavior towards new peptides.

This does not mean that peptide predictive models are inherently unreliable.

However, it means that the way they are evaluated should resemble the actual application for which they are intended to be used.

What does docking of peptides predict?

Peptide docking is a computer method used to propose the way in which a peptide can interact with a receptor or other molecular target.

Calculations can be analyzed in various ways:

• conformations of the peptide;
• orientation;
• positions;
• contacts with the recipient;
• the geometry of the interaction.

Then the proposed systems are evaluated using an appropriate point or energy function.

The docking result can therefore provide:

• the potential binding site;
• the proposed remaining participants in the interaction;
• structural hypotheses;
• ranking of candidates;
• initial structures for further computer analyses.

However, it should be remembered:

Beyond docking is a hypothesis, not an experimentally observed complex.

Similarly:

The docking point score is not an experimentally measured constant dissociation rate, cellular activity, or the probability of biological success.

The interpretation depends on:

• point function;
• the method of sampling conformations;
• receptor structures;
• receptor flexibility;
• the flexibility of the peptide;
• the initial assumptions;
• protonation and molecular representation;
• calibration for a specific task.

Peptide conjugation can be particularly demanding because peptides are often very flexible.

During binding, their conformation and position relative to the target can change significantly.

Rosetta FlexPepDock: refinement and global docking are different procedures

The original publication regarding the FlexPepDock server describes the high-resolution refinement of the model starting from a peptide located approximately in the receptor binding region.[10]

Therefore, it should not be described as a universal method that allows discovering any unknown binding site without structural information.

PIPPER-FlexPepDock extends this concept towards global docking by combining the placement of peptide fragments with subsequent flexible refinement of the model. [11]

The publication evaluated this method on a limited but representative set of experimentally defined peptide–protein complexes.

Success in the benchmark shows effectiveness in the tested conditions.

It does not guarantee correct prediction for every new receptor–peptide pair.

The official documentation of Rosetta distinguishes between:

• refinement;
• ab initio modeling near the approximate binding site;
• the global PIPER-FlexPepDock workflow [12]

Therefore, the computer report becomes much more informative if it specifies:

which FlexPepDock protocol was used + what initial data were used + how the obtained structures were evaluated.

The statement alone:

„Rosetta docking has been completed”

it leaves many important methodological information unspecified.

Modified peptides require appropriate representation

The computer model must correctly represent the molecular structure under study.

A sequence recorded using standard single-letter amino acid codes may omit important chemical information.

These may include:

• C-terminus amidation;
• acetylation of the N-terminus;
• customized amino acids;
• D-amino acids;
• cyclization;
• covalent lateral bonds;
• phosphorylation;
• other modifications.

The study using FlexPepDock analyzed peptide–MHC modeling involving non-standard residues and selected peptide modifications.[13]

The authors obtained useful benchmark results and observed a relationship between the calculated results and the relative binding measured experimentally in the studied sample.

This supports the practical application of the modeling that has been tested.

However, this does not mean that every possible modification will be represented correctly automatically, nor that docking replaces experimental binding measurements.

CHARMM-GUI, peptide models and C-terminus amidation

CHARMM-GUI provides tools for preparing molecular systems for simulation.

Official documentation emphasizes that modified rest terms require appropriate structural definitions and parameters of the force field.[14]

Successful generation of input data for the simulation therefore means that a computer representation of the system has been prepared.

It is not experimentally confirmed:

• the identity of the peptide;
• bindings;
• biological function;
• stability;
• activity.

In the case of an amidated peptide, the terminal chemical group at the C-terminus is an amide rather than a free carboxylic group.

This difference can affect:

• formal cargo;
• hydrogen bonds;
• electrostatic effects;
• conformity;
• interactions with the surrounding molecular environment.

The structure, topology, and parameters of the force field used in the simulation should therefore represent the same final form that is being experimentally studied.

The study on H5 maxims combined experimental and simulation methods to assess the consequences of a change in the C-terminus form. [15]

Differences in the structure and interactions with membranes have been observed for this particular peptide.

This does not mean that amidation improves the properties of each peptide.

It does, however, show why the chemistry of terminal groups should not be overlooked in the molecular description.

Therefore, the search for amidation options in the peptide building program should begin with the chemical question:

What exactly form of the molecule is to be modeled?

Only after defining this issue does the choice of software settings become significant.

Experimental validation and reproducibility

Independently research the appropriate molecule

The selected candidate should ultimately be examined as defined molecular material.

Important characteristics may include:

• the exact sequence;
• the form of endings;
• specific modifications;
• molecular identity;
• other properties relevant to the experiment.

If the initial screening used phage display or peptides attached to beads, subsequent experiments can verify whether the isolated peptide retains the measured property beyond the original presentation system.

In an early phage display study, peptides presented on phages were compared with their chemically synthesized counterparts.[1]

This shows how an independent study can combine the results of library selection with the preservation of the peptide molecule itself.

The quality and characteristics of the sample also matter.

Differences regarding:

• the designated concentration;
• solubilities;
• purity;
• aggregation;
• final form;
• chemical modifications

they can influence experimental comparisons.

Two samples designated by the same peptide name do not have to be chemically equivalent if their corresponding molecular forms have not been defined. Therefore, before interpreting the experiment, it is useful to distinguish between the concepts of identity, purity, and actual amount of material, which are discussed in more detail in the article on Purity of peptides and laboratory studies.

An additional test has the greatest value when it answers a specific question.

The orthogonal test examines a specific property using a different measurement principle or a different method of experimental presentation.

Its greatest value is the ability to examine a specific alternative explanation.

A second test can, for example, help determine whether the initial result resulted from:

• bonding to surfaces;
• interference of the marker;
• carrier-dependent behavior;
• effects of immobilization;
• another artifact characteristic of the first test.

Just using two different models of devices does not yet mean that the measurements are orthogonal.

If both tests use the same immobilization method or the same detection principle, they may be susceptible to the same source of error.

The most important question is:

What alternative explanation does the second experiment verify?

Consistency between actually complementary methods usually provides stronger support than repeated repetition of the same error-prone reading.

Reproduction and biological context

Different types of replication allow us to assess different sources of uncertainty.

Technical repetitions help to assess repeatability within the same experimental setup.

Independent repetitions of the experiment can include additional variability between preparations or series of experiments.

Experiments using biological samples may also require consideration of differences between:

• donors;
• cellular preparations;
• organisms;
• tissues;
• other sources of biological material.

In a study using overlapping peptides, the recognition by T lymphocytes in samples from donors was analyzed and a library was used to identify the recognized regions.[5]

The observed responses depended on the biological context.

The results obtained for a single donor or a single experimental system should not be generalized automatically to every person or every biological system.

Repetitiveness therefore means the possibility of reproducing a specific statement under appropriately described conditions.

It does not just mean getting another chart with a similar appearance.

The documentation should allow the analysis to be reproduced.

A good study of the peptide library requires sufficient documentation so that another researcher can understand how the result was obtained.

Useful experimental data include:

• the exact sequence of the peptide;
• the length of the peptide;
• final groups;
• chemical modifications;
• the architecture of the library;
• sample identifiers;
• screening conditions;
• measured endpoint;
• inspections;
• the structure of repetitions;
• the way of determining the background;
• the qualifying threshold for hitting;
• criteria for selecting candidates;
• the analytical methods used.

In computer studies, the following may be important:

• name of the software;
• version;
• protocol;
• input structure;
• the way the molecule is represented;
• set of parameters;
• data processing;
• the method of dividing the training and test sets;
• the scoring method;
• the procedure for ranking candidates;
• selection criteria.

Maintaining only the highest-rated model, the best docking image, or the final sequence may prevent later reproduction of the method used to obtain the result.

The characteristics of the library itself also constitute an element of repeatability.

Studies of the composition of phage libraries and the loading of beads show that variability may exist even before screening begins.[4][8]

Documenting only the final hit may therefore omit important information regarding the initial material.

Positive predictions and negative results are equally important

A candidate predicted by computer that did not prove true experimentally can still provide useful information.

It may reveal:

• model limitations;
• a mechanistic false assumption;
• insufficient representation of the chemical property;
• problem with the hypothesis regarding the test;
• the region of the sequential space where the model weakly generalizes.

Such results should not be discarded from the assessment just because they did not confirm the original prediction.

Iterative design becomes more valuable when experimental observations—including unexpected and negative results—influence subsequent decisions.

The study on the design of enzymatic substrates used a general cycle:

prediction → experimental measurement → refinement of the model based on data. [3]

Its scientific value stemmed from the combination of calculations with observations, rather than from treating the first ranking as the final answer.

The strongest claim should therefore clearly state what has been demonstrated:

• enrichment;
• binding;
• the substance as the substrate of the enzyme;
• inhibition;
• cellular response;
• computer prediction;
• effectiveness in the structural benchmark.

Each of these results can be scientifically valuable.

However, they should not replace one another.

The most common errors in the interpretation of peptide libraries

The results of peptide library studies can be misleading if the conclusions go beyond what was actually measured.

One common mistake is treating the theoretical size of a library as experimentally confirmed diversity. Physical libraries can be characterized by unequal representation, missing elements, duplicates, or burdens resulting from the synthesis process.[4]

Another mistake is treating enrichment as evidence of kinship. Enrichment can also result from an advantage in reproduction, attachment to a surface, or other properties related to the selection process.[6][7]

The third problem is the conflation of binding with function. A peptide can interact with a target without eliciting the proposed biological effect.

Similarly, the docking result should not be treated as an experimental measure of affinity. The point results come from computer models and require validation before they can be considered quantitative evidence of binding.[10][11][13]

The virtual peptide is also not automatically a peptide that can be easily obtained in its correct chemical form. Synthesis, solubility, molecular identity, and sample characteristics remain experimental questions.

Finally, the result obtained for a peptide presented on a peg, immobilized on a surface, or bound to a ball may not be reproduced by the corresponding free molecule, because the very way in which the peptide is presented can influence its conformation, availability, local density, and avidity.

These limitations do not reduce the value of peptide libraries.

They, however, determine the way in which the evidence provided by them should be interpreted.

Frequently asked questions

1. What is the library of overlapping peptides?

The overlapping peptide library is a collection of fragments comprising a longer parent sequence, with adjacent peptides having a shared amino acid sequence. This approach helps researchers identify regions associated with recognition or other measured response. A study using overlapping fragments from a meningococcal protein showed the application of this method to map the responses of T lymphocytes in samples from donors.[5] However, a positive result for a fragment does not automatically identify the minimal epitope, does not prove that every common residue is necessary, and does not mean that the fragment accurately reproduces the structure and processing of the entire protein.

2. Can computer predictions replace experiments?

No. Computer predictions can help prioritize candidates, identify patterns, create structural hypotheses, and reduce the number of sequences requiring immediate experimental investigation. However, their reliability depends on the model, training data, the way molecules are represented, the assumptions made, and the specific task at hand. In the study of the design of enzymatic substrates, peptides selected by the algorithm were repeatedly tested, and the experimental results were used for subsequent rounds of selection. [3] This shows the collaboration between computational methods and experimentation, rather than replacing experimentation with a model. The candidate remains a computer prediction until its corresponding property is evaluated experimentally.

3. Is the peptide library a physical collection or a database?

It can be both. Physical libraries contain molecules that can be directly investigated experimentally, whereas virtual libraries contain digital sequences, molecular structures, or other computer representations. The first one-bead-one-peptide studies concerned physical synthetic libraries, whereas computer design methods analyze molecular representations before the synthesis of selected candidates. [2] [3] The mere presence of a sequence in a virtual library does not confirm that the corresponding peptide was properly synthesized, possesses the intended chemical form, or exhibits the predicted property. Therefore, when discussing the size of the library or the results of the screening, its format must always be specified.

4. Does a larger library always lead to finding better candidates?

No. The size of the library matters only when it contains a substantial, accessible, identifiable, and appropriately represented diversity. Deep sequencing of the phage library showed unequal sequence abundance and factors that reduce effective diversity despite a broad planned project.[4] Very large libraries can also increase sampling and interpretation problems because not every theoretically possible element must actually be present or properly examined. The quality of candidates therefore depends on the composition of the library, the screening design, the quality of the test, the control and validation, and not only on the advertised number of possible sequences.

5. Are random peptide libraries really free of biases?

Not always. The randomized design can still lead to unequal representation of nucleotides, amino acids, and entire sequences. Biological production, replication, synthesis chemistry, and sampling can introduce additional sources of bias. Deep sequencing of a commercial phage library showed unequal representation even before selection began relative to the target. [4] This does not mean that every random library is unreliable. It does, however, show that „random design” and „experimentally confirmed uniformity” are two different statements. The characteristics of the initial library help distinguish a large initial population from the actual selection-dependent enrichment.

6. How does phage display allow the identification of a peptide sequence?

Phage display combines the peptide presented on the surface of the bacteriophage with the genetic information contained within the same molecule. Selected phages can therefore be recovered, and the corresponding peptide sequences can be determined based on the associated genetic material. Early studies using random peptides presented on filamentous phages used this relationship between genotype and phenotype to recover and characterize binding ligands.[1] However, phage recovery alone does not prove specificity towards the target, as the number may be influenced by interactions with the surface, differences in replication, and other selective pressures. The genetic link allows the identification of the candidate, and further experiments help explain the reason for its selection.

7. Why can a fag be enriched without binding with the intended target?

Several mechanisms can lead to enrichment unrelated to the intended goal. Fag can interact with the test surface, exceptionally well tolerate selection conditions, or multiply more efficiently than competing clones. In separate studies, recovery associated with polystyrene has been demonstrated, as well as enrichment related to dominance in propagation. [6] [7] These are different mechanisms and require further control. Enrichment should therefore be interpreted as evidence that the clone survived and increased its representation during the selection process. Demonstrating specific recognition of the target requires additional evidence to distinguish this possibility from alternative explanations.

8. What does one-bead–one-compound mean?

One-bead–one-compound means that each bead is intended to represent multiple copies of a specific compound, allowing the selected bead to ultimately be linked to a single element in the library. This concept enables the creation of large combinatorial synthetic libraries and was presented in early one-bead–one-peptide studies.[2] However, the term describes the intended organization of the library rather than guaranteeing identical loading, perfect synthesis, or chemical homogeneity of each bead. Subsequent measurements of individual beads revealed significant differences in the amount of peptide present.[8] Quality of material and loading remain therefore characteristics requiring experimental evaluation.

9. Can the amount of peptide per bead affect the screening result?

Yes. A bead that presents a greater amount of available peptide can generate a stronger test signal even when the molecular interaction itself is not stronger. In a study using Raman spectroscopy to analyze libraries of one-bead-one-compound interactions, significant variability in peptide quantity was observed, especially in longer sequences.[8] The study does not specify the effect of loading in any possible screening system, but identifies the source of variability that can affect the signal intensity. Therefore, the results of the screening on the beads must be interpreted both in the context of the molecular interaction itself and the density of the material presented.

10. Is binding the same as biological activity?

No. Binding means the interaction of molecules under specific conditions. Biological or biochemical activity requires a defined endpoint. A peptide can bind to a protein but not activate or inhibit it, and the cellular response may arise through a mechanism other than the interaction assumed to occur. Phage display studies focused primarily on binding, whereas the machine learning-based study concerned the behavior of peptides as enzymatic substrates. [1] [3] These are different scientific questions. Therefore, the success of a screen should be described according to the property that was actually measured, without attributing a broader effect to it.

11. Why is it worth synthesizing and examining the hit again individually?

Independent synthesis helps determine whether a specific defined peptide actually reproduces the observed behavior outside the original library system. Attachment to the phage, the amount of material on the bead, the effects of the mixture, interactions with the carrier, or other screening elements can influence the initial result. In early phage display studies, the displayed ligands were compared with peptides obtained by chemical synthesis, illustrating this confirmatory step. [1] Re-synthesis is most informative when the obtained material is properly characterized and tested in a test related to the original claim. Synthesis itself does not yet confirm the function or mechanism.

12. What does Rosetta FlexPepDock provide?

The Rosetta FlexPepDock model simulates the structures of peptide–protein complexes, but the individual procedures associated with FlexPepDock assume a different range of initial information. The original server publication primarily describes the high-resolution refinement of a peptide model that is already approximately close to the receptor binding site. [10] PIPER-FlexPepDock uses a broader global docking strategy followed by flexible refinement of the model. [11] Benchmark results do not guarantee accuracy for every new peptide–protein pair. Therefore, a useful report should provide the protocol used, initial information, and the way the predicted structures are evaluated.

13. Does the favorable docking result confirm a strong binding?

No. The docking score is a value generated by a computer-based scoring function and is not automatically equivalent to experimentally measured affinity, such as a dissociation constant. Its significance depends on the docking protocol, molecular representation, the method of sampling conformations, the target being studied, and the calibration. In the FlexPepDock study of peptide–MHC complexes, the computational results were compared with relative binding measurements in a specific dataset, showing that such relationships need to be investigated rather than assumed. [13] A favorable docking score can help prioritize candidates or support a structural hypothesis, but strong binding requires appropriate experimental evidence.

14. Why is C-terminus amidation important in peptide modeling?

C-terminal amidation changes the chemical group located at the end of the peptide and can affect the formal charge, hydrogen bonds, electrostatic interactions, and the conformation of the molecule. Therefore, a computer model should therefore present the same chemistry of the end as that of the experimentally studied peptide. In the studies of H5 maximin, differences in structure and interactions with membranes associated with the change in the terminal form were observed. [15] This does not prove that amidation always improves the properties of peptides. It does, however, show why the chemistry of the terminal groups should be properly defined and modeled.

15. Does the generation of a model in CHARMM-GUI confirm the properties of the peptide?

No. CHARMM-GUI helps prepare molecular systems for simulation. The successful creation of the structure or input files means that a computer representation of the system under study has been created; it does not experimentally confirm binding, biological activity, stability, or clinical significance. Official documentation also indicates that modified residues require appropriate structural definitions and parameters.[14] Preparing a computer model is therefore the starting point for further modeling. Whether the actual molecule behaves in accordance with predictions remains a separate experimental question.

16. Why could the predictive model fail for a new peptide?

The new peptide may differ from the examples used during training in a way that the model did not learn to represent correctly. Differences may relate to sequences, structure, length, modifications, molecular purpose, or other characteristics. A study of models predicting the binding of T-lymphocyte receptors showed that the assessment of efficacy changed depending on how similarity in sequential and structural similarity between training and test peptides was taken into account.[9] This does not mean that every prediction of a new sequence will be incorrect. It does, however, show that the overall statistical performance of the model should be interpreted in the context of how closely the test conditions resemble its intended use.

17. What makes the results of peptide library studies repeatable?

Repeatability requires sufficient information to define the library, the candidate molecule, the experimental conditions, controls, the method of analysis, and the criteria for accepting a result as conclusive. Studies of the composition of phage libraries and the amount of peptide on beads show that variability may exist even before screening begins.[4][8] Therefore, providing only the final sequence and the point score ignores potentially important sources of variation. Repeatable studies require clear reporting of sequences, modifications, library architecture, sample identifiers, test endpoints, controls, analytical criteria, computer parameters, and independent confirmation of the main result.

18. Does the most enriched peptide have to be the best ligand?

No. High enrichment can result from binding to the target, but initial sequence abundance, effective replication, non-specific interactions with the surface, and other selection-related properties can also influence it. Deep sequencing has shown an uneven initial representation of phage libraries, and other studies have documented enrichment associated with dominance in replication.[4][6] Therefore, the most abundant final sequence should not be automatically equated with the ligand with the highest affinity. Stronger evidence is provided by comparing initial and final abundance, appropriate controls, and independent binding measurements.

19. Does the predicted peptide sequence indicate that it can be effectively synthesized?

No. The computer project generates a sequence or molecular hypothesis, but does not confirm the success of the synthesis. Some sequences may cause performance problems related to coupling, aggregation, purification, solubility, or chemical modifications. The virtual library may therefore contain candidates with very different practical synthetic feasibility. Only experimental material acquisition and its analytical characterization allow determining whether the intended molecule was indeed obtained. Computer prioritization and chemical feasibility are related but separate stages of research.

20. Can a peptide selected in phage display behave differently as a free molecule?

Yes. In phage display, the peptide is found in an environment associated with the carrier, which can influence its conformation, orientation, availability, local density, and avidity. The free peptide lacks any of these properties and therefore may exhibit other apparent behaviors during binding. In early phage display studies, ligands presented on phages were compared with their chemically synthesized counterparts, demonstrating the importance of this experimental distinction.[1] Successful phage selection thus provides valuable evidence regarding the presented sequence, whereas independent testing is necessary to determine the behavior of the corresponding free peptide.

Limitations of evidence derived from peptide libraries and predictions

When interpreting the results of peptide library studies, several limitations should be kept in mind.

First, theoretical diversity is not the same as real experimental diversity. Unequal representation, synthetic problems, differences in multiplication and the way of sampling can change the actual composition of the library.[4]

Second, enrichment during screening does not automatically confirm a specific link to the target. Interactions with the surface and the advantage in multiplication can lead to enrichment through different mechanisms.[6][7]

Thirdly, the way the peptides are presented matters. Peptides attached to phages, beads, or surfaces may behave differently than the same molecules in a solution due to availability, orientation, local density, or avidity.[1][8]

Fourthly, binding, selectivity and function are different properties. Demonstrating one of them does not automatically confirm the others.

Fifthly, the computer prediction depends on the model. Training data, the evaluation method, molecular representation, and the assumptions made determine the significance that can be attributed to the result.[3][9]

Sixthly, docking results are not experimental measures of affinity. Structural models and point functions can help in selecting hypotheses, but the actual behavior during binding requires experimental measurements. [10] [11] [13]

Seventhly, the chemical form matters. Terminal groups, non-standard amino acids, stereochemistry, and other modifications should be properly represented in both experimental descriptions and computer models.[13][14][15]

Finally, screening results are the beginning of validation, not the end of it. Independent material preparation, molecular characterization, controls, complementary tests, and replication gradually strengthen the available evidence.

Summary

Peptide libraries allow for the systematic study of the vast space of possible molecular sequences.

They allow you to compare multiple candidates and analyze issues related to:

• molecular recognition;
• substrate of enzymes;
• sequential motives;
• binding;
• dependencies structure–activity;
• cellular responses;
• by computer predictions.

The term „peptide library” however encompasses several fundamentally different research formats.

The library of overlapping peptides helps to map the regions of the larger sequence.

A targeted library of templates allows the meaning of individual items to be analyzed.

The phage display library combines the presented peptide with the genetic information and enables selection.

The one-bead–one-compound library organizes synthetic candidates in single beads.

The virtual library exists in digital form and can be used to prioritize sequences before starting experimental studies.

These formats should not be considered interchangeable.

The size of the library also requires a cautious interpretation.

The theoretical collection of millions of possible sequences does not prove that every element of it is physically present, uniformly represented, chemically correct, and actually screened. Studies using deep sequencing show why the actual composition of the library may differ from the planned diversity. [4]

Similarly, enrichment is not synonymous with a specific affinity for the goal.

Phage clones can be enriched as a result of mechanisms unrelated to the actual target, such as interaction with the surface or an advantage in replication.[6][7]

That is why proper checks and independent research are so important.

The same caution applies to the computer design of peptides.

Sequential models and docking tools can help identify promising candidates and create mechanistic hypotheses, but their predictions depend on training data, structural representations, point functions, and initial assumptions.[3][9][10][11]

The favorable docking result does not in itself confirm a strong experimental binding.

The computer-generated sequence does not prove that it can be correctly synthesized.

The predicted interaction does not confirm the biological function.

The most informative process of peptide discovery thus treats the different types of evidence as separate but mutually complementary stages:

design → library design → screening → hit identification → independent confirmation → characterization → computational refinement → experimental validation.

The strength of the final conclusion depends on how carefully these stages are interconnected.

The most important principle can be summed up simply:

The library indicates possibilities, the screening identifies signals, the model generates predictions, and the validation determines which conclusions are actually supported by experimental evidence.

Disclaimer

The article is purely educational and scientifically informative in nature. It describes general principles regarding peptide libraries, phage display technology, screening of synthetic peptides, sequence design, computer prediction, peptide docking, molecular modeling, experimental validation, and reproducibility of studies. It does not constitute medical advice, therapeutic recommendations, instructions for the use of peptides, information regarding dosage, or guidance regarding their administration to humans. Computer predictions, hits obtained during screening, docking results, and preclinical experimental results should not be interpreted as evidence that a given peptide is safe, effective, clinically validated, or suitable for human use.

References

1. Barrett, R. W., Cwirla, S. E., Ackerman, M. S., Olson, A. M., Peters, E. A., & Dower, W. J. (1992). Selective enrichment and characterization of high affinity ligands from collections of random peptides on filamentous phage. Analytical Biochemistry, 204(2), 357–364. https://doi.org/10.1016/0003-2697(92)90252-3
2. Lam, K. S., Salmon, S. E., Hersh, E. M., Hruby, V. J., Kazmierski, W. M., & Knapp, R. J. (1991). A new type of synthetic peptide library for identifying ligand-binding activity. Nature, 354(6348), 82–84. https://doi.org/10.1038/354082a0
3. Tallorin, L., Wang, J., Kim, W. E., Sahu, S., Kosa, N. M., Yang, P., Thompson, M., Gilson, M. K., Frazier, P. I., Burkart, M. D., & Gianneschi, N. C. (2018). Discovering de novo peptide substrates for enzymes using machine learning. Nature Communications, 9(1), Article 5253. https://doi.org/10.1038/s41467-018-07717-6
4. Sloth, A. B., Bakhshinejad, B., Jensen, M., Stavnsbjerg, C., Liisberg, M. B., Rossing, M., & Kjaer, A. (2022). Analysis of compositional bias in a commercial phage display peptide library by next-generation sequencing. Viruses, 14(11), Article 2402. https://doi.org/10.3390/v14112402
5. Wiertz, E. J., van Gaans-van den Brink, J. A., Gausepohl, H., Prochnicka-Chalufour, A., Hoogerhout, P., & Poolman, J. T. (1992). Identification of T cell epitopes occurring in a meningococcal class 1 outer membrane protein using overlapping peptides assembled with simultaneous multiple peptide synthesis. Journal of Experimental Medicine, 176(1), 79–88. https://doi.org/10.1084/jem.176.1.79
6. Kamstrup Sell, D., Sloth, A. B., Bakhshinejad, B., & Kjaer, A. (2022). A white plaque, associated with genomic deletion, derived from M13KE-based peptide library, is enriched in a target-unrelated manner during phage display biopanning due to propagation advantage. International Journal of Molecular Sciences, 23(6), Article 3308. https://doi.org/10.3390/ijms23063308
7. Bakhshinejad, B., & Sadeghizadeh, M. (2016). A polystyrene binding target-unrelated peptide isolated in the screening of phage display library. Analytical Biochemistry, 512, 120–128. https://doi.org/10.1016/j.ab.2016.08.013
8. Tang, Y., Thillier, Y., Liu, R., Li, X., Lam, K. S., & Gao, T. (2017). Quantification of peptide loading distribution using single-bead confocal Raman spectroscopy for one-bead one-compound library synthesis. Analytical Chemistry, 89(13), 7000–7008. https://doi.org/10.1021/acs.analchem.7b00516
9. Castorina, L. V., Grazioli, F., Machart, P., Mösch, A., & Errica, F. (2025). Assessing the generalization capabilities of TCR binding predictors via peptide distance analysis. PLOS ONE, 20(5), Article e0324011. https://doi.org/10.1371/journal.pone.0324011
10. London, N., Raveh, B., Cohen, E., Fathi, G., & Schueler-Furman, O. (2011). Rosetta FlexPepDock web server – high resolution modeling of peptide-protein interactions. Nucleic Acids Research, 39(Web Server), W249–W253. https://doi.org/10.1093/nar/gkr431
11. Alam, N., Goldstein, O., Xia, B., Porter, K. A., Kozakov, D., & Schueler-Furman, O. (2017). High-resolution global peptide-protein docking using fragment-based PIPER-FlexPepDock. PLOS Computational Biology, 13(12), Article e1005905. https://doi.org/10.1371/journal.pcbi.1005905
12. RosettaCommons. (n.d.). FlexPepDock. Retrieved September 20, 2026, from Official documentation
13. Bloodworth, N., Barbaro, N. R., Moretti, R., Harrison, D. G., & Meiler, J. (2022). Rosetta FlexPepDock to predict peptide-MHC binding: An approach for non-canonical amino acids. PLOS ONE, 17(12), Article e0275759. https://doi.org/10.1371/journal.pone.0275759
14. CHARMM-GUI. (n.d.). Frequently asked questions. Retrieved September 20, 2026, from Official documentation
15. Dennison, S. R., Mura, M., Harris, F., Morton, L. H. G., Zvelindovsky, A., & Phoenix, D. A. (2015). The role of C-terminal amidation in the membrane interactions of the anionic antimicrobial peptide, maximin H5. Biochemistry and Biophysical Acta, 1848(5), 1111–1118. https://doi.org/10.1016/j.bbamem.2015.01.014