🤖 machine learning

ProteoKnight: Convolution-based Phage Virion Protein Classification and Uncertainty Analysis

The paper introduces ProteoKnight, a novel image-based encoding method that adapts the DNA-Walk algorithm for protein sequences to achieve competitive accuracy in classifying Phage Virion Proteins using pre-trained convolutional neural networks while also evaluating prediction uncertainty through Monte Carlo Dropout.

Original authors: Samiha Afaf Neha, Md. Ishrak Khan, Abir Ahammed Bhuiyan

Published 2026-08-18

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Samiha Afaf Neha, Md. Ishrak Khan, Abir Ahammed Bhuiyan

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Life on Earth relies on a vast, invisible machinery built from proteins, the molecular workhorses that construct and operate every living cell. Among the most abundant biological entities are bacteriophages, often simply called phages, which are viruses that specifically infect bacteria. These tiny invaders are not just biological curiosities; they are potential keys to fighting antibiotic-resistant infections and reshaping our understanding of microbial ecosystems. However, to harness them, scientists must first understand their structure, particularly the proteins that make up the phage's outer shell, known as virion proteins. Identifying these proteins is a critical step, yet it is a difficult task because the genetic instructions for building them are often short, highly variable, and lack the clear patterns that traditional computer programs look for. As the amount of genetic data from these viruses explodes, the need for fast, accurate ways to sort and identify these specific proteins has become urgent, driving researchers to find new ways to teach computers how to see what the human eye cannot.

In a recent study, a team of researchers from Bangladesh tackled this challenge by developing a new method called ProteoKnight. Their goal was to create a system that could accurately distinguish between the structural proteins of a phage and other types of proteins found in the same genetic data. The core of their innovation lies in how they translate a string of genetic letters into a picture that a computer can analyze. Instead of treating the protein sequence as a simple list of characters, they converted it into a visual map. They imagined the sequence as a journey across a grid, where each type of amino acid—the building block of a protein—corresponds to a specific step in a particular direction and a specific color. As the computer reads the sequence, it draws a path, placing colored dots on a canvas to form a unique image for every protein. This approach, which they named "Knight Encoding," preserves the spatial order of the sequence, ensuring that the relationship between one part of the protein and another remains visible in the final picture, a detail that some older methods had lost.

To test this new visual language, the researchers fed these generated images into powerful, pre-trained computer vision systems. These systems, originally designed to recognize objects in photographs, were fine-tuned to recognize the patterns within the protein maps. The results were striking. When asked to perform a binary classification—determining simply whether a protein was a phage structural protein or not—the system achieved an accuracy of roughly 90 percent. This performance matched or exceeded the capabilities of the most advanced tools currently available, proving that turning a biological sequence into an image is a viable and powerful strategy. The team found that the method worked particularly well for shorter sequences and for proteins that were not part of the phage structure, suggesting that the visual patterns were distinct and learnable.

However, the researchers did not stop at simply measuring success; they also wanted to understand how confident the computer was in its answers. In the world of artificial intelligence, a model can sometimes be very wrong while feeling very sure of itself. To address this, the team employed a technique that introduces a small amount of randomness into the decision-making process, allowing them to measure the stability of the prediction. They discovered that the computer's confidence varied depending on the length of the protein and its specific type. For longer sequences, the model showed more uncertainty, likely because the visual map became crowded with overlapping dots, making the pattern harder to distinguish. This finding is crucial because it highlights exactly where the system might need extra human attention, providing a safety net for future applications where reliability is paramount.

While the new method excelled at the binary task of separating phage proteins from non-phage proteins, the researchers noted that it faced more challenges when asked to sort the proteins into specific categories, such as distinguishing between the head of the virus and its tail. In these more complex scenarios, the accuracy dropped to a moderate level, indicating that while the visual encoding is a strong foundation, it still has room for refinement. The study concludes that ProteoKnight offers a significant step forward in the field, providing a fast and efficient way to annotate phage proteins. By bridging the gap between genetic sequences and visual recognition, and by honestly measuring the limits of its own confidence, this work offers a robust tool for scientists aiming to unlock the secrets of bacteriophages and their potential to revolutionize medicine and environmental science.

Technical Summary: ProteoKnight

Problem Statement
Accurate prediction of Phage Virion Proteins (PVPs) is critical for genomic studies, drug synthesis, and disease diagnosis, yet it remains challenging due to the high diversity and low sequence conservation of bacteriophage proteins. Traditional alignment-based methods often fail due to horizontal gene transfer and high mutation rates, while existing alignment-free machine learning approaches frequently rely on one-dimensional feature extraction (e.g., k-mer frequencies) that may discard valuable spatial information. Furthermore, while deep learning has improved classification accuracy, there is a significant gap in understanding prediction uncertainty, particularly in safety-critical biological applications where overconfident predictions on unseen data can be detrimental.

Methodology
The study introduces ProteoKnight, a framework comprising a novel image-based encoding technique and an uncertainty analysis pipeline.

Knight Encoding:
- Adapting the classical DNA-Walk algorithm, the authors developed a "Knight Encoding" strategy to convert 1D protein sequences into 2D RGB images.
- Mechanism: Each of the 20 amino acids is mapped to a unique color and assigned an angle on a 20-sided polygon (Icosagon), with vertices separated by $18^\circ$ .
- Process: Starting from the image center, the algorithm iteratively plots points based on polar coordinates. The displacement for each amino acid is calculated using $x = r \times \cos(\theta)$ and $y = r \times \sin(\theta)$ , where $r$ is a fixed radius (15 pixels) and $\theta$ is the amino acid's assigned angle. If a point hits the image boundary, the walk resets to the center.
- Goal: This approach aims to preserve spatial sequence information that is often lost in frequency-based encodings like Frequency Chaos Game Representation (FCGR).
Classification Architecture:
- The encoded images were classified using pre-trained Convolutional Neural Networks (CNNs) via transfer learning to address data scarcity and computational efficiency.
- Models evaluated included GoogLeNet, EfficientNet, and MobileNet.
- The study focused on both binary classification (PVP vs. non-PVP) and multi-class classification (8 specific PVP structural categories).
Uncertainty Analysis:
- The authors employed Monte Carlo Dropout (MCD) to quantify prediction uncertainty. By activating dropout layers during inference and performing multiple stochastic forward passes, they calculated the variance and entropy of the prediction distributions.
- Uncertainty was analyzed across different sequence lengths (short vs. long) and classes (PVP vs. non-PVP) to identify data points where the model exhibits low confidence despite potential accuracy.

Key Results

Binary Classification: ProteoKnight achieved competitive performance, with GoogLeNet reaching approximately 88.6% accuracy (91.1% sensitivity, 92.7% specificity) and EfficientNet v2 achieving ~91.2% accuracy. These results are comparable to state-of-the-art methods like PhaVIP and DeePVP.
Multi-class Classification: Performance was suboptimal, with accuracy ranging between 72% and 78%. The authors attribute this to the encoding method's limitation where amino acid points may overlap at the same coordinates, causing spatial information loss in complex multi-class scenarios.
Uncertainty Findings:
- The model demonstrated higher confidence (lower variance) when predicting non-PVP sequences compared to PVP sequences.
- Shorter sequences consistently showed lower prediction variance (higher confidence) compared to longer sequences.
- Entropy analysis corroborated these findings, with lower entropy values observed for non-PVP and shorter sequences.

Significance and Contributions
The paper claims three primary contributions:

Novel Encoding: The introduction of "Knight Encoding" provides an alternative to FCGR, specifically designed to mitigate spatial information loss in protein sequence image representation.
Efficient Classification: The study demonstrates that parameter-efficient, pre-trained CNNs combined with this new encoding can achieve high-accuracy binary PVP classification without the computational cost of training models from scratch.
Uncertainty Quantification: The research bridges a gap in protein classification by applying MCD to identify sequence types (specifically long PVPs) that yield high accuracy but low confidence. This provides a validation standard for researchers to flag sequences requiring further manual assessment before annotation.

Limitations and Future Directions
The authors modestly acknowledge that while binary classification is robust, multi-class performance requires improvement. They suggest that the point-overlap issue in the current encoding scheme is a primary bottleneck. Future work is proposed to extend the encoding to higher dimensions (e.g., separating frames for each amino acid) or optimizing hyperparameters (radius, dot size) to prevent information loss in longer sequences. The study positions ProteoKnight not as a final solution, but as a complementary tool that offers a fresh perspective on sequence classification and uncertainty in bio-engineering and phage therapy applications.

Technical Summary: ProteoKnight

More like this