Dataset for: Comparative study of nanopore phenylalanine clamp variants reveals unique peptide biosensing and classification properties

1 Comparative study of nanopore phenylalanine clamp variants reveals unique peptide biosensing and classification properties Jennifer M. Colby1 and Bryan A. Krantz2* 1Present address: Premier Biotech Labs, 723 Kasota Ave SE, Minneapolis, MN 55414, U.S.A. 2Department of Microbial Pathogenesis, School of Dentistry, University of Maryland, Baltimore, 650 W. Baltimore Street, Baltimore, MD 21201, U.S.A. * Corresponding author Bryan A. Krantz Department of Microbial Pathogenesis School of Dentistry, University of Maryland, Baltimore 650 W. Baltimore Street Baltimore, MD 21201, U.S.A (410) 706-1656 [email protected] Running title: Nanopore clamp variants as biosensors Preprint server: deposited at bioRxiv under doi: https://doi.org/10.1101/2025.08.22.671566 with a CC-BY 4.0 International license. Classification: Major: Biological Sciences, Minor: Biophysics and Computational Biology Keywords: peptide biosensor, machine learning, nanopore, anthrax toxin, protective antigen, electrophysiology, translocation 2 Abstract Rationally engineering biological nanopores is critical for advancing single-molecule biosensing. Here, we investigate the phenylalanine clamp active site (ϕ clamp) of the anthrax toxin protective antigen (PA) nanopore, a key site for molecular interaction, to test if engineering this site can improve peptide classification. We performed a comparative analysis of wild-type PA and two ϕ-clamp mutants (F427A, F427Y). We report the paradoxical finding that the F427A mutant—known to be a defective large protein translocase—is a superior peptide biosensor. Using a machine learning framework with engineered biophysical features, the F427A pore classifies a diverse peptide set with 93% accuracy. Our analysis suggests this enhanced performance arises because the F427A mutation, while weakening specific interactions, produces more consistent, lower-variance kinetic ‘fingerprints’ that are more easily distinguished by computational models. These findings establish a principle for biosensor design and enable a strategy where engineered pores with complementary specificities are deployed in multiplexed arrays for robust diagnostics. 3 Significance Statement Developing nanopore biosensors for accurate peptide detection is a major goal in biotechnology. Here, we engineered the ‘ϕ-clamp’ active site of the anthrax toxin nanopore to improve its performance as a sensor. We report a paradoxical discovery: a mutant pore (F427A) known to be a defective protein translocase is a far superior peptide biosensor, classifying peptides using machine learning with 93% accuracy. This counterintuitive result arises because the mutation produces weaker but more consistent electrical signals, creating a more reliable fingerprint for computational analysis. This work reveals a powerful new principle for rational biosensor design—that deliberately attenuating a natural biological function can enhance an engineered one—providing a clear strategy for creating next-generation diagnostic tools. 4 Introduction The precise identification and quantification of peptides and proteins are fundamental to a wide range of applications, from medical diagnostics to environmental monitoring. Peptide biomarkers hold immense promise for early detection of diseases, like heart disease, cancer, and infectious diseases (1-3). However, effectively analyzing complex and often dilute peptide mixtures demands analytical platforms with extraordinary sensitivity and specificity. While many conventional techniques struggle with these requirements, single-molecule nanopore biosensing offers a powerful, label-free alternative (4). This technology operates by detecting discrete changes in ionic current as individual molecules pass through a nanometer-scale pore, providing a real-time high-resolution readout of molecular properties. Beyond simple detection, this platform is a leading candidate for the ambitious goal of direct, high-throughput peptide and protein sequencing—an area of research that still lacks a widely applicable solution (5-7). At the heart of nanopore sensing are biological protein channels, such as anthrax toxin PA, which provide a highly controlled and tunable environment for molecular interactions (8-13). These pores, when embedded in a membrane, can capture the dynamic, picoamp-scale current fluctuations, as translocating molecules interact with the nanopore's internal features. The resulting ionic current signatures are rich in information, but their complexity presents a significant challenge, namely in extracting adequate analytical detail to enable reliable discrimination of diverse sets of peptide analytes. Critically, a single unmodified biosensor, regardless of its performance, may not possess the necessary discriminatory power to accurately classify the wide structural and chemical diversity found in real-world peptide mixtures. This suggests a need for a new paradigm in nanopore biosensor design. The anthrax toxin PA channel has emerged as a premier biological nanopore model system for polypeptide analysis (14). As a natural protein translocase, PA is exceptionally robust and facilitates the processive, voltageand/or proton gradient-driven transport of peptides (15, 16). It achieves this without the need for DNA tethers or other labels (17-20). PA uses multiple 5 internal loops and clefts, called ‘peptide-clamp’ sites, such as the phenylalanine clamp (ϕ clamp), to interact with the translocating molecule (9, 12, 13, 18, 20, 21). These interactions generate complex, multi-state ionic current signatures with distinct kinetic and conductance characteristics (20), which contain the information necessary for both peptide classification (22) and, in principle, sequencing. The inherent structural modularity of the PA nanopore makes it an ideal platform for protein engineering to fine-tune its biosensing properties. Analyzing the voluminous and often noisy data from these multi-state systems necessitates advanced computational methods. Machine learning (ML) and deep learning are particularly well-suited for this task, as they can identify subtle, high-dimensional patterns in translocation kinetics that are inaccessible to traditional analysis (22-28). However, the full potential of these computational tools has yet to be realized, especially when dealing with the challenge of analyzing complex peptide mixtures. In this work, we hypothesize here that a single unmodified sensor, optimized for certain types of peptides, may not perform as successfully across a wider array of different analytes. To address this, we are leveraging the PA nanopore as a model system, employing protein engineering of the ϕ-clamp site to generate a series of novel biosensors. We then use advanced ML methods to characterize the unique classification strengths of each variant. Our results demonstrate that an ensemble or multiplexed array of these engineered biosensors, with their complementary detection characteristics, can achieve a well-balanced performance, surmounting the performance of the wild-type (WT) pore. This ensemble approach represents a significant step forward for protein nanopore-based analysis in biotechnology, paving the way for more accurate peptide and protein sequencing capabilities. Results Nanopore variant study design. Our prior study leveraged the wellcharacterized anthrax toxin PA nanopore as a single-molecule peptide biosensor (22). The PA 6 nanopore's architecture includes a key narrow constriction site, known as the ϕ clamp, which is formed by a radially symmetric arrangement of phenylalanine residues (F427) (13). This site is positioned to make close contact with translocating proteins, effectively ‘clamping’ them and thereby promoting productive protein unfolding and translocation. Moreover, the interaction leads to full and partial blockade of the channel to ion flow in a highly dynamic and peptidespecific manner (20, 22). Our central hypothesis was that tuning the identity of the residue at this F427 position would alter the nanopore's selectivity and, consequently, its capabilities as a peptide classifier. To test this, we engineered two specific variants (i) F427A (replacing Phe with the much smaller Ala) and (ii) F427Y (replacing the site with the slightly bulkier, yet more hydrophilic Tyr). We then compared their classification performance to the WT PA nanopore. To evaluate our hypothesis, we used a previously characterized set of seven 10residue guest-host peptides of the sequence KKKKKXXSXX (where X is the guest residue) (17, 20, 22). The guest residues included non-aromatic side chains (Ala, Leu, Thr) and aromatic side chains (Phe, Trp, Tyr). A final variant, guest-host TrpDL, featured an alternating pattern of ᴅand ʟ-stereochemistry to probe how peptide backbone dynamics influence the translocation signal. Prior work has demonstrated that these peptides can be efficiently classified by the WT PA nanopore using ML with ~0.90 accuracy (22), and we used these results as a benchmark for the engineered F427A and F427Y nanopores. Data collection and preprocessing. In previous work, we collected WT PA nanopore translocation event streams for the guest-host peptide series. These experiments were performed under a constant driving force of 70 mV (cis positive), a condition known to strongly favor productive translocation events and produce robust current signals (22). We similarly collected single-channel peptide translocation recordings for the F427A and F427Y nanopore variants under the same 70 mV driving force, with data streams extending up to 30 minutes per peptide. The raw data from these experiments consisted of complex ionic current signals (Fig. 1). To prepare this data for ML, we developed a preprocessing pipeline. First, using a K-Means 7 clustering algorithm, the raw event streams were labeled according to their corresponding conductance states. While rare intermediate states were occasionally observed, our analysis identified four robust and well-populated states that consistently described most of the current signal. These states are enumerated 0 through 3, corresponding to a fully blocked (state 0), partially blocked (states 1 and 2), and fully open (state 3). This state labeling procedure was benchmarked against expert electrophysiology software, CLAMPFIT, and deemed to be highly consistent yet computationally much more efficient. These state-labeled records were then carefully segmented into individual translocation events. We settled on using only events greater than or equal to 15 ms in length in all our subsequent analyses to capture the most data while excluding lower-information short duration spiky events. Finally, a set of biophysical features, such as dwell time, current blockade level, and signal variance, were computed to describe each individual event (see Materials and Methods in SI Appendix for a complete list of features). These comprehensive feature sets, representing the unique biophysical signature of each peptide translocation event, were then used as the training data for our ML classification models. Unsupervised clustering analysis. To assess the intrinsic separability of the peptide features extracted from each nanopore variant, we performed an unsupervised clustering analysis using Uniform Manifold Approximation and Projection (UMAP) (Fig. 2). UMAP is a dimensionality reduction technique that is well-suited for visualizing the high-dimensional data, revealing the underlying structure and grouping of feature sets. To quantitatively evaluate the quality of the clustering, we calculated the Adjusted Rand Index (ARI) and Normalized Mutual Information (NMI) for each nanopore, comparing the UMAP-derived clusters against the ground truth peptide labels. In the analysis, using a 15 ms minimum event duration filter, WT PA nanopore had an ARI of 0.0648 and NMI of 0.0864; PA F427A had an ARI of 0.0536 and NMI of 0.0775; and PA F427Y had an ARI of 0.0947 and NMI of 0.1424. A higher ARI and NMI value indicates that the feature set of a given nanopore more effectively clusters the peptides into 8 groups that align with their true identities. Visual UMAP plots for the F427Y variant corroborated these metrics, showing less scatter and cleaner clusters compared to the F427A variant (Fig. 2). Interestingly, these results suggest that the features from the PA F427Y nanopore, which will be shown to have lower performance in supervised classification tests, demonstrate superior inherent clustering capabilities. This highlights a critical distinction between unsupervised and supervised learning tasks. While the features of the F427Y nanopore may create tight, welldefined clusters, these clusters may be positioned too closely to one another in the feature space, making it difficult for a supervised classifier to establish a clear decision boundary. Conversely, a nanopore with less tightly clustered data (like WT or F427A) may still have greater cluster separation, which is a more favorable condition for a supervised model to learn and generalize from. This observation underscores the fact that the ultimate metric of success in this study is not the inherent separability of the data, but the performance of the classification models built upon it. The following section will therefore focus on the supervised learning performance, where the true discriminative power of each nanopore variant is directly measured. Initial single-stage classification performance of nanopore variants. To establish a performance baseline, we first trained individual single-stage eXtreme Gradient Boosting (XGBoost) classifiers for each nanopore variant: WT, F427A, and F427Y (Table S1). This preliminary analysis revealed distinct performance profiles for each nanopore, suggesting a potential for synergistic gains through an ensemble approach. The WT nanopore achieved a mean macro-averaged F1-score of 0.8652 (±0.0087) (N=5), while the F427A variant demonstrated superior overall performance with a mean F1-score of 0.8869 (±0.0089). In contrast, the F427Y nanopore performed the lowest, with a mean F1-score of 0.7829 (±0.0088). A detailed, class-by-class analysis of the individual F1-scores revealed the specific strengths of each nanopore (Fig. 3). The F427A nanopore demonstrated a marked increase in predictive power for guest-host peptides containing small non-aromatic side chains. Specifically, 9 its F1-scores for guest-host Ala (0.95) and guest-host Thr (0.94) were substantially higher than those achieved by the WT nanopore (0.80 and 0.74, respectively) (Table S1). Conversely, the WT nanopore showed a slight advantage in classifying some aromatic guest-host peptides, such as guest-host TrpDL, with an F1-score of 0.96 compared to 0.91 for F427A. The WT nanopore also performed well on guest-host Phe, Trp and Tyr. Confusion matrices similarly show the high degree of guest-host Ala and Thr confusion for the WT nanopore relative to the F427A version (Fig. S1). These complementary strengths—F427A excelling with small nonaromatic side chains and WT performing somewhat better on aromatic ones—motivated the design of a more sophisticated, multistage ensemble classifier to leverage the unique information from each nanopore variant. Multistage ensemble classifier architecture. To effectively maximize the distinct single-molecule information from each nanopore variant, we designed a multistage ensemble classifier based on a probabilistic blending approach (Fig. 4A). In the first stage, individual XGBoost classifiers were trained for each nanopore variant (WT, F427A, or F427Y). These models were configured to output a probabilistic vector (softmax scores) indicating the confidence of an event being from an aromatic peptide class. In the second stage, a separate XGBoost model was trained on the augmented features, which now included the Stage 1 probabilities, to predict peptide-specific probabilities for both aromatic and non-aromatic peptides. Finally, a meta-classifier took as input the combined feature set from the raw data and the probabilistic outputs from both stages 1 and 2 to produce the final peptide predictions. For instance, in a two-nanopore combination (e.g., WT/F427A), the meta-classifier's input vector was a blend of all softmax scores from both nanopores from both stages. This architecture allowed the meta-classifier to optimally blend the confidence scores from the different nanopore classifiers, effectively identifying and amplifying the unique and complementary information provided by each variant to improve overall classification accuracy. 16 Table 1. Multistage ensemble probabilistic blending classifier performance1. Nanopore(s) Overall accuracy2 Macroaveraged F1-score2 F1-score per guest-host peptide class3 Ala Leu Phe Thr Trp TrpDL Tyr WT 0.8874 (±0.0094) 0.8870 (±0.0094) 0.81 0.95 0.93 0.77 0.95 0.98 0.91 F427A 0.9323 (±0.0041) 0.9322 (±0.0042) 0.96 0.96 0.89 0.97 0.96 0.94 0.89 F427Y 0.8365 (±0.0042) 0.8367 (±0.0047) 0.97 0.92 0.67 0.94 0.82 0.86 0.72 WT/F427A 0.9112 (±0.0019) 0.9112 (±0.0018) 0.90 0.96 0.90 0.88 0.95 0.93 0.88 WT/F427Y 0.8632 (±0.0060) 0.8632 (±0.0062) 0.86 0.93 0.83 0.83 0.89 0.92 0.84 F427A/F427Y 0.8909 (±0.0061) 0.8909 (±0.0061) 0.97 0.94 0.81 0.96 0.90 0.91 0.82 WT/F427A/F427Y 0.8848 (±0.0033) 0.8847 (±0.0032) 0.89 0.94 0.85 0.88 0.91 0.91 0.86 1Performance based on test set evaluation metrics of multistage ensemble probabilistic classification model. 2Overall accuracy and macro-averaged F1-score values are means and std. dev. (N=5). 3Best set of per-class F1-scores out of N=5 replicates. 17 Figures Fig. 1. Nanopore system and translocation event streams for F427X variant nanopores. (A) Sagittal section of the anthrax toxin PA7 nanopore cryo-EM structure (PDB: 3J9C) (8) rendered in Chimera (29) as a molecular surface. Peptide clamps and loop active sites are colored and labeled: α clamp (magenta), 397-loop (cyan), ϕ clamp (green), and charge clamp (red). Overall scale of the upper vestibule and elongated lower β barrel are indicated. Narrowest point, at the ϕ clamp, has a luminal diameter of 6 Å. Direction of translocation from cis to trans is indicated by an arrow. Membrane bilayer position is indicated with a solid gray rectangle. Representative recordings of guest-host peptide translocations carried out at 70 mV (cis positive) in symmetric succinate buffer, 100 mM KCl, pH 5.6 for (B) F427A and (C) F427Y nanopores. To the left are the standard three letter name for the guest residue. These nanopore-peptide systems populate multiple discrete partially or fully blocked intermediates (approximate locations indicated by dashed lines). From bottom to top of each record: fully blocked (state 0), partially blocked intermediates (state 1 and state 2), and fully open baseline (state 3). Scalebar at the upper right of each panel denotes 4 pA by 100 ms for guest-host Ala, Leu, Phe, Thr, and Tyr peptides. For guest-host Trp and TrpDL peptides, it is 4 pA by 500 ms to show their characteristically longer events. Note well that F427A nanopores conduct ~30% more than F427Y channels (13). 18 Fig. 2. Unsupervised clustering analysis of feature set describing translocation events from mutant nanopores. UMAP clustering analysis of event-level features. Data points, representing individual translocation events, are colored by guest-host peptide identity: Ala (black), Leu (red), Phe (green), Thr (blue), Trp (yellow), TrpDL (magenta), and Tyr (cyan). (A) UMAP embedding of F427A nanopore with ARI of 0.0536 and NMI of 0.0775. (B) UMAP embedding of F427Y nanopore events with ARI of 0.0947 and NMI of 0.1424. Events in either panel were filtered at a minimum event duration of 15 ms, and UMAP settings were n_neighbors = 15 and min_dist = 0.5. 19 Fig. 3. Per-peptide class F1-scores of individual nanopore ϕ-clamp variants. Per-class F1scores for single-stage XGBoost classification for WT (black), F427A (red), and F427Y (blue) nanopores. Plotted values represent the best performing evaluation out of 5 replicate trainings. 20 Fig. 4. Multistage ensemble probabilistic blending classifier. (A) Block diagram of multistage ensemble probabilistic blending classifier. The model processes raw translocation features (F) through a two-stage classification and blending pipeline. Stage 1 classifiers, one for each nanopore variant, predict the aromatic class probability (PA) for a given translocation event. These probabilities are then used to augment the original features (F) for Stage 2 classification. The Stage 2 classifiers predict peptide-specific probabilities for both aromatic (PApeptide) and non-aromatic (PNApeptide) peptides. Finally, a Meta-Classifier blends the probabilities from both Stage 1 (dashed line connections) and Stage 2 to produce the final peptide prediction. Normalized confusion matrices from most successful trainings of multistage ensemble probabilistic blending model for (B) a single F427A nanopore and (C) a WT/F427A ensemble. Rows represent true peptide labels, and columns represent predicted labels. Values indicate the proportion of events from a given true class that were predicted as each class. 1 Supporting Information for Comparative study of nanopore phenylalanine clamp variants reveals unique peptide biosensing and classification properties Jennifer M. Colby1 and Bryan A. Krantz2* 1Present address: Premier Biotech Labs, 723 Kasota Ave SE, Minneapolis, MN 55414, U.S.A. 2Department of Microbial Pathogenesis, School of Dentistry, University of Maryland, Baltimore, 650 W. Baltimore Street, Baltimore, MD 21201, U.S.A. * Corresponding author Bryan A. Krantz Department of Microbial Pathogenesis School of Dentistry, University of Maryland, Baltimore 650 W. Baltimore Street Baltimore, MD 21201, U.S.A (410) 706-1656 [email protected] This PDF file includes: Supporting text Figure S1 Table S1 SI References 1 Supporting Information Text Materials and Methods Nanopore and peptides. Monomeric 83-kDa PA (PA83) preprotein mutants, F427A and F427Y, and their homoheptameric prepore oligomers (PA7) were produced as described (1, 2). PA83 mutants were overexpressed in Escherichia coli BL21(DE3), using a pET22b plasmid, which directs expression to the periplasm. Cell cultures were grown at 37 °C in a custom 5 L fermentor using ECPM1 growth media (3), which was supplemented with carbenicillin (50 mg/L). Once reaching an OD600 of 3-10, the cultures were then induced with 1 mM isopropyl β-d-thiogalactopyranoside for ~3 h at 30 °C. PA83 was released from the periplasm by resuspending pelleted cells on ice using a wire whisk with 1 L of hypertonic sucrose buffer (20% sucrose, 20 mM Tris-Cl, 0.5 mM EDTA, pH 8) followed by osmotic shock of centrifuged/pelleted cells using a wire whisk in 1 L of hypotonic solution (5 mM MgCl2). Released PA83 monomer, isolated after centrifugation to remove cellular debris, was purified on Q-Sepharose anion-exchange chromatography in 20 mM Tris-Cl, pH 8.0 by binding and then eluting with a linear salt gradient using 20 mM Tris-Cl, pH 8.0 with 1 M NaCl. To make PA7 prepore oligomers of either mutation, purified PA83 at a concentration of 1 mg/ml was treated with trypsin (1:1000 wt/wt trypsin:PA) for 30 min at room temperature to form nicked PA. Trypsin was subsequently inhibited with soybean trypsin inhibitor at 1:100 dilution (wt/wt soybean trypsin inhibitor:PA). Nicked PA was applied to Q-Sepharose to then isolate the PA7; oligomer was bound to the column in 20 mM Tris-chloride, pH 8.0 and eluted by a linear salt gradient using 20 mM Tris-Cl, 1 M NaCl, pH 8.0. PA7 was concentrated and frozen in small aliquots to maintain reproducible nanopore insertion activity in planar bilayer experiments. Ten-residue guest-host peptides of the general sequence, KKKKKXXSXX, where X = A, L, F, T, W, and Y, were synthesized with standard ʟ amino acids (4, 5) (Elim 2 Biopharmaceuticals). One stereochemical variant of X = W (called TrpDL) was produced, where instead of synthesizing the peptide with uniform ʟ amino acids, an alternating pattern of ᴅ and ʟ amino acids was used (5). Single-channel electrophysiology. Planar lipid bilayer currents were recorded using an Axopatch 200B amplifier interfaced by a Digidata 1440A acquisition system (Molecular Devices) (2, 5, 6). Membranes were formed by painting across a 50-μm aperture of a 1-mL white Delrin cup with 3% (wt/vol) 1,2-diphytanoyl-sn-glycero-3phosphocholine (Avanti Polar Lipids) in n-decane. The cis (side to which the PA7 is added) and trans chambers were bathed in symmetric single-channel buffer (SCB: 100 mM KCl, 1 mM EDTA, 10 mM succinic acid, pH 5.60). Recordings were acquired at 500600 Hz using PCLAMP10. The applied voltage is defined as Δψ = ψcis - ψtrans (where ψtrans is 0 mV). Single-channel recordings of the guest-host peptide translocations via the PA nanopore were carried out as described (5) with some slight differences. A single PA channel was inserted into a painted bilayer at a Δψ of 20-30 mV by adding ~2 pM of PA7 (freshly diluted from a 2-μM stock) to the cis side of the membrane. The prepore oligomer converts to the nanopore state by inserting into the membrane in an oriented manner. Once a single channel inserted, the cis chamber was perfused by fresh SCB to remove excess uninserted PA7. Then the desired peptide analyte was added to the cis chamber at 5 to 20 nM. Translocation data were acquired by stepping the applied Δψ to a higher positive value and collecting recordings of the translocation event stream for up to thirty minutes. Minor processing as well as conductance state labeling of the raw single-channel event stream recordings was subsequently performed. Rare transient out-of-range current spikes, insertion of second channels, and inactivated channels were removed by a ‘force values’ routine in CLAMPFIT. Translocation recordings acquired at 500 or 600 3 Hz were downsampled to 400 Hz by decimation in Python using the scipy.signal library. 400 Hz was chosen to be consistent with prior WT PA datasets (7), maintaining a consistent time step for ML peptide classification models. Four discrete conductance states were then detected in these recordings using K-Means clustering, where baseline current drift was corrected by applying a moving window average offset (4000 time point window). While rare additional states were noted, the data were labeled for the four dominant states for either the F427A or F427Y mutants. By convention, the fully blocked peptide-bound state was state 0, the intermediate closest to the fully blocked state was state 1, the intermediate closest to the open state was state 2, and the open state was state 3. Ultimately, the state labeling produced a three-column CSV file of the stream with columns ‘Time’, ‘Current’, and ‘State’. >90% of time points were labeled consistently when comparing our K-Means algorithm and the more traditional CLAMPFIT. All labeled CSV stream files for the seven peptides were entered into a local annotated peptide database to aid in in situ loading/preprocessing for each tested ML model. Hardware and software used for ML. Anaconda was used to create a Python 3.10.16 environment, where XGBoost (3.0.0) (8) and other standard modules were installed. The hardware used in preprocessing and the ML peptide classifications was a 2025 MacBook Pro with M4 Apple Silicon and 24 GB of RAM. All source code is available at GitHub (https://github.com/bakrantz/Pept-Class). Preprocessing, event segmentation, and feature extraction. Raw state-labeled event streams were segmented into translocation events, enabling feature extraction, as described (7). The minimum event duration, which served as an effective filter for excluding very short-duration events, was set to 15 ms to optimally capture the most information-rich data. This filtering parameter was previously systematically varied in the range of 5 to 20 ms (7). Each segmented event was defined as initiating when the current changed from the fully open state (state 3, corresponding to baseline current) to 4 any peptide-bound state (state 0, 1, or 2) and terminating when the current returned to the open state. From these segmented events, both raw current sequences and corresponding state sequences were extracted. A comprehensive set of event-level features was then computed from these two different sequences using a custom segmentation core. This core maintains a generalizable framework to process peptide translocation events from systems exhibiting diverse mechanisms and an arbitrary number of states. These features were initially grouped into scalar, vector, and matrix data structures, with values in vector and matrix features being state or transition enumerated. Scalar features included: Shannon entropy of state sequence, event duration, number of transitions, time of the first transition, total number of states visited during the event, skewness of the current sequence, and kurtosis of the current sequence. Vector features included: observed conductance state Boolean, observed conductance levels, probability of residing in each state, and longest dwell time in each state. Matrix features included: average dwell time for specific state-to-state transitions, variance of dwell time for transitions, and ratio of probabilities between states. For downstream ML classification, all matrix features were flattened into one-dimensional arrays and appended with the vector and scalar features to form a single feature vector for each translocation event. These flattened descriptive key names for the features were generated to maintain traceability in subsequent applications. All processed event sequences, their flattened features, and associated feature key names were saved as a Python pickle object for efficient storage and cached retrieval. A local peptide events database was employed to track these preprocessed cached pickle files for efficiency. Initial single nanopore ML peptide classification. ML-based classification of peptide translocation events from single individual nanopore variants was performed using the gradient boosting framework, XGBoost (8), which was implemented as described (7). For this, peptide event data, previously extracted and characterized into