Capsule Networks Push Antimicrobial Peptide Prediction Past 97 Percent Accuracy

A deep learning model built around a capsule neural network has reported some of the strongest classification numbers yet for identifying antimicrobial peptides, the short amino acid sequences that many researchers view as promising candidates for the next generation of antibiotics. The model, named AMP-CapsNet, was introduced by Ali Ghulam and colleagues in a study published on 7 February 2026 in the journal Genomics & Informatics. Using a feature encoding scheme called dipeptide composition, the model achieved an accuracy of 97.29 percent and an area under the receiver operating characteristic curve of 98.91 percent on its test set, outperforming convolutional and recurrent neural network baselines as well as several machine learning classifiers.

Antimicrobial peptides, or AMPs, are small, biologically active peptides found across plants, insects, and animals, typically ranging from ten to one hundred amino acids in length. They generally kill microbes by damaging microbial cell membranes, a mechanism that tends to be non-specific rather than targeted at particular proteins. According to the authors, bacteria acquire resistance to AMPs to a lesser extent than to conventional antibiotics for precisely this reason. Some naturally occurring AMPs are already used clinically, and the paper notes reported applications in wound healing, infectious disease management, and the treatment of diabetic foot infections. Screening genomes of diverse organisms for novel AMP sequences is considered a promising route to faster antibiotic discovery, but experimental identification remains slow, laborious, and expensive.

The core technical challenge the researchers addressed is sequence classification: given a peptide sequence, decide whether it is antimicrobial. Conventional biochemical assays such as agar diffusion tests and minimum inhibitory concentration assessments can evaluate extracted peptides, but the paper argues these approaches are costly, slow, limited in the range of peptides they can screen, and constrained by computational and stability issues when simulated. Machine learning offers an alternative, and the team positioned capsule networks, described as a next-generation neural architecture, as particularly well suited to the task because they replace scalar neurons with vectors of neurons, called capsules, whose lengths encode confidence in a prediction.

To turn raw peptide sequences into numerical inputs, the researchers tested two feature encoding methods. Amino acid composition, or AAC, represents a sequence by the frequency of each of the twenty standard amino acids, producing a compact twenty-dimensional vector that indirectly captures physicochemical properties such as polarity and hydrophobicity. Dipeptide composition, or DPC, goes further by counting all four hundred possible adjacent amino acid pairs, thereby preserving local sequence order information that AAC discards. In the proposed fusion scheme, the normalized AAC and DPC vectors were horizontally concatenated into a single 420-dimensional hybrid feature vector that was fed directly into the capsule network architecture.

The AMP-CapsNet pipeline itself consists of three modules. A sequence encoding module splits the raw protein sequence into fragments and encodes each amino acid as numerical vectors. A multi-dimensional feature representation module applies convolutional operations together with a channel attention mechanism to detect hidden multi-scale features. Finally, the capsule network module combines and refines these features into capsule vectors for classification, using a dynamic routing mechanism to update the network’s internal parameters and a margin loss function that penalizes excessive capsule activations when a class is absent. The authors state that this design lets the model capture both local shapes and higher-order interactive structures within the encoded sequence data.

The datasets were drawn from UniProt and a benchmark referenced from earlier work. After screening, the team assembled 1,085 AMP sequences and 1,316 non-AMP sequences, a total of 2,401 data points, with negative samples matched in length to the positives so that sequence length could not serve as a distinguishing feature and duplicates were removed. The model was trained and validated internally with five-fold cross-validation and then evaluated on an external test. Performance was measured with accuracy, sensitivity, specificity, precision, F1-score, Matthews correlation coefficient, and AUC derived from ROC curves plotted across all classification thresholds.

The headline results favor dipeptide composition decisively. On the test set, AMP-CapsNet with DPC reached 97.29 percent accuracy, 96.80 percent sensitivity, 97.70 percent specificity, and an MCC of 94.54 percent, while the AAC version managed 84.42 percent accuracy with 76.20 percent sensitivity and 90.49 percent specificity, and an MCC of 68.58 percent. Training set accuracies were lower, at 88.95 percent for DPC and 81.42 percent for AAC, and the model recorded a training accuracy of 98.09 percent with a validation accuracy of 98.91 percent and a validation loss of 4.76 percent on the DPC-encoded data. The authors attribute the gap between encodings to DPC’s ability to encode sequence correlations, arguing that dipeptide relationships suit the capsule network’s dynamic routing better than the simpler amino acid frequencies.

Comparative experiments reinforced the pattern. Against other deep learning architectures on DPC features, CapsNet posted an AUC of 0.9777, ahead of CNN at 0.9762, BiLSTM at 0.9266, and LSTM at 0.8639. Against classical machine learning classifiers, the picture was more mixed: with AAC features, LightGBM, XGBoost, and AdaBoost each scored ROC-AUC values around 0.95 to 0.96 while CapsNet scored 0.94, but with DPC features CapsNet reached 0.99, ahead of LightGBM at 0.97 and XGBoost and AdaBoost at 0.91.

The team also probed why the model makes its decisions. A SHAP, or Shapley Additive exPlanations, analysis ranked feature importance and identified features 200, 209, and 214 as the most influential, pushing predictions toward the AMP class when their values were high, while features such as 315 and 28 showed SHAP values near zero and little effect on outputs. A t-SNE visualization of the 400-dimensional DPC training features, computed with two components, perplexity 50, and a learning rate of 1000, showed AMPs and non-AMPs as partially separable but with considerable overlap, which the authors interpret as evidence that AMP prediction is not a trivially separable problem and benefits from advanced feature extraction and classification methods.

The authors conclude that CapsuleNet combined with dipeptide composition is the most effective method for AMP classification in their comparisons and may represent the state of the art in this context, positioning the approach as a tool for accelerating deep learning-based AMP drug discovery. They acknowledge room for improvement, planning to add further feature types and fine-tune hyperparameters, and they note that differences in dataset complexity, distribution, and noise caused somewhat weaker performance on the external dataset than on internal validation. The code and data supporting the study have been released in a public GitHub repository, allowing other groups to test whether these laboratory-benchmark gains translate into genuinely faster discovery of usable antimicrobial candidates.

Subject of Research: Deep learning prediction of antimicrobial peptides using capsule networks and multi-view feature fusion

Article Title: AMP-CapsNet: a multi-view feature fusion approach for antimicrobial peptide prediction using capsule networks

Article References: Ghulam, A., Rehman, M., Fida, H., Zhao, P.-Y., Noroze, R., Qi, Y.-C., & Yu, X.-L. (2026). AMP-CapsNet: a multi-view feature fusion approach for antimicrobial peptide prediction using capsule networks. Genomics & Informatics, 24(1), Article 4. https://doi.org/10.1186/s44342-026-00067-6

Image Credits: AI Generated

DOI: 10.1186/s44342-026-00067-6

Keywords: antimicrobial peptides, capsule neural network, deep learning, dipeptide composition, amino acid composition, antibiotic resistance, peptide classification, SHAP analysis, machine learning, drug discovery, bioinformatics, Genomics & Informatics