Identification of Macrocyclic Peptide Families from Combinatorial Libraries Containing Noncanonical Amino Acids Using Cheminformatics and Bioinformatics Inspired Clustering

Abstract

In the past decade, macrocyclic peptides gained increasing interest as a new therapeutic modality to tackle intracellular and extracellular therapeutic targets that had been previously classified as “undruggable”. Several technological advances have made discovering macrocyclic peptides against these targets possible: 1) the inclusion of noncanonical amino acids (NCAAs) into mRNA display, 2) increased availability of next generation sequencing (NGS), and 3) improvements in rapid peptide synthesis platforms. This type of directed-evolution based screening can produce large numbers of potential hit sequences given that DNA sequencing is the functional output of this platform. The current standard for selecting hit peptides from these selections for downstream follow-up relies on the frequency counting and sorting of unique peptide sequences which can result in the generation of false negatives due to technical reasons including low translation efficiency or other experimental factors. To overcome our inability to detect weakly enriched peptide sequences among our large data sets, we wanted to develop a clustering method that would enable the identification of peptide families. Unfortunately, utilizing traditional clustering algorithms, such as ClustalW, is not possible for this technology due to the incorporation of NCAAs in these libraries. Therefore, we developed a new atomistic clustering method with a Pairwise Aligned Peptide (PAP) chemical similarity metric to perform sequence alignments and identify macrocyclic peptide families. With this method, low enriched peptides, including isolated sequences (singletons), can now be clustered into families providing a comprehensive analysis of NGS data resulting from macrocycle discovery selections. Additionally, upon identification of a hit peptide with the desired activity, this clustering algorithm can be used to identify derivatives from the initial data set for structure–activity relationship (SAR) analysis without requiring additional selection experiments.

Introduction

As drug discovery transitions to a new era of developing therapeutics for challenging targets previously thought to be “undruggable”, investments in the discovery and development of new therapeutic modalities are critical in this endeavor. Macrocyclic peptides represent one such modality where recent technological advances have made their discovery and optimization possible for a broad range of intracellular and extracellular therapeutically relevant targets.¹⁻³ Due to their increased molecular weight over traditional small molecules, macrocyclic peptides engage intracellular and extracellular targets by binding to shallow grooves and clefts with incredibly high potency and specificity, similar to antibodies. Additionally, the incorporation of noncanonical amino acids (NCAAs) has significantly increased the probability of identifying macrocyclic peptides with better drug-like properties, such as stability, solubility, and in vivo half-life, broadening their use as a therapeutic modality.^4,5

While it remains challenging to develop a macrocyclic peptide into a therapeutic, many technological advances have been developed to allow the robust discovery of macrocyclic peptides as early stage leads against any therapeutic target of interest. One such technology is mRNA display⁶ in combination with the RAPiD platform⁷ that identifies high affinity peptide binders that contain NCAAs (Figure 1). In this technology, tRNAs are acylated in vitro with an NCAA of choice and incorporated into macrocyclic peptide libraries using in vitro translation resulting in libraries containing approximately 10¹⁴ molecules. After library translation, peptides are covalently attached to their cognate mRNA through the use of a puromycin-nucleic acid linker, thereby ensuring that chemotype and genotype are coupled. A reverse transcription reaction is performed to generate a cDNA/mRNA complex allowing the PCR amplification of eluted peptides round after round. For each round of selection, target proteins are incubated with the in vitro translated peptide libraries, nonspecific binders are washed off, and the bound macrocyclic peptides are heat eluted and the cDNA is PCR amplified for the next round of in vitro translation and selection. Rounds are continued for each selection until a desired level of enrichment is achieved as compared to background, and the hit peptides are identified through the use of next generation sequencing (NGS).

Figure 1.

Workflow of macrocyclic peptide discovery using mRNA Display. To start, mRNA-Puromycin libraries are translated in vitro and subsequently reverse transcribed to make RNA/cDNA hybrids that are covalently attached to their cognate macrocyclic peptide. The target protein (dark red) is incubated with the peptide libraries in solution or on beads and free macrocyclic peptides are removed from the solution with a series of wash steps. The bound macrocyclic peptides are heat eluted, and the linked mRNA/cDNA is reamplified for subsequent rounds of selection until the library is enriched with peptides against the target protein. Next generation sequencing is performed on the elution fractions, and frequency counting of unique peptide sequences is performed to identify binders.

At present, counting peptide frequency round by round is the primary approach to analyze macrocyclic peptide selection NGS data sets and select hit peptides for synthesis. Although this method identifies the top enriched peptides to a target protein, it utilizes less than 1% of the total sequencing data and ignores families of peptides that may be weakly enriched individually while still having individual members that may be bona fide binders of interest. However, if these data sets are clustered by similarity, the resulting hit lists would represent relevant and diverse binders from families of varied enrichments. Unfortunately, traditional substitution matrices, such as Blocks Substitution Matrix⁸ (BLOSUM) and Point Accepted Mutation (PAM),⁹ currently used in sequence alignment algorithms for clustering protein sequences, such as ClustalW,¹⁰ are not suited for computing the similarity of peptides containing NCAAs as these matrices are derived from the mutation probabilities of natural amino acids in proteins to measure evolutionary distance.

In this work, we describe the development and validation of a new atomistic clustering method for cyclic peptides that contain NCAAs. We developed a new similarity metric, Pairwise Aligned Peptide (PAP) similarity, that uses dynamic programming¹¹ for sequence alignment and a chemical similarity matrix instead of a substitution matrix for similarity computation. The chemical similarity matrix contains the molecular structure similarities of pairs of amino acids, both natural and noncanonical, thereby expanding the ability to cluster peptides with building blocks beyond natural amino acids. In addition, the atom-level description is well-suited for describing differences relevant for protein–ligand interactions, such as stereochemistry and pharmacophoric features, thereby allowing structure–activity relationship (SAR) analysis to be probed from these initial NGS data sets. To validate our algorithm, we show that the new PAP similarity complies with the similarity property principle,¹² i.e., peptides with similar PAP similarity have similar activity. In a retrospective analysis of published peptides, with Major Histocompatibility Complex (MHC) activity data, we demonstrate that an active peptide can be used as a query to retrieve other active peptides with similar structures in the data set. We then present two prospective test cases: 1) using PAP similarity to identify novel macrocyclic peptides from a macrocycle discovery campaign against PSMD2; and, 2) using PAP similarity to probe SAR of known peptide binders from an existing NGS data set. In both cases, PAP similarity was successfully used in identifying novel chemical matter that was previously ignored with standard frequency analysis. Altogether, the new atomistic clustering method using PAP similarity for macrocycle discovery utilizing the RAPiD technology provides a robust platform for identifying and comprehensively characterizing novel macrocyclic peptides, thereby improving the speed and efficiency of discovering novel therapeutics for hard-to-drug targets.

Results

Development of PAP Similarity for Macrocyclic Peptides

To mine the results of the mRNA display selection experiments analyzed using NGS, we require (1) a peptide similarity metric able to recognize molecular structure and sequence differences of peptides of varying sizes containing amino acid building blocks that are noncanonical or have different stereochemistry (Figure 2a) and (2) a scalable clustering method amenable to processing tens of thousands of macrocyclic peptides and allowing for prioritization of the clusters.

Figure 2.

Details of the new clustering method. Lower case letters indicate a d-amino acid. (a) Short description of the algorithms. (b) Linear combination of similarity matrices encoding similarities of amino acids without and with consideration of the stereochemistry on the Cα atom. (c) Alignment of two peptide sequences and similarities of amino acid pairs; the optimal alignment requires one gap in each sequence. (d) Formula for computing the similarity of two peptides from the similarities of the paired amino acids. (e) Schematic display of the DISE clustering results where (+) denote cluster seeds and (■) the cluster members.

To meet the first requirement, we utilized the Needleman-Wunsch algorithm¹³ for sequence alignment but replaced the traditional amino acid substitution matrix with a custom matrix containing the chemical similarities of both natural and noncanonical amino acids which allows us to account for the molecular structure differences of amino acids relevant to molecular interactions. To identify the appropriate matrix to use in our application, three small molecule similarity matrices were compared: Atom–Atom Path¹⁴ (AAP) similarity, Tanimoto coefficients¹⁵ based on linear fingerprints¹⁶ (LFP), and Morgan fingerprints¹⁷ (AFP2). Fingerprint-based methods compare matching and nonmatching fragments, linear fragments for LFP, and circular fragments for AFP2. AAP similarity, on the other hand, compares the size of the common substructures of the two molecules.

Unfortunately, these three similarity metrics, like other small molecule chemical similarity metrics, do not take stereochemistry into account. Therefore, such metrics consider d and l amino acid isomers as equal. However, changing the stereochemistry of even a single amino acid can have a large influence on the 3D conformation of the peptide and its properties.¹⁸ To amend the lack of stereochemistry perception, we introduced a heuristic to differentiate d and l alpha amino acids. We substitute the Cα atom of d amino acids with a germanium (Ge) atom and l amino acids with a silicon (Si) atom. The resulting final matrix is created as an equally weighted linear combination of the similarity matrices generated with the nonmutated and with Ge/Si mutated amino acids (Figure 2b). It should be noted that the choice of Ge and Si was arbitrary as the sole purpose for their introduction is to render the Cα atoms of the d and l alpha amino acid differentially in our analysis.

To account for peptides with varying lengths, we used the Needleman-Wunsch algorithm to align peptide sequences through the introduction of gaps (Figure 2c). We then used each of our final amino acid matrices for scoring the alignment instead of the identity or BLOSUM matrix used for scoring protein alignment. To account for gaps, we chose a penalty of 0.25 for the opening of the gap and a penalty of 0.0625 for each extension of the gap based on following reasoning: The penalty of 0.25 is relatively high compared to the chemical similarity of two amino acids and reflects the fact that a gap has a strong influence on the 3D conformation of the macrocyclic peptides, while the small penalty of 0.0625 was chosen to not over penalize larger gaps. For each peptide pair, the aligned sequence pair with the highest score was picked from all sampled alignments for computation of the PAP similarity. We used a Tanimoto-like equation to compute the PAP similarity value by normalizing the sum of the similarities of the amino acid pair which allows us to compare the similarity of peptides with varying lengths together. The similarity of an amino acid paired to a gap is defined as 0 (Figure 2d). Overall, PAP similarity enables the alignment and similarity calculation of two peptides with varying lengths that may contain NCAAs or amino acids of differing stereochemistry.

For the clustering aspect of our application, we chose the Directed Sphere Exclusion¹⁴ (DISE) algorithm, which was developed for clustering small molecule hits from high throughput screening. In this case, DISE reads in the NGS sequencing files from selections against a given target, orders the peptides by their frequency counts, and utilizes PAP similarity with a similarity cutoff to build clusters of related peptides (Figure 2e). The ordering ensures that peptides with high sequence counts have greater probability of becoming cluster seeds. The highest frequency peptide is chosen as the first seed sequence. The second peptide is compared to the seed. If their similarity falls below the cutoff value, the second peptide becomes a new seed, otherwise it is set aside because it is within the sphere of the first seed peptide. Subsequent peptides are compared to all previously selected seeds. Finally, the peptides that were set aside are assigned to the cluster with the most similar seed. The total number of clusters is directly related to the diversity of the data set and the cutoff value chosen by the user. This means the number of clusters can get quite large for massive data sets. To account for this, a maximum number of clusters may be set up-front such that once the maximum number of clusters is achieved, any peptide that does not pass the similarity cutoff will be stored in a reserve list. Generally, we have found that similarity cutoff values between 0.25 and 0.45 allow all peptides to be clustered within 300 clusters for a typical screening set when we used the amino acid similarity matrix generated with AAP similarity metric.

Validation of PAP Similarity Matrices with MHC Peptide Binding Data

A good similarity metric for macrocyclic peptides must conform to the similarity property principle, i.e., peptides similar to an active peptide should have increased likelihood of being active themselves. To evaluate our three chemical similarity matrices and to determine if any of them meet the similarity property principle, we performed a retrospective analysis using 15 previously reported data sets of small linear peptides binding to MHC class I alleles (Table S1).¹⁹ For each data set, five active peptides were selected at random as queries with all other peptides considered as candidates. The similarity of each query to all candidates was computed using our AFP2_mix, AAP_mix, or LFP_mix amino acid similarity matrices. The candidates were then rank-ordered by similarity to the most similar query peptide. The Area Under the Curve (AUC) of the Receiver Operator Curve (ROC)¹⁹ was used as a metric to compare enrichment of active analogs retrieved by each similarity search (Table S2). Additionally, each search was repeated 50 times with different query peptides to gather statistical data on the performance of each similarity matrix.

The average AUC-ROC values across the 15 MHC assays are 0.84, 0.82, and 0.76 for the AFP2, AAP, and LFP-based amino acid similarity rankings, respectively (Figure 3a). In most cases, ranking by the AFP2-based similarity performed slightly better than the AAP-based similarity but within the error of the ranking. Ranking by the LFP-based similarity showed a significantly lower AUC-ROC value than the AFP2 or AAP-based similarities in all but two cases (HLA-B_4601 and Mamu-B_01). Given the minimal performance difference between AFP2 and AAP-based similarities, we decided to use the chemical similarity matrix generated with AAP similarity for computing PAP similarities moving forward. We chose this because the size of common substructures in a molecule has a large impact on calculated AAP similarity; therefore, it is more scaffold-biased. Altogether, PAP similarities rankings computed with any of the three chemical similarities resulted in significant enrichment of peptides active toward MHC I proteins indicating that PAP similarity adheres to the similarity property principle and is therefore a suitable similarity metrics for peptides.

Figure 3.

Result of retrospective analyses. (a) Median AUC-ROC by allele of 50 similarity search experiments for each allele as described in the method section. AUC-ROC values of 0.5 correspond to a random selection while values of 1.0 correspond to perfect enrichment. Error bars are standard deviations of 50 repeats. Numeric values and standard deviations are given in Table S2. (b–c) Result of one Mamu-B_01 similarity search using the AAP-based similarity. (b) ROC curve. (c) Activity of candidates versus similarity to the most similar query peptide. (d) Selected peptides (■ in plot c) aligned to their most similar query peptide 173.

Figure 3b–d shows the result of an exemplar similarity search within the Mamu-B_01 data set using the AAP similarity. This data set has the highest average AUC-ROC, and therefore has a steep ROC compared to, for example, HLA-A_0101 and HLA-A_6901 (Figure S1). In Figure 3c, there are a number of candidate peptides having PAP similarity greater than 0.5 to the most similar query peptides. Visual inspection of the candidate peptides aligned to the query peptide 173 shows that peptide 223 with the three mutated amino acids and a PAP similarity of 0.55 could still be considered a family member to the query peptide 173. The peptide alignment also shows that PAP similarity is not only driven by the number of substituted amino acids but also by the structural difference between the substituted amino acid and its substitution. For example: peptide 186 with two substitutions has a higher PAP similarity than peptide 175 having only one because the structural difference between phenylalanine and valine is much larger than between phenylalanine and tyrosine and between isoleucine and leucine.

Identification of Novel Peptide Macrocycle Binders to PSMD2 Using Clustering with PAP Similarity

We next tested if the PAP similarity algorithm could identify macrocyclic peptide hits from our library selections that were previously disregarded by the traditional frequency counting. For this study, we leveraged the selection data from our recently published macrocycle discovery campaign against PSMD2,²⁰ an essential subunit of the 26S proteasome complex, which is responsible for binding, unfolding, and degrading proteins tagged for removal from the cell. The PSMD2 macrocycle discovery campaign was performed using a library consisting of peptides having 10 to 14 amino acids with the variable positions encoding all natural amino acids except methionine. For this experiment, PSMD2 was incubated with the macrocyclic peptide library either preimmobilized to beads, or in solution before immobilization to maximize the number of accessible epitopes in the screen. After 6 rounds of selection, both selection experiments showed strong library enrichment, and NGS analysis with frequency counting identified multiple peptides as potential PSMD2 binders (Tables S3, S4). As previously reported, the top two peptides were synthesized from each selection, PSMD2_0000129056 (MC1) and PSMD2_0000000204 (MC3), and their binding affinity to PSMD2 was measured as 1.5 nM and 37 nM, respectively.

We reanalyzed our NGS data set that was used in the initial frequency counting analysis with our PAP similarity algorithm using four similarity cutoffs for the DISE clustering, 0.25, 0.3, 0.35, and 0.4, to understand how different cutoffs affect the ability to detect novel peptides (Figure 4). As expected, the number of clusters increased as we increased the cutoff value with the 0.4 cutoff hitting the maximum number of clusters allowed at 300 (Figure 4A). At the cutoff of 0.4, 338 “outlier” peptides were not assigned to the 300 clusters, suggesting that the lower cutoffs were better able to encapsulate all of the data from our selections. For the remaining three cutoff data sets, we calculated both the frequency contribution from all peptides (Sum of Frequency) and the number of unique peptides (Cluster Size) for each cluster and cutoff and plotted them together (Figure S2). These cluster characteristics allow us to determine the sequence diversity of each cluster as those that contain a high number of total peptides (Sum of Frequency) but a lower number of unique peptides (Cluster Size) represent hits that would have been identified in our initial frequency analysis which would not be novel. From this data, we observed that clusters containing the highest frequency peptides identified from the top 50 highest frequency peptides only represent a small subset of the actual data sequenced from our selections which contained 4,181 unique peptides from over 175,000 sequences.

Figure 4.

SeqSim algorithm enables rapid identification of new chemical matter through clustering. A) Number of clusters across different similarity cutoffs. B) Target binding and cluster distribution of hit picks with a similarity cutoff of 0.35 (top panel), 0.3 (middle panel), and 0.25 (bottom panel). Graphs represent PSMD2 binding as a function of the number of sequencing reads per cluster (Sum of reads, i.e. Sum of Frequency), and hits are colored by the number of individual members for each cluster (Count, i.e. Cluster Size). ELISA data represent the mean of two independent experiments with three technical replicates each. Clusters containing the two original hits MC1 and MC3 and top clustering hits are labeled next to their data point. C) Sequence enrichment profiles of previously published hits (top panel, Original Picks) compared to profiles of hits picked through clustering analysis (bottom panel, Clustering Hit Picks). Hits are colored according to PSMD2 binding measured by ELISA. Clustering hits with equivalent or better binding (2, 3, 5, 11; Table 1) than the published macrocyclic peptide MC3 are labeled.

We chose 19 peptides that represented clusters identified using all three cutoffs that either (a) had a large number of unique peptides, or (b) a large number of total peptides and ensured that any peptide chosen was not previously identified in the top 10 of our previous frequency analysis (Table 1). Each of the 19 peptides were in vitro translated individually and used in a binding ELISA assay with PSMD2 to compare their binding signal to the in vitro translated versions of MC1 and MC3. Excitingly, all but one peptide had a binding ELISA signal >2 signal over noise ratio (S/N) with half of the newly identified peptides exhibiting a S/N > 10, indicating true binders. Additionally, four PAP similarity-identified peptides, i.e., hit 2, 3, 5, and 11, with ELISA values of 33.5, 49.0, 41.6, and 44.2, respectively, (Table 1) showed ELISA S/N values >30 which is equivalent to the values observed from our previous hits MC1 and MC3 identified by frequency counting. The diversity of these sequences varies in both length and chemical property space, showcasing that our PAP similarity and clustering algorithms are able to identify novel binders that were previously buried deep in our NGS data sets when looking solely at round over round frequency calculations (Table 1).

Table 1. Sequences Selected for Binding Studies after the Clustering Analysis and Compared to Two Previously Identified Molecules, MC1 and MC3^a.

^aDisplay information including sequence frequency across rounds of selection, macrocyclic peptide sequence, and PSMD2 binding ELISA signal/noise are shown. ELISA data represent the mean of two independent experiments with three technical replicates each. Amino acids are colored by property: red = aromatic, purple = basic, blue = polar, green = acidic, orange = aliphatic, yellow = cysteine. ClAc = Chlororacetyl-l-phenylalanine.

As each representative peptide was clustered independently at each cutoff value, we wanted to understand how the cutoff values affected the identification of these peptides with the hope of narrowing down the most relevant cutoff value for future analyses. We plotted the trajectory of each peptide by its representative cluster size and ELISA binding signal for each cutoff value (Figure 4B). Although there was no immediately discernible feature that suggested one cutoff may be more favorable than another, it was apparent that when the hits were plotted by frequency at each round, some of the best binders were detected at much earlier rounds with much lower frequencies than MC1 and MC3 (Figure 4C, Table 1). This suggests that the biggest impact our clustering algorithm provides is that it groups sequences from all of the rounds, rather than by filtering solely on frequency at the last round of selection.

Using PAP Similarity to Probe SAR of Peptide Families with Noncanonical Amino Acids

Since the discovery of macrocyclic peptides using mRNA display is based solely on binding kinetics and not functional activity, we wanted to see if we could use PAP similarity to probe the structure–activity relationships (SAR) of previously identified active peptides from our NGS data sets. While similar peptides are expected to have both similar activity and properties, understanding how activity and properties vary in macrocyclic peptide pairs with PAP similarity may yield new hits with different characteristics and inform hit-to-lead optimization strategies.

For this work, we turned to an unnamed active therapeutic program focused on identifying macrocyclic peptides that can inhibit enzymatic activity. We had previously performed selections against Target X using macrocyclic peptide libraries that contained a variety of natural and noncanonical amino acids including, N-a-Methyl-L-Norleucine (MeNle), Sarcosine (Sar), N-(2-Phenylethyl)-glycine (PEtG), and d-Phenylalanine (f). Using our previously described selection and frequency analysis methods, we identified and confirmed three lead peptides: MC831, a 10 amino acid all natural peptide; MC832, a 12 amino acid peptide with two MeNle amino acids; and MC835, a 10 amino acid macrocyclic peptide with an f amino acid. Because these peptides are still active in an ongoing therapeutic project, the parent sequences have not been revealed. However, this does not preclude the analysis and reporting of the use of this algorithm to identify derivatives with varying property and activity profiles from our NGS data sets.

We reprocessed our NGS data using our PAP similarity algorithm with a cutoff of 0.4 using the three parent peptides as individual clustering seeds. Similarities for all of the peptides in our data set were then compared to these three peptides. Interestingly, a very small number of the 54,595 total peptides sequenced showed a similarity greater than 0.4, i.e., 504 (0.92%), 187 (0.34%), and 4,464 (8.17%) for MC831, MC832, and MC835, respectively (Figure S3). For each parent peptide, we chose between 10 and 15 peptides with a similarity cutoff of at least 0.4 that displayed a range of properties, sequence length, and amino acid composition for further characterization (Figure 5). As this protein target is an enzyme, we used a single concentration enzymatic assay (undisclosed) in combination with the in vitro translation of our peptides to characterize each peptide’s inhibitory activity. While the total concentration of each peptide is unknown when peptides are in vitro translated, we assume that the translation efficiency will be similar to the parent peptide, allowing us to compare enzyme inhibition relative to the parent sequence.

Figure 5.

Correlation of enzyme inhibition and PAP similarity for reference compounds MC831 (top), MC832 (middle), and MC835 (bottom) and members of their clusters. A) Peptides more similar to the active references are likely to also inhibit the enzyme. Black dashed line indicates 0.6 similarity cutoff. Red circle highlights the parent peptide with PAP similarity of 1.0. B) Amino acid differences between MC831 (top), MC832 (middle), and MC835 (bottom) and their PAP similarity (PAPS) identified derivatives with PAPS, AlogP, and enzymatic inhibition reported for each in vitro translated peptide. ‘-’ indicates a derivative peptide with an inserted or deleted amino acid at that position relative to the parent sequence. Percent Inhibition (% Inhib.) is reported as a normalized value to a control compound.

Excitingly, for all three parent peptides, many of the derivatives identified by our method showed increased inhibitory activity compared to the parent molecule (Figure 5A). In all three cases, peptides with PAP similarities greater than 0.6 were likely to inhibit the target protein at a similar or better level as the parent peptide, while peptides with PAP similarities below 0.5 display varying degrees of activity loss. In the cases of both the MC831 and MC832 series, the drop from high to low inhibition is quite steep as peptides drop in similarity, whereas the MC835 series peptides showed a much shallower correlation between similarity and inhibition, most likely due to the lower activity of the MC835 parent peptide to the target protein.

This data also shows how PAP similarity can be used as a design tool for peptide optimization (Figure 5B). For instance, in the MC831 series, mutating MC831 with a single Ile improves its inhibition to 94% from 79% but also lowers AlogP by nearly two units, highlighting this position as a spot for future optimization. Conversely, the data generated from the MC835 series suggests that there is a lower potential to optimize this compound for better inhibition or property space with the limited activity improvements measured over two AlogP units. Additionally, PAP similarity can also highlight specific amino acid positions that are essential and thus may make direct interactors with our target. For instance, positions 2, 3, 6, and 7 in MC351 display a lower mutational frequency and diversity of replacement amino acids and in most cases, mutations cause a significant loss in inhibitory activity. Because PAP similarity can account for both NCAA incorporation and gaps, this method interrogates how varying ring size and incorporating NCAAs may modulate functional activity as seen in the MC832 and MC835 series data sets. Altogether, using PAP similarity to identify similar peptides from selection data sets will be important for the design of future affinity maturation and optimization strategies of macrocyclic peptides.

Discussion

In small molecule high throughput screens, initial hits are commonly clustered to characterize and understand screening results. Analysis of the molecules within their clusters enables an assessment of how property and activity are related by similar compounds and establishes an SAR relationship for that series against a given therapeutic target. Here, we describe an analogous approach for the selection data generated in our mRNA peptide macrocycle discovery pipeline that is uniquely suited to peptides with NCAAs. We specifically developed a new peptide similarity measure that combines sequence alignment and a new amino acid similarity metric that enables useful clustering for peptide discovery. With this algorithm, we were able to identify peptide binders to PSMD2 that were previously undetected using standard frequency analysis. Additionally, we were able to use this method to probe the SAR and property space of previously identified active macrocyclic peptides which contained NCAAs, highlighting the importance in developing a tool like this for directed evolution, peptide discovery technologies.

In Bioinformatics, modern clustering methods utilize a substitution matrix such as BLOSUM for scoring alignments. The values in these substitution matrices are derived from the relative frequencies of amino acids and their substitution probabilities in evolutionarily divergent protein sequences, and therefore they are not suited for measuring chemical similarity. Additionally, substitution matrices are restricted to natural amino acids, which prevents their use on data sets where NCAAs have been incorporated into library design. Using a chemical similarity matrix instead of a substitution matrix allows us to compute chemical similarities for peptides with any NCAA and account for the importance of stereochemistry in the function and structure of an identified peptide hit. Importantly, previously developed similarity metrics for small molecules were not useful for comparing peptides, as the common backbone atoms and large size of peptides combine to reduce the dynamic range of these similarity metrics.

One limitation of our current method is the use the “Cα atom mutation” approach to introduce stereochemistry awareness in PAP Similarity. Specifically, our approach does not consider the impact of stereocenters in amino acid R-groups, such as threonine, but we believe these stereocenters to not be as significant as the Cα atom in identifying clusters from naïve libraries, however, are worth being taken into consideration for structure–property relationship analysis. One such integration considering for future development is the Rapid Overlay of Chemical Structure Similarities (ROCS) for Multialignment Using Fast Fourier Transform (MAFFT) method.²¹ Since ROCS considers the shape of the molecule conformation, any stereocenters present in an amino acid will be taken into account.

With PAP Similarity, we can now analyze NGS data sets from mRNA macrocycle discovery efforts in full, rather than relying on frequency counting methods employed previously. Due to the fact that most directed evolution platforms, like mRNA Display, identify hit peptides solely by binding affinity and not functional activity, it is very important to mine the data of all selection rounds and identify as many potential hits as possible, since not all identified binders may be functional. Although the frequency of peptides observed does correlate with binding, there are other selection pressures, including translation efficiency, efficiency of NCAA incorporation, epitope competition, and other technical aspects that may preclude the enrichment of a peptide or peptide family, thereby rendering it unidentifiable with traditional frequency methods. Importantly, because our algorithm is able to read NGS data from multiple rounds of selection, peptides that might have been enriched in earlier selection rounds but were not detected in later rounds are now captured.

We also showed that PAP similarity can be used to assess the SAR and the property landscape of peptide leads using the NGS data set where they were initially identified. When paired with an activity assay, we showed that it was also possible to identify more active peptides that are similar in sequence to the initial hits. Additionally, these similarity data sets can help prioritize the optimization of one peptide scaffold over another based on the property space of the clustered families. These data sets can also be used as roadmaps for guiding subsequent affinity maturation or medicinal chemistry efforts to optimize hits into leads toward identifying a clinical candidate. Importantly, while this method was developed for the specific application of mRNA display peptide macrocycle discovery, this algorithm can be used for any amino acid based discovery technology including antibodies, miniproteins, and cysteine-knot peptides. Altogether, PAP similarity represents a more holistic approach to analyzing NGS data sets resulting from peptide discovery experiments with the goal of identifying all potential binders in a screen to a given target of interest.

Methods

MHC 1 Data Sets

The MHC1 database assembled by Peters et al.¹⁹ was processed into 15 data sets (Table S1) with small peptides classified into active and inactive based on the MHC allele. Details are given in the Supporting Information.

Generate Chemical Similarity Matrix

Command line programs from the Chemalot package²² were used to compute AAP similarity and AFP2 and LFP-based Tanimoto similarity of amino acids. The replacement of Cα atoms in amino acids with Ge and Si atoms was performed with chemical transformations, i.e., by substructure matching and subsequent replacement of the identified Cα atoms of d and l α-amino acid isomers, respectively. The chemical similarity matrices computed with nonmutated and mutated amino acid structures, Sim_i, j^{no – stereo} and Sim_i, j, are combined to generate the final “mix” matrix according to eq 1 with the weighting parameter c set to 0.5, in which i and j denote the two amino acids used for computation of the amino acid similarity.

1

Computing PAP Similarity

The Needleman–Wunsch algorithm in conjunction with a chemical similarity matrix is used for pairwise peptide alignment. The gap opening and extension penalty were set to 0.25 and 0.0625. The alignment with the highest score is used for computing the similarity of the given peptide pair by summing up the similarities of the paired amino acids and normalizing the resulting sum by the length (len) of the peptides using eq 2, where i and j denote paired amino acids in peptides A and B to yield the overall peptide similarity, i.e., PAP similarity. This formula assumes that the similarities of amino acids is between 0 and 1, which is true for all chemical similarities evaluated here.

2

Cluster Peptides

The DISE algorithm was used for clustering based on the PAP similarity. The procedure consists of (a) sorting the peptides by a property of choice, (b) compiling a list of cluster seeds using the Sphere Exclusion diverse subset selection algorithm, and (c) assigning the remaining peptides to the most similar cluster seed. The number of clusters depends on the chosen similarity cutoff. (Figure 2). For the prospective clustering experiments (cf. Results section) described in this paper the macrocyclic peptides were sorted by frequency, i.e., the enrichment count.

Comparing Similarity Matrices

We followed the procedure described by Riniker et al.²³ to validate the PAP similarity method and to compare the various similarity matrices. For each MHC data set, five active peptides were selected at random as queries. All other peptides were considered as candidates. The similarity of each query to all candidates was computed. The candidates were ranked based on the similarity to the most similar query peptide. The Area Under the Curve (AUC) of the Receiver Operator Curve (ROC) was used to compare enrichment of active analogues retrieved by each similarity search. Each search was repeated 50 times with different query peptides to gather statistical data on the performance of each similarity matrix.

Protein Purification

Recombinant PSMD2 was purified from BL21 gold (DE3) cells (Invitrogen) transformed with a pET 52b plasmid containing full-length human PSMD2 fused to an N-terminal His₆ tag and TEV site, or to N-terminal His₆-TEV-Avi tags. For biotinylated PSMD2, cells were transformed with both the N-terminal His₆-TEV-Avi-tagged PSMD2 plasmid and a plasmid containing untagged BirA enzyme with a chloramphenicol marker. 3 L of cells were cultured at 37 °C for 24 h in TB autoinduction media supplemented with carbenecillin for His₆-tagged PSMD2, and carbenecillin and chloramphenicol for Avi-tagged PSMD2. Cell pellets were lysed with BPER (Thermo Fisher) supplemented with 150 mM NaCl, 5% (v/v) glycerol, 25 mM Imidazole, 0.5 mM TCEP, and Complete EDTA-free protease inhibitor tablets (Roche). Lysates were clarified by centrifugation and incubated in batch with 5 mL Ni-NTA agarose (Qiagen). Resin was loaded onto columns, washed, and eluted with 250 mM imidazole. Eluates were then dialyzed into 50 mM HEPES 7.5, 100 mM NaCl, 25 mM imidazole, 5% (v/v) glycerol, and 0.5 mM TCEP with 1 μg·mL^–1 TEV protein to cleave the His₆ tag. Dialysates were then passed over another 5 mL of Ni-NTA agarose (Qiagen) to remove any remaining tagged protein. The cleaved protein was then passed over a MonoQ 5/50 (Cytiva) and eluted with a gradient of 0.05 to 1000 mM NaCl over 50CV. Peak fractions were concentrated in a 30K MWCO concentrator (EMD Milipore) and injected onto a Superdex 200 16/60 (Cytiva), and the peak was concentrated in a 30K MWCO concentrator (EMD Millipore).

In Vitro Translation of Macrocyclic Peptides

Briefly, each 20 μM oligo encoding a peptide sequence with a C-terminal FLAG tag was translated at 37 °C for 30 min in a natural amino acid in vitro translation system²⁴ with N-chloroacetyl-l-phenylalanine (ClAc-F) as the initiator amino acid (tRNA^fMet aminoacylated with ClAc-l-Phe). For our binding assays, translation was performed at a 5 μL scale. After the translation, the reaction was quenched with 17 mM EDTA. The product was subsequently reverse-transcribed using RNase H minus reverse transcriptase (Promega) at 42 °C for 30 min and buffer was exchanged for HBS-T: 25 mM HEPES–NaOH, 150 mM NaCl, 0.05% Tween-20.

Binding ELISA

Biotinylated PSMD2 was immobilized on streptavidin-coated plates (Nunc) by incubating 60 nM of protein solutions for 0.5 h at RT. After washing the plate, 1 μL of in vitro translated FLAG-tagged peptides were incubated with 50 μL HBS-T in the plate for 1 h. After washing by HBS-T (300 μL, 3 times), the plate was incubated with anti-Flag-HRP antibody (Monoclonal ANTI-FLAG M2-Peroxidase (HRP) antibody produced in mice, Sigma) for 0.5 h. Color development was achieved by adding TMB substrate (Sera Care, USA), and the reaction was stopped by adding an equal volume of TMB stop solution (Sera Care, USA). Absorbances were recorded at OD 450.

Peptide Synthesis

Thioether macrocyclic peptides were synthesized using standard Fmoc solid phase peptide synthesis (SPPS). After the peptide assembly was completed, the N-terminus was capped with chloroacetyl in solid-phase. The peptide was then released from the resin by the treatment with a trifluoroacetic acid (TFA) cocktail, followed by precipitation with diethyl ether. The obtained crude peptide was dissolved in DMSO, and triethylamine was added for intramolecular cyclization via formation of a thioether bond between the thiol of the cysteine and N-terminal chloroacetyl group. Upon completion of cyclization, the reaction was quenched with AcOH. The cyclized peptide was purified using standard reverse-phase high-performance liquid chromatography (HPLC) methods and characterized by liquid chromatography–mass spectrometry (LC-MS) (see Supporting Information for details).

Supporting Information Available

The Supporting Information is available free of charge at https://pubs.acs.org/doi/10.1021/acschembio.3c00159.

• Overview of MHC 1 data sets, and additional experimental tables and figures. (PDF)

• The chemical similarity matrices of d and l amino acids and MCH 1 data set files (ZIP)

Author Contributions

^† Equal Contribution

The authors declare the following competing financial interest(s): M-L.L, A.G., and C.N.C. are current employees of Genentech, Inc. and shareholders of Roche.