SCMTHP: A New Approach for Identifying and Characterizing of Tumor-Homing Peptides Using Estimated Propensity Scores of Amino Acids

Phasit Charoenkwan, Wararat Chiangjong, Chanin Nantasenamat, Mohammad Ali Moni, Píetro Lió, Balachandran Manavalan, Watshara ShoombuatongPharmaceutics 2022 · DOI

TL;DR

This paper introduces SCMTHP, a novel and interpretable computational method for identifying and characterizing tumor-homing peptides (THPs). By leveraging a scoring card method and generating amino acid propensity scores, SCMTHP not only achieves competitive accuracy but also provides valuable insights into the biophysical properties that drive THP activity, making it a useful tool for cancer therapy research.

What Problem Does This Paper Address?

Tumor-homing peptides (THPs) are promising candidates for targeted cancer diagnosis and therapy due to their ability to specifically bind to cancer cells. However, accurately identifying and understanding the mechanisms behind their tumor-homing capabilities is crucial for their effective development. While existing computational methods exist, a significant drawback is their lack of interpretability, making it difficult for researchers to understand why a peptide is predicted to be a THP. This hinders the rational design of new and improved THPs. This paper aims to bridge this gap by proposing a method that is both accurate and interpretable, offering insights into the underlying biochemical and biophysical properties that govern THP function.

Methods

The core of this work is the development of the SCMTHP (Scoring Card Method for Tumor-Homing Peptides) approach. It builds upon the Scoring Card Method (SCM), a technique known for its simplicity and interpretability. The key innovation lies in:

  • Amino Acid Propensity Scores: Instead of relying solely on traditional sequence features, SCMTHP generates "propensity scores" for each of the 20 standard amino acids. These scores represent how likely each amino acid is to contribute to a peptide's tumor-homing activity. These scores are optimized using a genetic algorithm (GA) to maximize predictive performance.
  • Weighted-Sum Classifier: SCMTHP uses a simple weighted-sum function to calculate a "THP score" for any given peptide. This score is then compared against a threshold to classify the peptide as either a THP or a non-THP. This simplicity contributes to the method's interpretability.
  • Physicochemical Property (PCP) Analysis: The generated propensity scores are then used to identify informative physicochemical properties of amino acids that are strongly correlated with tumor-homing activity. This allows researchers to gain a deeper understanding of why certain peptides exhibit this behavior.

Datasets

The study utilizes two benchmark datasets previously established for THP identification:

  • Main Dataset: Comprises 1302 sequences (651 THPs and 651 non-THPs).
  • Small Dataset: A subset of the Main dataset, containing peptides ranging from 4 to 10 residues in length (938 sequences: 469 THPs and 469 non-THPs).

For evaluation, these datasets were split into training (Main-TRN, Small-TRN) and independent test sets (Main-IND, Small-IND). The paper notes that these datasets are publicly available.

Key Insights and Ideas

  • Interpretability is Key: The paper strongly emphasizes that interpretability is as important as accuracy. By providing propensity scores for individual amino acids, SCMTHP offers a window into the "decision-making" process of the model, which is crucial for guiding peptide design.
  • Cysteine's Role: The analysis suggests that cysteine (Cys) residues may play a significant role in the stabilization and function of THPs, potentially through disulfide bond formation. This is a concrete, actionable insight for peptide engineers.
  • Refractivity as a Predictor: The study highlights "Refractivity" as a physicochemical property strongly correlated with THP activity, suggesting that this property might influence peptide accumulation in tumor cells.
  • Simplicity Can Be Powerful: The success of SCMTHP, a relatively simple weighted-sum approach, demonstrates that complex machine learning models are not always necessary for achieving high performance, especially when interpretability is a priority.

Key Results

SCMTHP demonstrated strong performance across various evaluations:

  • Accuracy: Achieved accuracies of 0.827 on the Main-IND dataset and 0.798 on the Small-IND dataset, comparable to state-of-the-art methods.
  • Outperformance: SCMTHP outperformed several conventional machine learning classifiers (e.g., decision tree, k-nearest neighbor, multi-layer perceptron, naive Bayes, and partial least squares regression) in terms of accuracy, cost-effectiveness, and simplicity.
  • Web Server: A user-friendly online web server for SCMTHP has been developed and made freely available, facilitating its use by the research community.
  • Physicochemical Property Insights: The analysis identified key amino acids and physicochemical properties (like refractivity and extinction coefficients) that are important for THP activity, providing valuable biological context.

Limitations and Future Work

  • Dataset Size: While the datasets are standard, the authors acknowledge the need for more extensive datasets to further validate and improve the model.
  • Generalizability: The study focuses on THPs. Future work could explore extending this approach to other types of therapeutic peptides.
  • Experimental Validation: While the method provides computational insights, experimental validation of the predicted THPs and the identified key properties would be a crucial next step.
  • Further PCP Exploration: The paper suggests that further investigation into the relationship between SCMTHP-derived propensity scores and other physicochemical properties could yield even deeper insights into THP mechanisms.

Summary updated

Abstract

Tumor-homing peptides (THPs) are small peptides that can recognize and bind cancer cells specifically. To gain a better understanding of THPs' functional mechanisms, the accurate identification and characterization of THPs is required. Although some computational methods for in silico THP identification have been proposed, a major drawback is their lack of model interpretability. In this study, we propose a new, simple and easily interpretable computational approach (called SCMTHP) for identifying and analyzing tumor-homing activities of peptides via the use of a scoring card method (SCM). To improve the predictability and interpretability of our predictor, we generated propensity scores of 20 amino acids as THPs. Finally, informative physicochemical properties were used for providing insights on characteristics giving rise to the bioactivity of THPs via the use of SCMTHP-derived propensity scores. Benchmarking experiments from independent test indicated that SCMTHP could achieve comparable performance to state-of-the-art method with accuracies of 0.827 and 0.798, respectively, when evaluated on two benchmark datasets consisting of Main and Small datasets. Furthermore, SCMTHP was found to outperform several well-known machine learning-based classifiers (e.g., decision tree, k-nearest neighbor, multi-layer perceptron, naive Bayes and partial least squares regression) as indicated by both 10-fold cross-validation and independent tests. Finally, the SCMTHP web server was established and made freely available online. SCMTHP is expected to be a useful tool for rapid and accurate identification of THPs and for providing better understanding on THP biophysical and biochemical properties.
BibTeX
@inproceedings{charoenkwan2022scmthp,
  author = {Phasit Charoenkwan and Wararat Chiangjong and Chanin Nantasenamat and Mohammad Ali Moni and Píetro Lió and Balachandran Manavalan and Watshara Shoombuatong},
  title = {{SCMTHP: A New Approach for Identifying and Characterizing of Tumor-Homing Peptides Using Estimated Propensity Scores of Amino Acids}},
  booktitle = {Pharmaceutics},
  year = {2022},
  doi = {10.3390/pharmaceutics14010122},
}

Loading related content...

Figure