ProteoKnight: PVP Classifier with Knight Encoding

• ProteoKnight is a PVP classification framework that uses Knight Encoding to convert protein sequences into ordered colored trajectories.
• It performs binary discrimination and multiclass prediction over 8 PVP subclasses using transfer learning with pre-trained convolutional neural networks.
• The framework integrates Monte Carlo dropout for uncertainty analysis, achieving competitive accuracy with efficiency-focused models like GoogLeNet despite challenges in long sequences.

ProteoKnight is a phage virion protein (PVP) classification framework centered on a protein-to-image encoding called Knight Encoding, followed by classification with pre-trained convolutional neural networks and uncertainty analysis using Monte Carlo Dropout. It targets binary discrimination between PVP and non-PVP sequences and multiclass prediction over 8 PVP subclasses. The method is motivated by the claim that an image representation preserving sequential spatial structure better than frequency chaos game representation (FCGR) can yield competitive PVP classification while exposing uncertainty patterns associated with protein class and sequence length (Neha et al., 10 Aug 2025).

1. Biological scope and problem formulation

ProteoKnight is situated in the problem of annotating phage virion proteins, the structural proteins of bacteriophages associated with the head, tail, baseplate, portal, capsid, and related virion components. The paper frames accurate PVP annotation as important for bacteriophage genomics because high-throughput sequencing produces phage sequences faster than they can be experimentally characterized, and because structural protein annotation affects downstream applications including phage therapy, antibacterial discovery, diagnostics, host interaction studies, and broader metagenomic analysis. It also emphasizes that conventional laboratory approaches such as mass spectrometry and protein arrays are expensive and laborious, while alignment-based computational methods are often unreliable for phages because of low sequence conservation, horizontal gene transfer, and lack of collinearity (Neha et al., 10 Aug 2025).

Within that setting, ProteoKnight addresses two linked problems. The first is representational: how to encode protein sequences for deep learning-based PVP classification without losing sequence-order information. The second is inferential: how to quantify prediction uncertainty in a domain where low conservation and curated training data may encourage overconfident but unreliable predictions. The paper therefore treats ProteoKnight not only as a classifier but also as a framework for uncertainty-aware annotation.

The classification tasks are explicitly split into binary classification, defined as PVP versus non-PVP, and multiclass classification, defined over 8 PVP subclasses. The benchmark derives from PhaVIP, using a dataset curated from the RefSeq viral protein database with annotations up to December 2022. The original curation process included removing low-confidence labels, extracting structural proteins via diverse keyword search, collecting non-structural proteins by identifying enzymes with names ending in “ase,” and applying CD-HIT clustering at 90% sequence similarity, keeping the longest sequence per cluster. The final binary dataset contains 35,213 PVP sequences and 46,883 non-PVP sequences. The PVP subset is partitioned into 8 classes: Baseplate 3362, Portal 2770, Tail Fiber 2305, Major Capsid 2443, Minor Capsid 398, Major Tail 5083, Minor Tail 1458, and Others 17,385 (Neha et al., 10 Aug 2025).

2. Knight Encoding

The defining component of ProteoKnight is Knight Encoding, an adaptation of the classical DNA-Walk concept to protein sequences. Instead of compressing sequence information into k-mer frequencies, the method maps a protein into a colored 2D trajectory intended to preserve positional evolution through sequence space. The amino-acid alphabet is fixed as:

$18^\circ$2

These 20 residues are placed on the vertices of an icosagon, so adjacent residues differ by a constant angular increment of

$18^\circ = \frac{360^\circ}{20}.$

If a residue has index $i$ in the amino-acid list, its direction is

$\theta = i \times 18^\circ.$

The paper gives examples such as C at index 1 receiving $18^\circ$, and G at index 5 receiving $90^\circ$ (Neha et al., 10 Aug 2025).

The image is drawn on an $M \times M$ canvas initialized at the center,

$(x,y) = (M/2, M/2).$

For each amino acid, ProteoKnight uses a fixed radius $r=15$ and point size 2, with directional displacement

$x = r \times \cos(\theta) \tag{1}$

$y = r \times \sin(\theta) \tag{2}$

and updates the current position by

$i$0

$i$1

where $i$2 is the residue-specific displacement. A circular point of size 2 is then drawn at the updated coordinates. If the walk reaches an image boundary, the algorithm resets to the center $i$3 and continues encoding from there (Neha et al., 10 Aug 2025).

Each residue is also assigned a fixed RGB color. The paper lists a full mapping, including examples such as A: (255, 0, 0), C: (255, 255, 0), D: (0, 234, 255), E: (170, 0, 255), W: (204, 204, 204), and Y: (0, 64, 255). The resulting image is therefore a colored ordered walk through sequence space rather than a frequency heatmap. The paper’s principal representational claim is that this mitigates the “spatial information loss” associated with FCGR. It also identifies a new limitation: if multiple residues map to exactly the same coordinates during the walk, points can overlap, which can obscure information, especially in longer or more complex sequences and in multiclass classification (Neha et al., 10 Aug 2025).

From an implementation standpoint, the preprocessing is intentionally light. The original encoded image size is 512 × 512, and images are resized to 224 × 224 before CNN input. No explicit normalization, standardization, or data augmentation procedure is described. The method also does not require sequence truncation or padding, since variable-length sequences are naturally represented as longer or shorter trajectories; the sequence-length issue instead reappears through overlap and boundary resets (Neha et al., 10 Aug 2025).

3. Classification architecture and optimization

After encoding, ProteoKnight applies transfer learning with pre-trained convolutional neural networks. The paper states that several pre-trained architectures from the PyTorch library were evaluated, highlighting the top four models and specifically naming the GoogLeNet, EfficientNet, and MobileNet families. Exact identities of all top four are not fully specified in the text, but the main comparison emphasizes GoogLeNet and EfficientNet V2 (Neha et al., 10 Aug 2025).

The training configuration is concise. The reported settings are: 25 epochs, batch size 32, learning rate 0.001, original image size 512 × 512, and CNN input size 224 × 224. The paper states that binary_crossentropy was used as the optimization loss function, although it does not explicitly specify the multiclass loss. Evaluation uses accuracy as the primary metric, together with precision, recall/sensitivity, specificity, and F1-score, with the standard definitions

$i$4

$i$5

$i$6

$i$7

$i$8

The paper presents EfficientNet V2 as the best raw classifier, but selects GoogLeNet as the most practical model because it achieved nearly comparable performance with far fewer parameters. GoogLeNet is reported to have 5.5M parameters, described as roughly one-fourth of EfficientNet V2. This efficiency-performance tradeoff becomes important because the paper couples GoogLeNet with the uncertainty-analysis stage (Neha et al., 10 Aug 2025).

4. Empirical performance

The empirical results distinguish between framework-level best performance and the specific GoogLeNet configuration chosen for practical deployment. The paper reports binary classification accuracy of 90.8% in the abstract and says EfficientNet V2 reached around 91.2%, while GoogLeNet achieved 88.60% accuracy, 89.87% F1, and 88.54% recall on the binary task. For multiclass prediction, GoogLeNet achieved 76.37% accuracy, 76.37% F1, and 76.37% recall. The authors describe multiclass performance as “moderate,” with accuracy in the 72–78% range across models (Neha et al., 10 Aug 2025).

Setting	Reported result	Note
Binary, framework-level	90.8% accuracy	Abstract headline
Binary, EfficientNet V2	around 91.2%	Best raw classifier
Binary, GoogLeNet	88.60% accuracy	Chosen for efficiency
Multiclass, GoogLeNet	76.37% accuracy	8 PVP subclasses

In class-wise binary evaluation, ProteoKnight achieved, for PVP, precision 0.83, recall 0.91, and F1 0.87; for non-PVP, precision 0.94, recall 0.89, and F1 0.93. Compared with prior methods, the paper reports PhANN with PVP F1 0.84 and non-PVP F1 0.83; DeePVP with PVP precision 0.99 but recall 0.51 and PVP F1 0.68, non-PVP F1 0.80; VirionFinder with PVP F1 0.84 and non-PVP F1 0.81; Meta-iPVP with PVP F1 0.80 and non-PVP F1 0.77; PVPred-SCM with PVP F1 0.66 and non-PVP F1 0.74; and PhaVIP with PVP F1 0.94 and non-PVP F1 0.94. ProteoKnight therefore outperformed several older feature-engineered and 1D approaches, but did not surpass PhaVIP on binary classification (Neha et al., 10 Aug 2025).

The paper interprets these results cautiously. Its claim is not that ProteoKnight establishes dominant binary performance across all baselines, but that it offers a conceptually different encoding that preserves sequence progression better than FCGR while remaining competitive in binary PVP prediction. The weaker multiclass results are attributed partly to the overlap behavior of the walk representation, which can obscure class-discriminative information when many residues land on the same coordinates (Neha et al., 10 Aug 2025).

5. Uncertainty analysis with Monte Carlo Dropout

A major secondary contribution of ProteoKnight is uncertainty analysis using Monte Carlo Dropout (MCD), presented in the paper as the first such analysis for PVP/protein sequence classification in this setting. During MCD, dropout remains active at test time, with stochastic retention or removal of weights modeled as

$i$9

and predictive mean and variance across $\theta = i \times 18^\circ.$0 stochastic forward passes given by

$\theta = i \times 18^\circ.$1

$\theta = i \times 18^\circ.$2

For the uncertainty experiments, the authors selected 100 randomly chosen sequences from each of four categories: PVP short, PVP long, non-PVP short, and non-PVP long. Each sequence was predicted 100 times, so the uncertainty study used $\theta = i \times 18^\circ.$3 stochastic forward passes per sample. The class-specific length thresholds were defined as $\theta = i \times 18^\circ.$4 for PVP and $\theta = i \times 18^\circ.$5 for non-PVP, yielding the following training subsets: PVP $\theta = i \times 18^\circ.$6: 12,589, PVP $\theta = i \times 18^\circ.$7: 12,060, non-PVP $\theta = i \times 18^\circ.$8: 16,502, and non-PVP $\theta = i \times 18^\circ.$9: 16,316 (Neha et al., 10 Aug 2025).

Using GoogLeNet, the paper concludes that predictions were more confident for non-PVP than PVP and more confident for short than long sequences. At dropout 0.2, the reported variances were:

• PVP short: 0.08558
• non-PVP short: 0.07555
• PVP long: 0.09119
• non-PVP long: 0.07663

The same trend generally holds at dropout 0.1 and 0.3. Across those settings, the highest uncertainty appears for long PVPs, and the lowest for short non-PVPs. The paper offers two explanations: data imbalance, since there are more non-PVP training sequences than PVPs, and intrinsic sequence complexity, since PVPs may be harder because of lower conservation and more variable structure-related composition. It also links uncertainty in long sequences to the encoding itself, because longer walks create more opportunities for point overlap and boundary resets (Neha et al., 10 Aug 2025).

The paper additionally computes predictive entropy for selected high-variance and low-variance examples using

$18^\circ$0

where $18^\circ$1 is the predicted probability for the class of interest. The reported entropy values are numerically irregular: PVP short high-var 0.1122, low-var 0.1259; non-PVP short high-var 0.1369, low-var 0.1398; PVP long high-var 0.2111, low-var 0.3943; non-PVP long high-var 0.06539, low-var 0.1716. The paper nevertheless draws the broader qualitative conclusion that long PVPs are more uncertainty-prone and should be treated with more caution before annotation. A plausible implication is that ProteoKnight is intended not merely to label proteins, but also to identify cases where automatic annotation should be deprioritized or manually reviewed (Neha et al., 10 Aug 2025).

6. Limitations, interpretation, and relation to adjacent protein-AI systems

The paper identifies several limitations. First, multiclass classification accuracy remains suboptimal relative to binary classification. Second, the walk representation can suffer from point overlap, especially in longer sequences. Third, boundary resets to the image center may distort trajectory continuity. Fourth, there is no detailed ablation isolating the impact of colors, radius, image size, or reset behavior, nor a controlled comparison to FCGR under the same backbone beyond broad qualitative claims. Fifth, reproducibility is partial: hardware and software are reported as Python 3.11, TensorFlow, Keras, PyTorch, running on an AMD Ryzen 7 3700X 8-core CPU, 16 GB DDR4 RAM, NVIDIA GeForce RTX 3060 Ti, and Windows 11 Pro, but the exact train/test split protocol and code availability are not given in the provided text (Neha et al., 10 Aug 2025).

Within the broader protein-AI landscape, ProteoKnight is substantially narrower than systems for protein question answering, ontology-scale annotation, or retrieval-centric protein reasoning. Prot2Chat is a multimodal generative protein question-answering system that integrates sequence, structure, and question text through early fusion and question-conditioned virtual protein tokens, aiming at free-text answers rather than fixed-label classification (Wang et al., 7 Feb 2025). ProtGO is a sequence-only transformer-based fusion model for multi-label Gene Ontology prediction over the top 100 terms in each of the MF, BP, and CC branches, using ProtBert-style transfer rather than image encoding (Tamir et al., 2024). DistNBLP is a distributed graph-based label propagation method for large-scale protein annotation in similarity graphs, emphasizing scalability and reviewed-label transfer rather than learned sequence representations (Aridhi et al., 2017). These comparisons suggest that ProteoKnight belongs to a distinct methodological line: it is an alignment-free, sequence-to-image PVP classifier with explicit uncertainty analysis, not a general protein assistant, ontology predictor, or graph propagation engine.

That specialization clarifies both its value and its limits. ProteoKnight’s core contribution is the proposal that PVP classification can be approached through a colored spatial walk that preserves sequence-order information more directly than FCGR and remains compatible with lightweight transfer-learned CNNs. Its uncertainty analysis extends that proposal into a confidence-aware annotation setting. At the same time, the framework does not attempt multimodal protein reasoning, structure-grounded generation, ontology-aware function prediction, or open-ended retrieval and explanation. In that sense, it is best understood as a focused classifier for phage virion protein annotation whose novelty lies primarily in Knight Encoding and in the integration of Monte Carlo Dropout into the PVP classification workflow (Neha et al., 10 Aug 2025).