1. Bibliographic Information

1.1. Title

iBitter-SCM: Identification and characterization of bitter peptides using a scoring card method with propensity scores of dipeptides

1.2. Authors

Phasit Charoenkwan, Janchai Yana, Nalini Schaduangrat, Chanin Nantasenamat, Md. Mehedi Hasan, Watshara Shoombuatong

1.3. Journal/Conference

Published at (UTC): 2020-03-28T00:00:00.000Z. The specific journal is not explicitly stated in the provided abstract or paper content, but it is implied to be a peer-reviewed scientific journal in the field of bioinformatics or computational biology, given the nature of the research and the mention of "ygeno.2020.03.019" in the supplementary data link, which often indicates journal abbreviation (e.g., Genomics). The publication date suggests it has undergone peer review and is officially published.

1.4. Publication Year

2020

1.5. Abstract

This paper introduces iBitter-SCM, a novel computational model designed for the prediction and characterization of bitter peptides directly from their amino acid sequences. The motivation stems from the crucial role of bitter taste perception in protecting animals from toxins and the increasing volume of peptides in the post-genomic era, necessitating rapid computational discrimination of bitter from non-bitter peptides. iBitter-SCM is built upon the scoring card method (SCM), leveraging estimated propensity scores of amino acids and dipeptides. Benchmarking on an independent dataset demonstrated its effectiveness, achieving an accuracy of 84.38% and a Matthews Correlation Coefficient (MCC) of 0.688. The study highlights iBitter-SCM's superiority over other widely used machine learning classifiers (such as k-nearest neighbor, naive Bayes, decision tree, and random forest) due to its simplicity, interpretability, and ease of implementation. Furthermore, the analysis of propensity scores provides valuable insights into the biophysical and biochemical properties of bitter peptides. A public web server is provided, aiming to facilitate high-throughput prediction and de novo design of bitter peptides.

1.6. Original Source Link

/files/papers/69135bc8430ad52d5a9ef439/paper.pdf (This appears to be a local path or an internal identifier within a larger system, indicating the PDF is available but the exact public URL is not directly provided in the abstract text. The supplementary data link https://doi.org/10.1016/j.ygeno.2020.03.019 suggests an official publication at a DOI resolver.)

2. Executive Summary

2.1. Background & Motivation

The perception of bitter taste is a fundamental biological mechanism that protects animals from potentially poisonous plants and environmental toxins. In the context of food science and pharmaceutical development, peptides derived from hydrolyzed proteins or synthetic sources can exhibit undesirable bitter tastes, impacting product palatability and drug efficacy. The post-genomic era has led to an explosion in the discovery and synthesis of diverse peptides, making high-throughput experimental identification of bitter peptides time-consuming and costly. Existing computational methods, primarily quantitative structure-activity relationship (QSAR) models, often rely on complex 3D structural information or generic chemical properties, and no dedicated sequence-based computational model existed specifically for predicting and analyzing bitter peptides from their primary amino acid sequence. This gap presented a significant challenge for efficient drug development and nutritional research.

The paper addresses this challenge by recognizing the urgent need for a computational tool that can rapidly and effectively discriminate bitter from non-bitter peptides using only their amino acid sequence. This approach bypasses the complexities and limitations of structural information, offering a more direct and scalable solution.

2.2. Main Contributions / Findings

The primary contributions and findings of this paper are:

• First Sequence-Based Computational Model: The paper presents iBitter-SCM, the first computational model specifically designed to predict and characterize the bitterness of peptides solely based on their amino acid sequence. This eliminates the dependency on functional domain or structural information, a significant advancement over previous QSAR or machine learning (ML) approaches that often required more complex descriptors.

• Novel Application of Scoring Card Method (SCM): iBitter-SCM leverages a scoring card method (SCM) that estimates propensity scores for both amino acids and dipeptides. This method is described as simple, effective, and interpretable, providing insights into the underlying biochemical properties.

• High Predictive Performance: iBitter-SCM demonstrated robust predictive performance, achieving an accuracy of 84.38% and a Matthews Correlation Coefficient (MCC) of 0.688 on an independent test dataset. Its Area Under the Receiver Operating Characteristic curve (auROC) was 0.904, indicating strong discriminative power.

• Superiority over Conventional ML Classifiers: Rigorous independent testing showed that iBitter-SCM outperformed other widely used ML classifiers (e.g., k-nearest neighbor (KNN), naive Bayes (NB), decision tree (DT), support vector machine (SVM), and random forest (RF)) in terms of overall effectiveness, stability, and interpretability, particularly for the independent test set.

• Mechanistic Interpretability: The analysis of the estimated propensity scores of amino acids and dipeptides, along with informative physicochemical properties (PCPs), provided a deeper understanding of the biophysical and biochemical characteristics contributing to peptide bitterness. Key findings included the importance of hydrophobic amino acids (especially Phe and Pro), the effect of hydrophobic amino acids at the C-terminus, and the influence of the number of carbon atoms in the amino acid side chain on bitterness intensity.

• Publicly Available Web Server: For practical utility, a user-friendly web server for iBitter-SCM was developed and made publicly available, enabling experimental scientists to easily predict and design bitter peptides.

  These findings solve the problem of needing a fast, accurate, and interpretable method for identifying bitter peptides from sequence information, facilitating drug development and nutritional research by streamlining the screening and design process.

3. Prerequisite Knowledge & Related Work

3.1. Foundational Concepts

To fully understand the iBitter-SCM model, a reader should be familiar with several core concepts from biology, chemistry, and machine learning:

• Peptides: Peptides are short chains of amino acids linked by peptide bonds. They are smaller than proteins and play diverse biological roles. The bitterness of peptides is often related to their amino acid composition and sequence.
• Amino Acids: The fundamental building blocks of peptides and proteins. There are 20 common natural amino acids, each with a unique side chain that confers specific chemical properties (e.g., hydrophobicity, charge, size).
• Dipeptides: A molecule consisting of two amino acids joined by a single peptide bond. The paper utilizes dipeptide composition as a key feature.
• Bitter Taste Perception: One of the five basic tastes. It is often perceived as unpleasant and serves as a warning system against toxins. In humans, bitter taste is mediated by a family of G protein-coupled receptors called TAS2Rs.
• Hydrophobic Amino Acids: Amino acids that tend to repel water and associate with other hydrophobic molecules. Examples include Phenylalanine (Phe), Proline (Pro), Isoleucine (Ile), Leucine (Leu), Valine (Val). The paper emphasizes their importance in bitterness.
• Quantitative Structure-Activity Relationship (QSAR) Modeling: A computational approach that seeks to find a mathematical relationship between the chemical structure of a compound (or peptide) and its biological activity. QSAR models use molecular descriptors (features derived from chemical structure) to predict properties like bitterness.
• Machine Learning (ML): A field of artificial intelligence that focuses on building systems that learn from data. ML algorithms are trained on known examples to make predictions or decisions on new, unseen data.
• Scoring Card Method (SCM): A general-purpose ML approach specifically designed for predicting and analyzing protein and peptide functions from their amino acid sequence. It works by estimating propensity scores for amino acids and dipeptides regarding a specific function. The method is valued for its simplicity and interpretability.
• Propensity Scores: In the context of SCM, propensity scores quantify the likelihood or tendency of a particular amino acid or dipeptide to be associated with a specific biological property (e.g., bitterness). Higher propensity scores indicate a stronger association.
• Dipeptide Composition (DPC): A sequence-based feature representation that counts the occurrences of all possible dipeptides (pairs of adjacent amino acids) in a peptide sequence. Since there are 20 amino acids, there are $20 \times 20 = 400$ possible dipeptides. This forms a 400-dimensional vector.
• Genetic Algorithm (GA): A metaheuristic inspired by the process of natural selection. GAs are used to find optimized solutions to search and optimization problems. In SCM, GA can be used to optimize the propensity scores to improve model performance.
• Cross-Validation (CV): A technique to assess how well a machine learning model generalizes to an independent dataset. k-fold cross-validation involves splitting the dataset into $k$ subsets, training on k-1 subsets, and testing on the remaining one, repeating this $k$ times.
• Independent Test Set: A portion of the dataset held back during model training and cross-validation, used only once at the very end to provide an unbiased evaluation of the final model's performance on unseen data.
• Evaluation Metrics (for Binary Classification):
  • Accuracy (Ac): The proportion of correctly classified instances (both bitter and non-bitter) out of the total number of instances. $ \mathrm{Ac} = \frac{\mathrm{TP} + \mathrm{TN}}{\mathrm{TP} + \mathrm{TN} + \mathrm{FP} + \mathrm{FN}} $ where $\mathrm{TP}$ is True Positives, $\mathrm{TN}$ is True Negatives, $\mathrm{FP}$ is False Positives, and $\mathrm{FN}$ is False Negatives.
  • Sensitivity (Sn) (also known as Recall or True Positive Rate): The proportion of actual bitter peptides that were correctly identified as bitter. $ \mathrm{Sn} = \frac{\mathrm{TP}}{\mathrm{TP} + \mathrm{FN}} $
  • Specificity (Sp) (also known as True Negative Rate): The proportion of actual non-bitter peptides that were correctly identified as non-bitter. $ \mathrm{Sp} = \frac{\mathrm{TN}}{\mathrm{TN} + \mathrm{FP}} $
  • Matthews Correlation Coefficient (MCC): A measure of the quality of binary classifications. It is considered a balanced measure even if the classes are of very different sizes. A value of +1 indicates a perfect prediction, 0 an average random prediction, and -1 an inverse prediction. $ \mathrm{MCC} = \frac{(\mathrm{TP} \times \mathrm{TN}) - (\mathrm{FP} \times \mathrm{FN})}{\sqrt{(\mathrm{TP} + \mathrm{FP})(\mathrm{TP} + \mathrm{FN})(\mathrm{TN} + \mathrm{FP})(\mathrm{TN} + \mathrm{FN})}} $
  • Receiver Operating Characteristic (ROC) Curve: A graphical plot that illustrates the diagnostic ability of a binary classifier system as its discrimination threshold is varied. It plots the True Positive Rate (Sensitivity) against the False Positive Rate (1-Specificity).
  • Area Under the ROC Curve (auAUC): A scalar value that summarizes the overall performance of a classifier. An auAUC of 1.0 indicates a perfect classifier, while 0.5 indicates a random classifier.

3.2. Previous Works

The paper contextualizes its contribution by referencing several previous studies, primarily QSAR models and ML classifiers, for predicting bitterness:

• Yin et al. (2010): Developed 28 QSAR models to predict the bitterness of dipeptides. They used quantitative multidimensional amino acid descriptors E1-E5, representing properties like hydrophobicity, steric properties, alpha-helix preferences, composition, and net charge. Their analysis was based on support vector regression (SVR) for ACE inhibitors and bitter dipeptides, achieving good leave-one-out cross-validation and root mean squared error values.

• Soltani et al. (2013): Analyzed 229 experimental bitterness values for 224 peptides and 5 amino acids, specifically bitter thresholds ($\log(1/T)$). They employed multiple linear regression (MLR), support vector machine (SVM), and artificial neural network (ANN) for model building, describing peptides with a set of 3D descriptors (1295 descriptors). This highlights the reliance on complex 3D information.

• Xu and Chung (2019): Proposed QSAR models for bitter peptides by integrating fourteen amino acid descriptors. Their dataset included di-, tri-, and tetrapeptides with bitter taste thresholds. They reported high cross-validation results for different peptide lengths.

• Huang et al. (2016) - BitterX: An open-access tool for identifying human bitter taste receptors, TAS2Rs, for small molecules. It used sequential minimal optimization (SMO), logistic regression (LR), and random forest (RF) to discriminate bitter from non-bitter compounds, achieving accuracies of 0.93 (training) and 0.83 (hold-out test). This tool was focused on small molecules rather than peptides.

• Dagan-Wiener et al. (2017) - BitterPredict: An ML classifier that predicts bitterness of compounds based on their chemical structures, achieving over 80% accuracy on a hold-out test. Similar to BitterX, this focused on chemical structures (small molecules), not specifically peptide sequences.

  Crucially, while these prior works contributed significantly to understanding bitterness prediction, they either focused on small molecules, relied heavily on 3D structural descriptors, or developed QSAR models that were not explicitly sequence-based for peptides. None provided a dedicated computational model for predicting and characterizing bitter peptides directly from sequence information in an interpretable manner, which is the gap iBitter-SCM aims to fill.

3.3. Technological Evolution

The prediction of biological activities like taste perception has evolved from purely experimental and labor-intensive methods to sophisticated computational approaches. Initially, such predictions were often made through direct experimental assays, which are costly and time-consuming, especially for large datasets. The advent of QSAR modeling marked a significant step, enabling the prediction of activity based on chemical structure, often using molecular descriptors that could be 2D or 3D. These methods, while powerful, could be complex to implement and interpret, particularly when dealing with the dynamic nature and conformational flexibility of peptides.

The rise of machine learning provided more flexible and powerful tools for pattern recognition in complex biological data. Early ML applications in this field often still relied on manually engineered descriptors or physicochemical properties derived from sequences or structures. However, there has been a growing trend towards developing simpler, more interpretable, and purely sequence-based methods, especially for peptides, where primary sequence dictates much of the behavior. iBitter-SCM represents an advancement in this trajectory by providing a highly interpretable, sequence-based model that leverages propensity scores to not only predict but also explain the underlying biochemical drivers of bitterness in peptides. It moves beyond black-box ML models by providing a clear link between amino acid and dipeptide composition and the predicted outcome.

3.4. Differentiation Analysis

Compared to the main methods in related work, iBitter-SCM offers several key differentiators and innovations:

• Sequence-Only Prediction for Peptides: The most significant difference is that iBitter-SCM is the first computational model dedicated to predicting and characterizing bitter peptides solely from their amino acid sequence. Previous QSAR models for peptides (e.g., Yin et al., Soltani et al., Xu and Chung) often relied on 3D descriptors or a wide array of amino acid attributes that could implicitly incorporate structural or functional domain information. Tools like BitterX and BitterPredict focused on small molecules and their chemical structures, not peptide sequences. iBitter-SCM simplifies the input requirement, making it broadly applicable without the need for complex structural calculations.

• Interpretability via Scoring Card Method (SCM): Unlike many black-box machine learning classifiers (SVM, RF, ANN) that offer little insight into why a prediction is made, iBitter-SCM uses the SCM to derive propensity scores for individual amino acids and dipeptides. These scores directly indicate the contribution of each amino acid or dipeptide to the bitterness, providing clear, biologically meaningful interpretations of the prediction. This is a crucial advantage for researchers looking to understand the underlying mechanisms or design new peptides.

• Simplicity and Effectiveness: The paper emphasizes iBitter-SCM's simplicity and effectiveness. While other ML models can be powerful, SCM offers a straightforward methodology that is easy to understand and implement, yet achieves competitive, if not superior, performance.

• Focus on Biophysical and Biochemical Properties: By analyzing propensity scores and their correlation with physicochemical properties (PCPs), iBitter-SCM directly unravels the specific characteristics (e.g., hydrophobicity, C-terminal location, side chain length) that contribute to bitterness, providing targeted insights for de novo peptide design.

• Public Web Server: The provision of a user-friendly web server makes the tool readily accessible to experimental scientists, facilitating high-throughput applications and bridging the gap between computational models and practical research.

  In essence, iBitter-SCM differentiates itself by offering a unique combination of sequence-based prediction, high interpretability, and robust performance, specifically tailored for bitter peptides, in contrast to the more general or structure-dependent approaches found in prior work.

4. Methodology

4.1. Principles

The core idea behind iBitter-SCM is to identify and characterize bitter peptides by quantifying the relative contribution of individual amino acids and dipeptides to the bitter taste. This is achieved through the Scoring Card Method (SCM), which calculates propensity scores. The principle is that certain amino acids or dipeptide combinations occur more frequently or have a stronger association with bitter peptides compared to non-bitter ones. By learning these propensity scores from a training dataset, the model can then predict the bitterness of an unknown peptide by summing the propensity scores of its constituent dipeptides (or amino acids) in a weighted-sum approach. The method inherently offers interpretability because the propensity scores directly reveal which amino acids and dipeptides are more indicative of bitterness.

4.2. Core Methodology In-depth (Layer by Layer)

The iBitter-SCM model development involves several structured steps, starting from dataset preparation to feature representation and model construction using the SCM.

4.2.1. Benchmark Datasets

The first step is to establish a high-quality benchmark dataset. The authors followed a rigorous procedure:

1. Collection: Experimentally confirmed bitter peptides were manually collected from various literature sources [3,9-14,17,20-26].
2. Filtering: Peptides containing ambiguous residues (e.g., $X$, $B$, $U$, $Z$) were excluded.
3. Deduplication: Duplicate peptide sequences were removed, ensuring that each peptide in the dataset was unique.
4. Positive Dataset: The cleaned collection of unique bitter peptides formed the positive dataset, comprising 320 bitter peptides.
5. Negative Dataset: Due to the scarcity of experimentally validated non-bitter peptides (often less scientific significance in publishing negative results), a standard procedure was adopted. 320 non-bitter peptides were randomly generated from the BIOPEP database [27], a well-known and universally accepted database for bioactive peptides.
6. Combined Dataset: The positive and negative datasets were combined to form the BTP640 benchmark dataset, containing a total of 640 peptides (320 bitter, 320 non-bitter).
7. Dataset Split: To prevent overestimation of the model's performance, BTP640 was randomly split into a training set (BTP-CV) and an independent test set (BTP-TS) with an 8:2 ratio.
   • BTP-CV (training set): Consisted of 256 bitter and non-bitter peptides (approximately 128 bitter, 128 non-bitter).
   • BTP-TS (independent test set): Consisted of 64 bitter and non-bitter peptides (approximately 32 bitter, 32 non-bitter).

4.2.2. Feature Representation

Once the dataset is prepared, each peptide sequence needs to be converted into a numerical format that machine learning models can process. The iBitter-SCM utilizes dipeptide composition (DPC) as its feature representation.

Given a peptide sequence $\mathbf{P}$, it can be represented as: $ \mathbf{P} = \mathtt{p}_1 \mathtt{p}_2 \mathtt{p}3 ... \mathtt{p}{\mathrm{N}} $ where $\mathtt{p}_i$ denotes the $i^{th}$ amino acid residue in the peptide $\mathbf{P}$, and $\mathrm{N}$ is the length of the peptide. Each $\mathtt{p}_i$ belongs to the set of 20 natural amino acids (A, C, D, E, F, G, H, I, K, L, M, N, P, Q, R, S, T, V, W, Y).

Dipeptide composition captures the frequencies of all possible adjacent pairs of amino acids. Since there are 20 amino acids, there are $20 \times 20 = 400$ possible dipeptides. The DPC feature vector counts the occurrences of these dipeptides within a given peptide sequence.

A peptide sequence $\mathbf{P}$ is thus expressed by a vector with 400 dimensions: $ \mathbf{P} = [ \mathrm{d}\mathbf{p}_1, \mathrm{d}\mathbf{p}2, ..., \mathrm{d}\mathbf{p}{400} ]^{\mathbf{T}} $ where $\mathbf{T}$ is the transposed operator, and $\mathrm{d}\mathbf{p}_j$ for $j=1, ..., 400$ represents the occurrence frequency of the $j^{th}$ dipeptide in the peptide sequence $\mathbf{P}$. This means that each element $\mathrm{d}\mathbf{p}_j$ is calculated as the number of times a specific dipeptide appears in $\mathbf{P}$, normalized by the total number of dipeptides in $\mathbf{P}$ (which is $\mathrm{N}-1$).

4.2.3. Scoring Card Method (SCM)

The SCM is the core algorithm used to build iBitter-SCM. It is a general-purpose approach that estimates propensity scores of amino acids and dipeptides for a specific function (in this case, bitterness). The SCM process involves six main steps:

1. Preparing Training and Independent Datasets: This refers to the BTP-CV and BTP-TS datasets created in the Benchmark Datasets step.
2. Calculating Initial Dipeptide Propensity Score (init-DPS): The init-DPS is derived using a statistical approach based on the normalized dipeptide composition of bitter and non-bitter peptides in the BTP-CV training set. While the exact statistical formula for init-DPS isn't explicitly provided in the "Scoring card method" subsection, generally, propensity scores are related to the ratio of a dipeptide's frequency in positive samples versus its frequency in negative samples, or some form of normalized log-odds ratio.
3. Optimizing init-DPS to obtain Augmented Dipeptide Propensity Score (opti-DPS) using a Genetic Algorithm (GA): The init-DPS values are further optimized using a Genetic Algorithm (GA). A GA is an iterative optimization heuristic inspired by natural selection. It works by maintaining a population of candidate solutions (in this case, sets of dipeptide propensity scores), which are then iteratively improved through processes like selection, crossover (recombination), and mutation, guided by a fitness function (e.g., MCC or Accuracy). The GA fine-tunes the propensity scores to maximize the model's performance on the training data, leading to the opti-DPS. The non-deterministic nature of GA means that running it multiple times might yield slightly different opti-DPS sets, thus the paper mentions performing ten experiments.
4. Estimating Amino Acid Propensity Scores using a Statistical Approach: Once the opti-DPS (for dipeptides) are determined, the amino acid propensity scores are estimated. This is typically done by summing the propensity scores of dipeptides that contain a particular amino acid, or by other statistical methods that infer individual amino acid contributions from the dipeptide scores. The paper references previous studies [46] for details.
5. Discriminating Bitter Peptides from Non-Bitter Peptides using a Weighted-Sum with opti-DPS: For a given query peptide sequence $\mathbf{P}$, a "bitter score" (BS) is calculated. This score is a weighted-sum of the opti-DPS values corresponding to the dipeptides present in $\mathbf{P}$. A peptide $\mathbf{P}$ is classified as bitter if its bitter score $S(\mathbf{P})$ exceeds a certain threshold; otherwise, it is classified as non-bitter. The formula for the weighted-sum score $S(\mathbf{P})$ for a peptide $\mathbf{P}$ is implicitly defined by the SCM and DPC representation. If $\mathbf{P} = \mathtt{p}_1 \mathtt{p}_2 ... \mathtt{p}_{\mathrm{N}}$, then the dipeptides are $(\mathtt{p}_1, \mathtt{p}_2), (\mathtt{p}_2, \mathtt{p}_3), ..., (\mathtt{p}_{\mathrm{N-1}}, \mathtt{p}_{\mathrm{N}})$. Let $DPS(\mathtt{a}_i\mathtt{a}_j)$ be the optimized dipeptide propensity score for the dipeptide formed by amino acid $\mathtt{a}_i$ followed by amino acid $\mathtt{a}_j$. The score for a peptide $\mathbf{P}$ of length $\mathrm{N}$ is calculated as: $ S(\mathbf{P}) = \sum_{k=1}^{\mathrm{N}-1} DPS(\mathtt{p}k \mathtt{p}{k+1}) $ The classification threshold is determined during the optimization phase (step 3) to achieve the best performance on the training set.
6. Bitter Peptides Characterization using the Propensity Scores of Amino Acids and Dipeptides: The calculated propensity scores for amino acids and dipeptides are then analyzed to understand the biochemical and biophysical properties that contribute to bitterness. This involves identifying amino acids and dipeptides with the highest (or lowest) scores and correlating them with known physicochemical properties (PCPs) from databases like AAindex.

4.2.4. Characterization of the Bitter Taste of Peptides

To gain deeper insights into the biophysical and biochemical basis of peptide bitterness, the propensity scores derived from SCM are further analyzed.

• Amino Acid and Dipeptide Propensity Scores: These scores, estimated using SCM, directly reflect the influence of each amino acid and dipeptide on the bitterness property. High scores indicate a strong positive contribution to bitterness.
• Informative Physicochemical Properties (PCPs): The paper identifies relevant PCPs from the AAindex database [49] by calculating the Pearson correlation coefficient (R) between the amino acid propensity scores and various PCPs. PCPs are fundamental attributes of amino acids (e.g., hydrophobicity, charge, size, polarity) that dictate their behavior in biological systems. By finding PCPs that strongly correlate with the propensity scores, the model can highlight which fundamental amino acid properties are most critical for bitterness. This step helps in mechanistically interpreting the findings beyond just identifying specific amino acids. The methods for determining informative PCPs are referenced from previous studies [28-30].

5. Experimental Setup

5.1. Datasets

The study utilized a manually curated benchmark dataset named BTP640, which comprised 320 bitter and 320 non-bitter peptides.

• Bitter Peptides (Positive Set): Manually collected from various literature sources [3,9-14,17,20-26]. After filtering ambiguous residues and removing duplicates, this set contained 320 unique experimentally validated bitter peptides.

• Non-Bitter Peptides (Negative Set): To compensate for the scarcity of experimentally validated non-bitter peptides, 320 peptides were randomly generated from the BIOPEP database [27]. This is a common practice in bioinformatics studies when negative data is limited.

  The BTP640 dataset was then randomly divided into:

• Training Set (BTP-CV): 256 peptides (128 bitter, 128 non-bitter). This set was used for model development, propensity score optimization via Genetic Algorithm, and 10-fold cross-validation.

• Independent Test Set (BTP-TS): 64 peptides (32 bitter, 32 non-bitter). This set was reserved for an unbiased evaluation of the final model, ensuring its generalization ability to unseen data.

  These datasets were chosen to represent a diverse collection of peptides relevant to bitterness research, and the split ensured robust validation against overfitting.

5.2. Evaluation Metrics

To assess the prediction ability of the iBitter-SCM model and compare it with baseline methods, the authors employed four widely used metrics for binary classification problems, along with ROC curves and auAUC for threshold-independent evaluation.

1. Accuracy (Ac):

   • Conceptual Definition: Accuracy measures the overall correctness of the model's predictions by calculating the proportion of both true positive and true negative predictions out of the total number of predictions made. It indicates how often the classifier is correct.
   • Mathematical Formula: $ \mathrm{Ac} = \frac{\mathrm{TP} + \mathrm{TN}}{\mathrm{TP} + \mathrm{TN} + \mathrm{FP} + \mathrm{FN}} $
   • Symbol Explanation:
     • $\mathrm{TP}$: True Positives (correctly predicted bitter peptides).
     • $\mathrm{TN}$: True Negatives (correctly predicted non-bitter peptides).
     • $\mathrm{FP}$: False Positives (non-bitter peptides incorrectly predicted as bitter).
     • $\mathrm{FN}$: False Negatives (bitter peptides incorrectly predicted as non-bitter).

2. Sensitivity (Sn) (also known as Recall or True Positive Rate):

   • Conceptual Definition: Sensitivity quantifies the model's ability to correctly identify all positive instances. In this context, it measures the proportion of actual bitter peptides that were successfully identified by the model.
   • Mathematical Formula: $ \mathrm{Sn} = \frac{\mathrm{TP}}{\mathrm{TP} + \mathrm{FN}} $
   • Symbol Explanation:
     • $\mathrm{TP}$: True Positives.
     • $\mathrm{FN}$: False Negatives.

3. Specificity (Sp) (also known as True Negative Rate):

   • Conceptual Definition: Specificity measures the model's ability to correctly identify all negative instances. Here, it indicates the proportion of actual non-bitter peptides that were correctly identified as non-bitter.
   • Mathematical Formula: $ \mathrm{Sp} = \frac{\mathrm{TN}}{\mathrm{TN} + \mathrm{FP}} $
   • Symbol Explanation:
     • $\mathrm{TN}$: True Negatives.
     • $\mathrm{FP}$: False Positives.

4. Matthews Correlation Coefficient (MCC):

   • Conceptual Definition: MCC is a robust and reliable statistical rate that produces a high score only if the prediction obtained good results in all of the four confusion matrix categories (true positives, false negatives, true negatives, and false positives), proportionally to the size of both positive and negative elements in the dataset. It is particularly useful for imbalanced datasets, though this dataset is balanced. It ranges from -1 (total disagreement) to +1 (perfect prediction), with 0 indicating random prediction.
   • Mathematical Formula: $ \mathrm{MCC} = \frac{(\mathrm{TP} \times \mathrm{TN}) - (\mathrm{FP} \times \mathrm{FN})}{\sqrt{(\mathrm{TP} + \mathrm{FP})(\mathrm{TP} + \mathrm{FN})(\mathrm{TN} + \mathrm{FP})(\mathrm{TN} + \mathrm{FN})}} $
   • Symbol Explanation:
     • $\mathrm{TP}$: True Positives.
     • $\mathrm{TN}$: True Negatives.
     • $\mathrm{FP}$: False Positives.
     • $\mathrm{FN}$: False Negatives.

5. Receiver Operating Characteristic (ROC) Curve and Area Under the ROC Curve (auAUC):

   • Conceptual Definition: The ROC curve is a probability curve that plots the True Positive Rate (Sensitivity) against the False Positive Rate (1-Specificity) at various threshold settings. The auAUC measures the entire two-dimensional area underneath the entire ROC curve. auAUC values typically range from 0.5 (random classifier) to 1.0 (perfect classifier), providing a single value that summarizes the model's performance across all possible classification thresholds, making it threshold-independent.

5.3. Baselines

The iBitter-SCM model was compared against several baseline methods to demonstrate its performance:

1. BLAST (Basic Local Alignment Search Tool): A widely used bioinformatics algorithm for comparing primary biological sequence information, such as amino acid sequences, based on sequence similarity. In this study, BLASTP was used, with the BTP-CV dataset serving as the database and BTP-TS as query sequences. Various E-value cut-offs (0.1 to 0.0001) were tested.
2. Conventional Machine Learning Classifiers: The following machine learning models were implemented using the Scikit-Learn package [62] with the same dipeptide composition feature representation and cross-validation methods as iBitter-SCM:

   • Support Vector Machine (SVM): A powerful supervised machine learning model used for classification and regression tasks. It works by finding an optimal hyperplane that best separates different classes in the feature space.

   • Random Forest (RF): An ensemble learning method that constructs a multitude of decision trees during training and outputs the class that is the mode of the classes (classification) or mean prediction (regression) of the individual trees.

   • k-Nearest Neighbor (KNN): A non-parametric, instance-based learning algorithm that classifies a data point based on the majority class of its $k$ nearest neighbors in the feature space.

   • Naive Bayes (NB): A probabilistic classifier based on Bayes' theorem with the "naive" assumption of conditional independence between features given the class label.

   • Decision Tree (DT): A tree-like model where each internal node represents a test on an attribute, each branch represents an outcome of the test, and each leaf node represents a class label.

     These baselines are representative of widely used similarity-based and machine learning approaches for sequence classification, providing a comprehensive comparison for the proposed iBitter-SCM.

6. Results & Analysis

6.1. Core Results Analysis

6.1.1. Prediction Results using BLAST

The study first assessed the performance of BLAST, a similarity-based search tool, for predicting bitter peptides. The BTP-CV dataset was used as the BLASTP database, and BTP-TS as query sequences.

The following are the results from Table 1 of the original paper:

The results show that BLAST achieved its highest accuracy of 65.08% at an E-value of 0.1. However, the sensitivity (Sn) was very low across all E-values, indicating that BLAST was poor at identifying actual bitter peptides (e.g., only 34.92% at 0.1 E-value). While specificity (Sp) was high, this often comes at the cost of sensitivity when dealing with imbalanced or hard-to-classify positive cases. The authors concluded that similarity-based search methods like BLAST are insufficient for accurately predicting bitter peptides, necessitating the development of more intelligent ML models.

6.1.2. Prediction Performance of iBitter-SCM

The iBitter-SCM model was developed using the scoring card method (SCM) with dipeptide propensity scores. Since the Genetic Algorithm (GA) used for optimizing these scores is non-deterministic, ten independent experiments were run, each producing a different set of optimized dipeptide propensity scores (opti-DPS). The performance of these ten SCM models was evaluated using 10-fold cross-validation (CV) on BTP-CV and an independent test on BTP-TS.

The paper mentions that the opti-DPS from experiment #5 (which was among the top-ranked for both 10-fold CV and independent test) was chosen as the optimal one.

The following are the results from Table 2 of the original paper, showing the comparison of ten SCM models over the independent test:

Experiment #5 achieved the highest accuracy (84.38%) and MCC (0.688) on the independent test set, along with a high auROC (0.904). This experiment's opti-DPS was thus selected for the final iBitter-SCM model. The consistency of opti-DPS from experiment #5 and #7 in the top ranks for both 10-fold CV and independent tests further validated their robustness.

6.1.3. Contribution and Effectiveness of the Estimated Propensities of Dipeptides

The authors further investigated the effectiveness of the optimized propensity scores (opti-DPS) compared to the initial, statistically derived ones (init-DPS).

The following are the results from Table 3 of the original paper:

| Method | 10-fold CV | | Independent test | | | | :------- | :--------- | :------ | :----------------- | :------ | :------ | :------ | | Ac (%) | MCC | Ac (%) | Sn (%) | Sp (%) | MCC | Init-DPS | 85.17 | 0.716 | 81.25 | 86.46 | 76.04 | 0.628 | opti-DPS | 87.11 | 0.751 | 84.38 | 84.38 | 84.38 | 0.688

As shown in Table 3, opti-DPS significantly outperformed init-DPS. On 10-fold CV, opti-DPS showed improvements of 2% in accuracy (87.11% vs 85.17%) and 4% in MCC (0.751 vs 0.716). On the independent test, opti-DPS yielded improvements of 3% in accuracy (84.38% vs 81.25%), 2% in sensitivity (84.38% vs 86.46%), 8% in specificity (84.38% vs 76.04%), and 6% in MCC (0.688 vs 0.628).

The improvement in specificity (correctly identifying non-bitter peptides) was particularly notable. This suggests that the Genetic Algorithm optimization process effectively refined the propensity scores to better distinguish between bitter and non-bitter peptides.

The visual representation in Fig. 3 (histogram of scores for bitter and non-bitter peptides) further supports this. Fig. 3(a) (using init-DPS) shows more overlap between the score distributions of bitter (blue) and non-bitter (red) peptides, indicating weaker discriminative power. In contrast, Fig. 3(b) (using opti-DPS) shows much clearer separation, with distinct peaks for bitter and non-bitter peptides, demonstrating the enhanced ability of opti-DPS to discriminate between the two classes.

6.1.4. Comparison of iBitter-SCM with Conventional Classifiers

To further validate the superiority of iBitter-SCM, its performance was compared against five widely used machine learning classifiers: SVM, RF, NB, KNN, and DT. These models were trained and tested on the same datasets (BTP-CV and BTP-TS) using the same dipeptide composition features.

The following are the results from Table 4 of the original paper:

On the BTP-CV training set, iBitter-SCM achieved the highest accuracy (87.11%), MCC (0.751), and auROC (0.903), significantly outperforming all other classifiers. SVM and RF followed as the next best performers.

On the crucial BTP-TS independent test set, iBitter-SCM again demonstrated strong performance, matching SVM in accuracy (84.38%) and MCC (0.688). While RF achieved a slightly higher auROC (0.916 vs 0.904) and sensitivity (90.63% vs 84.38%), its accuracy and MCC were slightly lower than iBitter-SCM. Given that MCC is a robust metric and the independent test is the most rigorous validation, the authors concluded that iBitter-SCM is more effective and stable than RF. The high specificity of iBitter-SCM (84.38% vs RF's 76.56%) indicates its balanced performance in correctly identifying both bitter and non-bitter peptides.

The ROC curves presented in Fig. 4 (not shown here, but described in text) would visually confirm iBitter-SCM's strong discriminative power against other ML models.

The superior performance of iBitter-SCM compared to other widely used ML models, combined with its simplicity, interpretability, and implementation, makes it a promising tool for bitter peptide prediction.

6.1.5. Identification of Peptides Having High Bitterness Intensities

One of the practical benefits of iBitter-SCM is its ability to quantify the "bitter score" (BS) for peptides. By applying iBitter-SCM to the BTP-CV dataset, the authors identified peptides with particularly high bitterness intensities. The classification threshold was set at 331, meaning peptides with a BS greater than 331 are classified as bitter.

The following are the results from Table 5 of the original paper:

Table 5 lists the top twenty peptides with the highest bitter scores. All of the top ten peptides exhibit bitter scores substantially above the threshold of 331, confirming their bitter nature. Notably, peptides like PF (BS=1000.00), RPF (BS=839.50), and GF (BS=823.00) are among the highest. The $log(1/T)$ values (bitter-tasting threshold concentration) from experimental literature, where higher values indicate stronger bitterness, generally align with the calculated bitter scores. For instance, RPFF has a very high $log(1/T)$ of 4.4, consistent with its high BS of 773.33. The inclusion of GG (Gly-Gly) which is noted as "tasteless" but has a high BS (773.00) is an interesting point, potentially indicating an overprediction or that GG can contribute to bitterness in specific contexts. However, the overall trend supports the model's ability to identify peptides with strong bitter taste.

6.1.6. Analysis of Bitter Peptides Using Propensity Scores of Amino Acids and Dipeptides

The interpretability of iBitter-SCM allows for an in-depth analysis of the amino acid and dipeptide contributions to bitterness using their propensity scores. These scores were derived from the optimal opti-DPS from experiment #5.

The following are the results from Table 6 of the original paper:

The propensity scores for amino acids (Table 6) reveal that Gly, Phe, Pro, Glu, and Asp are the top-five amino acids with the highest propensity for bitterness. Conversely, Tyr, Ser, Lys, Ala, and Gln have the lowest scores, suggesting they are less associated with bitterness or even act as attenuators.

A key observation from Table 6 is the high rank of Phe (2) and Pro (3). The paper highlights previous studies confirming the critical role of Phe and Pro in enhancing bitterness. For instance, Ishibashi et al. [22] showed that oligopeptides containing Phe (e.g., FG, FV, FIV, FPF) produced bitter tastes, with FPF showing very strong bitterness. Similarly, Pro residues, while sometimes associated with sweet taste, consistently contributed to bitterness in di- and triproline and peptides like PFP/FPF.

Hydrophobic amino acids are generally considered important for bitterness. The analysis identified Gly, Phe, Pro, Ile, Cys, Val, and Leu as highly informative hydrophobic amino acids (7 out of top 10). This aligns with the understanding that bitterness often arises from hydrophobic interactions with taste receptors.

Regarding dipeptide propensity scores (Fig. 2 heatmap and Table S2, which is mentioned but not provided in full content), the top-ranked dipeptides in bitter peptides included PF, IS, QL, DP, GF, NA, LE, GP, GG, and YV. Conversely, LP, YI, PN, LV, RK, TF, LK, ES, HS, and WM were among the top-ranked in non-bitter peptides. This further pinpoints specific dipeptide patterns contributing to or detracting from bitterness.

The low propensity score for Ala (ranked 19) is consistent with its low percentage in bitter peptides (2.901%) compared to non-bitter peptides (8.288%), suggesting it is not a significant contributor to bitterness.

6.1.7. Analysis of Bitter Peptides Using Informative Physicochemical Properties

To understand the fundamental properties driving bitterness, the authors correlated the amino acid propensity scores with physicochemical properties (PCPs) from the AAindex database.

The following are the results from Table 7 of the original paper:

Table 7 highlights three PCPs with strong Pearson correlation coefficients (R) to the amino acid propensity scores:

1. PONP800104 (R = 0.495): Described as "Surrounding hydrophobicity in alpha-helix". This PCP relates to the hydrophobic environment of amino acid residues.

2. MEIH800103 (R = 0.403): Not explicitly described in the paper, but its context suggests it's related to hydrophobicity.

3. COWR900101 (R = 0.396): Described as "Hydrophobicity index".

   The high correlation with these PCPs strongly indicates that hydrophobicity is a primary factor influencing bitterness. The paper explicitly mentions other hydrophobicity-related PCPs (e.g., WILM950101, EISD860103) also showing importance. This confirms previous experimental findings that hydrophobic amino acid residues (e.g., Phe, Pro, Ile, Arg, Val, Leu, Tyr) play a crucial role in bitterness [70]. The mechanism likely involves hydrophobic interactions with bitter taste receptors at specific binding and stimulating units [71].

6.1.7.1. Importance of Hydrophobic Amino Acid Residue for the Manifestation of Bitterness

The analysis strongly supported the notion that hydrophobic amino acids are key to bitterness. PCPs related to hydrophobicity (e.g., PONP800104, WILM950101, EISD860103, COWR900101) showed the strongest correlations with amino acid propensity scores. This aligns with the binding (BU) and stimulating (SU) unit theory for bitterness, where hydrophobicity drives interaction with the receptor's recognition zone [71]. The paper cited instances where Gly, initially associated with sweet or tasteless properties, formed bitter di- and tripeptides when combined with hydrophobic amino acids like Pro, Val, Ile, Leu, Phe, and Tyr. Furthermore, increasing the number of hydrophobic amino acids (e.g., diPhe, triPhe) led to stronger bitterness, even surpassing that of caffeine. These findings underscore hydrophobicity as a fundamental property governing bitter taste.

6.1.7.2. Importance of Hydrophobic Amino Acids Located at the C-terminus for Determining Bitterness Intensity

Beyond just the presence of hydrophobic amino acids, their position within the peptide sequence also matters. The paper discussed the long-standing assumption that hydrophobic amino acids located at the C-terminus (carboxyl-terminal end) exhibit higher bitterness compared to those at the N-terminus (amino-terminal end). Experimental studies by Ishibashi et al. [20,26,65,66] specifically investigated this. For example, dipeptides like FG, FF, and FV showed more intense bitterness when Phe was at the C-terminus (GF, FF, VF) rather than the N-terminus (FG, FF, FV). The Rcaf (ratio of bitterness to caffeine) values for FG and FV (C- vs N-terminus) clearly demonstrated this positional effect (e.g., $(0.83 vs 0.17)$ for FG). Similarly, oligopeptides like RPFF and RRPFF with Phe at the C-terminus showed exceptionally high bitterness. QSAR modeling by Xu and Chung [17] also identified N1-T-3 and N2HESH-2 as top informative variables describing C-terminal amino acid hydrophobicity, further supporting its critical role.

6.1.7.3. The Number of Carbon Atoms on the Amino Acid Side Chain Affects the Intensity of Bitterness

The structure of the amino acid side chain, particularly the number of carbon atoms and its branching pattern, was identified as another crucial factor.

• Carbon Count: Gly (no carbon in side chain), Ala (one carbon), and aminobutyric acid (Abu) (two carbons) often resulted in tasteless or sweet peptides. However, Val (three carbons) exhibited a mixed bitter and sweet taste, with bitterness more apparent when Val was at the C-terminus. Pro (three carbons in a ring structure) also showed both bitter and sweet tastes but predominantly contributed to bitterness in many peptides. Amino acids with larger side chains (four or more carbons) like Leu, Ile, Phe, and Tyr consistently formed bitter peptides when combined with Gly. This suggests a threshold effect where a certain number of carbons in the side chain is necessary for strong bitterness.
• Linear vs. Branched Chains: Studies comparing synthetic normal Val (n-Val) (linear propyl chain) with Val (isopropyl/branched chain) showed that n-Val peptides produced greater bitterness. Similarly, normal Leu (n-Leu) (linear butyl chain) peptides had higher bitterness intensity than Leu and Ile (branched butyl chains). This implies that linear hydrophobic side chains are more effective at eliciting bitterness than branched ones, likely due to better fitting into the bitter receptor's binding pocket.

6.2. Data Presentation (Tables)

All tables were transcribed directly in the sections above, providing a comprehensive view of the experimental results and analyses.

6.3. Ablation Studies / Parameter Analysis

The paper does not explicitly perform traditional ablation studies (e.g., removing a specific feature type to see performance drop). However, the comparison between init-DPS and opti-DPS (Table 3) serves as an indirect form of parameter analysis, demonstrating the crucial role of the Genetic Algorithm optimization step in refining the propensity scores and improving model performance. The ten experiments with GA (Table 2) also show the variability and optimization process, indicating that the selection of the optimal opti-DPS is critical. The analysis of propensity scores for amino acids and dipeptides itself is a form of intrinsic parameter analysis, as it highlights the learned weights (scores) and their importance.

7. Conclusion & Reflections

7.1. Conclusion Summary

The iBitter-SCM model marks a significant advancement in the computational prediction and characterization of bitter peptides. By introducing the first sequence-based model using the scoring card method (SCM) and dipeptide propensity scores, the authors successfully addressed the limitations of previous QSAR and ML approaches that often relied on complex structural information or lacked interpretability. iBitter-SCM demonstrated high predictive accuracy (84.38%) and MCC (0.688) on an independent dataset, outperforming conventional machine learning classifiers. Beyond prediction, the model's inherent interpretability allowed for a detailed analysis of amino acid and dipeptide propensity scores, revealing critical insights into the biophysical and biochemical drivers of bitterness, such as the paramount importance of hydrophobicity, the specific influence of C-terminal hydrophobic amino acids, and the role of side chain length and branching. The provision of a user-friendly web server further enhances the utility of iBitter-SCM, making it a valuable tool for accelerating drug development and nutritional research by facilitating high-throughput prediction and de novo design of bitter peptides.

7.2. Limitations & Future Work

The paper implicitly highlights some limitations and suggests future directions:

• Dataset Size and Diversity: While the BTP640 dataset was carefully curated, it consists of 640 peptides. The generalizability of the model could potentially be further enhanced with larger and more diverse datasets, especially for non-bitter peptides, which were partly generated randomly from BIOPEP due to scarcity. The reliance on manually collected experimental data is a common constraint in biological prediction tasks.

• Genetic Algorithm (GA) Variability: The GA used for opti-DPS optimization is non-deterministic. Although the paper conducted ten experiments and selected the most robust opti-DPS, this implies a degree of variability in the optimization process. Further research could explore more stable or globally optimal optimization techniques.

• Specific Bitter Receptor Interaction: While iBitter-SCM predicts general bitterness, it does not explicitly model interactions with specific bitter taste receptors (TAS2Rs). Future work could potentially integrate receptor-specific data if available to develop more nuanced prediction models.

• De Novo Design Guidance: Although the propensity scores provide valuable guidance for de novo design, translating these scores into concrete, novel peptide sequences with desired bitterness profiles might still require further iterative computational and experimental validation.

  The authors anticipate that iBitter-SCM will serve as an important tool to facilitate the high-throughput prediction and de novo design of bitter peptides, implying these as key areas of application and future utility.

7.3. Personal Insights & Critique

This paper presents a highly valuable contribution to the field of peptide research, especially for its emphasis on interpretability. In an era where many machine learning models are perceived as "black boxes," iBitter-SCM stands out by providing clear, biologically meaningful propensity scores that directly explain the contribution of each amino acid and dipeptide to bitterness. This level of interpretability is crucial for experimental scientists who need to understand why a peptide is bitter or how to design one that is not.

The choice of dipeptide composition as a feature, combined with the SCM, is elegant in its simplicity and effectiveness. It avoids the complexity of 3D structural modeling or extensive physicochemical descriptor computation, making the method practical and efficient. The rigorous benchmarking against BLAST and a suite of conventional ML classifiers, particularly on an independent test set, strengthens the credibility of iBitter-SCM's performance claims.

A potential area for improvement or future exploration could be to investigate the impact of peptide length more explicitly. While dipeptide composition intrinsically handles varying lengths, the scoring card method might be extended to incorporate tripeptide or k-mer propensities if longer-range interactions are deemed significant, though this would dramatically increase feature dimensionality. Additionally, while the negative dataset generation method is standard, its synthetic nature might slightly limit the diversity compared to a purely experimentally validated set, if one were available.

The application of this method could extend beyond bitterness. The SCM framework, with its focus on propensity scores, could be adapted to predict other peptide properties (e.g., antimicrobial activity, antigenicity, binding affinity) in a similarly interpretable manner, provided suitable datasets are available. The web server is a commendable practical outcome, ensuring accessibility and utility for the broader scientific community. Overall, iBitter-SCM represents a well-executed and impactful piece of research that effectively combines computational rigor with practical utility and interpretability.