🧬 biology

ProteoKnight: Image-Based Encoding for Phage Virion Protein Classification and Uncertainty Estimation

This paper introduces ProteoKnight, an image-based encoding method that adapts the DNA-Walk algorithm for accurate binary classification of Phage Virion Proteins and incorporates Monte Carlo Dropout to quantify prediction uncertainty, achieving competitive accuracy while revealing how protein class and sequence length influence confidence levels.

Original authors: Md. Ishrak Khan, Abir Ahammed Bhuiyan, Samiha Afaf Neha, Md. Khalilur Rahman, Jannatun Noor

Published 2026-11-05

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Md. Ishrak Khan, Abir Ahammed Bhuiyan, Samiha Afaf Neha, Md. Khalilur Rahman, Jannatun Noor

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

In the microscopic world of biology, viruses known as bacteriophages are the most abundant life forms on Earth, though they only infect bacteria. These tiny invaders are built from a core of genetic material wrapped in a protective shell made of proteins. Scientists call these structural building blocks phage virion proteins. Identifying which proteins serve as the shell and which do not is a critical task for researchers. Knowing the structure of these proteins helps in developing new treatments for bacterial infections, understanding how bacteria evolve, and even managing food safety. However, as modern technology allows scientists to read the genetic code of millions of these viruses at once, the sheer volume of data has outpaced the ability to manually sort through it. Traditional methods of identifying these proteins are slow, expensive, and often struggle when the genetic sequences are highly variable. To keep up, researchers are turning to computers to learn how to recognize these patterns automatically, treating the genetic code not just as a list of letters, but as a pattern that can be visualized and analyzed.

A team of researchers from BRAC University and United International University in Bangladesh has developed a new way to help computers understand these viral proteins. They created a tool called ProteoKnight, which transforms the linear sequence of amino acids that make up a protein into a visual image. Instead of feeding the computer a long string of text, the researchers translate each amino acid into a specific colored dot and arrange them on a grid based on a set of rules. This process, which they call "Knight Encoding," is inspired by an older method used for DNA sequences. In this new system, the computer starts at the center of an image and moves outward, placing a colored point for each amino acid in the sequence. The direction of the move depends on the type of amino acid, and the distance is fixed. If the path hits the edge of the image, it resets to the center and continues. The result is a unique, colorful map for every protein, where the arrangement of dots reveals the underlying structure of the molecule.

The researchers tested this visual approach by feeding these protein maps into powerful computer programs known as convolutional neural networks. These programs are already experts at recognizing patterns in photographs, and the team adapted them to recognize the patterns in these protein maps. They trained the system on a massive dataset containing over 35,000 examples of phage proteins and nearly 47,000 examples of non-phage proteins. When asked to simply decide whether a protein was a phage virion protein or not, the system performed with remarkable success, correctly identifying the proteins about 89 percent of the time. This level of accuracy is competitive with the best existing methods, yet it was achieved using a simpler, more efficient computer model that requires less computing power. The study also showed that while the system is very good at this binary choice, it is less certain when asked to sort the proteins into eight different specific categories, such as distinguishing between the tail of the virus and its head.

Beyond just getting the answer right, the researchers were interested in knowing how sure the computer was about its answer. In many real-world applications, a correct answer given with low confidence can be just as risky as a wrong one. To measure this, the team used a technique that runs the same protein image through the computer many times, each time with a slightly different internal setting. By watching how much the computer's answer fluctuates between these runs, they could calculate a score for uncertainty. They found that the computer was generally more confident when looking at shorter protein sequences and when identifying non-phage proteins. Conversely, it showed more hesitation and variability when analyzing longer sequences or actual phage proteins. This suggests that while the system is a powerful tool, there are specific types of data where it needs extra caution. The researchers noted that the visual encoding method sometimes causes dots to overlap in longer sequences, which might confuse the computer and lead to that lower confidence.

The work by the ProteoKnight team offers a fresh perspective on how to study viral proteins. By turning text into images, they have opened a new door for using visual recognition tools in biology. Their findings suggest that this method is a highly effective way to quickly sort through vast amounts of genetic data to find the structural proteins that are essential for the life of a virus. While the system is not yet perfect at distinguishing every single type of protein structure, it provides a strong foundation for future improvements. The ability to not only classify these proteins but also to measure the confidence of that classification gives scientists a new way to judge the reliability of their data. As the field of phage therapy and bacterial research continues to grow, tools like this will be essential for turning the deluge of genetic information into actionable knowledge about the microscopic world.

Technical Summary of "ProteoKnight: Image-Based Encoding for Phage Virion Protein Classification and Uncertainty Estimation"

Problem Statement
Accurate prediction of Phage Virion Proteins (PVPs) is critical for genomic studies, drug synthesis, and disease diagnosis, yet the annotation of phage proteins remains a bottleneck due to low sequence conservation and the high cost of traditional experimental methods. While machine learning offers a computational alternative, existing approaches face two primary limitations:

Encoding Inefficiency: Traditional methods often rely on one-dimensional feature extraction (e.g., k-mer frequencies) or image-based representations like Frequency Chaos Game Representation (FCGR). The authors note that FCGR-based encodings, while effective, tend to lose spatial information inherent in the sequence due to compact transformations.
Lack of Uncertainty Quantification: Deep learning models often produce overconfident predictions, particularly on unseen data. In safety-critical biological applications, there is a lack of research regarding prediction uncertainty in PVP classification, leaving researchers without a metric to gauge model reliability for specific sequence types.

Methodology
The study introduces ProteoKnight, a framework comprising a novel encoding strategy, a deep learning classification pipeline, and an uncertainty estimation module.

Knight Encoding: The core innovation is an image-based encoding technique adapted from the DNA-Walk algorithm. Unlike FCGR, which relies on k-mer frequencies, Knight Encoding treats protein sequences as a 2D random walk.
- Mechanism: Each of the 20 amino acids is assigned a unique color and a specific angle based on a 20-sided polygon (Icosagon) where vertices are separated by 18°.
- Process: Starting from the center of an $M \times M$ image, the algorithm plots a colored point for each amino acid. The position of the next point is calculated using polar coordinates ( $x = r \cos \theta$ , $y = r \sin \theta$ ) relative to the previous point, with a fixed radius step. If a point exceeds the image boundary, the walk resets to the center. This preserves the sequential order and spatial relationships of the amino acids.
Classification Model: The encoded images are fed into pre-trained Convolutional Neural Networks (CNNs). The study evaluates several architectures, including GoogLeNet, EfficientNetV2, and MobileNetV3, utilizing transfer learning to fine-tune these models on the encoded dataset.
Uncertainty Estimation: To quantify prediction confidence, the authors employ Monte Carlo Dropout (MCD). During inference, the dropout layer is kept active (with a 20% rate), and the model performs multiple stochastic forward passes ( $T$ $T$ ) for each input.
- Metrics: Prediction uncertainty is measured using the variance of the softmax probabilities and entropy across the $T$ samples. High variance or entropy indicates low confidence (epistemic uncertainty).
- Analysis: The study categorizes sequences by class (PVP vs. non-PVP) and length (short vs. long) to analyze how these factors influence model confidence.

Key Results

Binary Classification Performance: The proposed method achieved competitive results in binary classification (PVP vs. non-PVP).
- GoogLeNet: Achieved approximately 88.6% accuracy, with an F1-score of 89.87% for PVPs and 89.0% for non-PVPs.
- EfficientNetV2: Achieved a slightly higher accuracy of 91.2%.
- These results are comparable to state-of-the-art methods like PhaVIP (which uses FCGR and Vision Transformers) but utilize parameter-efficient CNNs.
Multi-class Classification: Performance was suboptimal, with accuracy ranging between 72% and 78%. The authors attribute this to the encoding method, where amino acid points may converge or overlap at the same location in the image, causing spatial ambiguity that is more detrimental in complex multi-class scenarios than in binary classification.
Uncertainty Findings:
- The model exhibited lower confidence (higher variance) when predicting PVP sequences compared to non-PVP sequences.
- Sequence Length: Longer sequences generally showed higher prediction variance (lower confidence) compared to shorter sequences.
- Entropy Analysis: Entropy values were lower for non-PVP and short sequences, corroborating the variance findings. High-variance samples (indicating low confidence) were identified, suggesting these specific sequences require further manual assessment before annotation.

Significance and Claims
The paper claims that ProteoKnight offers a novel, image-based alternative to existing sequence encoding methods that successfully mitigates the spatial information loss associated with FCGR. By adapting the DNA-Walk algorithm for proteins, the study demonstrates that 2D image representations can effectively capture features necessary for high-accuracy binary classification using standard CNNs.

Furthermore, the study pioneers the application of uncertainty quantification in PVP classification. It establishes that prediction confidence is not uniform across all protein types; specifically, PVPs and longer sequences are associated with higher uncertainty. This insight provides a validation standard for researchers, highlighting that while a model may achieve high accuracy, certain sequence types (e.g., long PVPs) may yield low-confidence predictions that necessitate additional scrutiny. The authors conclude that while the multi-class performance requires refinement to address point-overlap issues, the combination of Knight Encoding and uncertainty analysis offers a robust, efficient tool for annotating phage proteins in computational biology.

More like this