A new machine-learning method can spot the structural proteins that build a virus's shell with over 90 percent accuracy, and its makers also measured how much to trust its answers.
The method, called ProteoKnight, converts protein sequences into images by adapting the classical DNA-Walk algorithm, using pixel color and adjusted walk distances to preserve spatial patterns that older encodings lose. Those images are then fed into pre-trained convolutional neural networks for classification. In binary tests, sorting phage virion proteins from non-virion ones, the system hit 90.8% accuracy, competitive with existing state-of-the-art tools. When asked to sort proteins into more specific classes, accuracy dropped off. The researchers also ran a Monte Carlo Dropout analysis to flag which predictions the model was actually confident about, finding that confidence swings depending on protein class and sequence length.
This matters because bacteriophages, viruses that infect bacteria, are getting renewed attention as a tool against antibiotic-resistant infections, and identifying their structural proteins from raw sequencing data is a bottleneck in that research. ProteoKnight's image encoding is explicitly built to fix a known weakness in frequency chaos game representation, a prior encoding method that throws away spatial information in the sequence. Flagging low-confidence predictions, rather than just outputting a number, also gives researchers a way to know when to double-check the model instead of trusting it blindly.
Still, a method that only manages one class distinction reliably is a narrower win than the framing suggests, a reminder that genomic AI tools tend to work best on the easy binary questions and get shakier as the real biology gets messier.