Deep learning neural network development for the classification of bacteriocin sequences produced by lactic acid bacteria

Food industry

Some microorganisms can cause food and beverage contamination, leading to their deterioration, posing a constant concern in the food industry as it can spoil taste and cause foodborne illnesses in humans.^30,31 Bacterial pathogens transmitted through food are the primary cause of food poisoning. Chemical additives have been widely used for food preservation; however, their toxicity may raise human health issues. Some of the commercially used chemical preservatives include various synthetic chemicals.^32,33 Currently, there is a negative public perception towards chemical preservatives. This has led to a consumer preference for alternatives considered more “natural”.³⁴

In response to this demand for natural preservatives, bacteriocins show significant potential for use in the food industry, aiming to prevent food spoilage and hinder disease transmission by inhibiting the growth of pathogenic bacteria.^34,35 Certain LAB-derived bacteriocins, such as nisin, pediocin, enterocin, and leucocin, have been employed for this purpose.^36–38 They can be used in the preservation of dairy products, meats, vegetables, sourdough bread, wine, among others.² Furthermore, using bacteriocins as preservatives leads to the creation of tastier, less acidic, lower salt content, and higher nutritional value food products. Additionally, these bacteriocins can be used as antimicrobial films in food packaging to extend the shelf life and expiration dates of these products.^39,40

However, it’s important to note that while bacteriocins are a promising tool, their application is still under development and study, and they do not completely replace traditional antibiotics in all cases. Further research is needed to fully understand their potential and limitations.³⁴

Medicine

Currently, the growing resistance of bacterial pathogens poses a serious challenge to global public health, impacting not only humans but also animals, plants, and the environmental ecosystem.⁴¹ Drug resistance is on the rise worldwide due to the excessive and uncontrolled use of antimicrobial substances. According to the WHO, superbugs represent one of the most significant threats to public health, causing millions of deaths each year.⁴² It is projected that by 2060, at least 20 new types of antibiotics will be needed to effectively address the problem of bacterial drug resistance. However, developing new antibiotics involves a long and complex process, posing a significant barrier. Therefore, it is imperative to explore and develop new therapeutic strategies capable of effectively combating antibiotic-resistant microorganisms.^7,18

In clinical applications, some bacteriocins have demonstrated efficacy in treating infections, especially those caused by multidrug-resistant strains. Being produced by non-pathogenic bacteria that typically colonize the human body, they are of interest in the medical field.^43–45 Some identified bacteriocins applicable in the treatment of infectious diseases include nisin, lacticin, salivaricin, subtilosin, mersacidin, enterocin, gallidermin, epidermin, and fermentin.³⁰ Furthermore, bacteriocins have been explored for potential use in treating conditions such as diarrhea, dental caries, mastitis, and cancer.^46–48

Livestock animal husbandry

Livestock, comprising domestic animals raised in agricultural settings, play a crucial role in providing labor and a wide range of products such as milk, meat, eggs, hides, and leather. Maintaining livestock health and improving the economy through optimal production requires proper feeding and effective hygiene practices. However, farm animals remain susceptible to infections caused by viruses and bacteria despite these measures.^49–51

In the quest to safeguard animal health on farms, novel techniques are being explored as alternatives to antibiotics. This search becomes especially relevant due to various infectious diseases caused by bacteria in cattle, including conditions like mastitis, post-weaning diarrhea, meningitis, arthritis, endocarditis, pneumonia, and septicemia. Despite this pressing need, the range of bacteriocins evaluated for maintaining livestock health is limited, primarily focusing on nisin, lacticin, garvicin, and macedocin.^52–54

The application of bacteriocins in livestock food or water has ensured food safety by reducing the presence of foodborne pathogens in the gastrointestinal tract.^55,56 This application of bacteriocins has not only been used to improve the productivity of cattle but also probiotic strains capable of producing bacteriocins have been explored to increase the growth rate of pigs. Furthermore, efforts have been made in the poultry industry to control Salmonella.⁵⁷ Maintaining a diet with bacteriocin-producing bacteria can reduce existing populations of foodborne pathogens such as Salmonella and Escherichia coli and prevent the reintroduction of these pathogenic bacteria.⁵⁵ Additionally, they can be used in other forms such as the development of intra-mammary formulations for mastitis, which act as germicidal preparations applied to cows’ udders.^58,59

Aquaculture

Aquatic cultures face similar challenges to livestock, dealing with potential pathogenic risks and requiring preventive measures such as various breeding techniques, vaccination, and antibiotic use.^55,60 Bacteriocins function as probiotics, leveraging the interconnected ecosystem shared by animals and microorganisms within the aquatic environment. This interaction promotes probiotic competition against pathogenic bacteria, facilitating the production of inhibitory compounds. As a result, it improves water quality, strengthens the immune response of host species, and enhances species nutrition by producing additional digestive enzymes.^61–63

Studies involving photosynthetic bacteria like Rhodobacter sphaeroides and bacteriocins derived from Bacillus spp. have investigated their impact as probiotics on shrimp growth and digestive enzyme activity.^64,65 Likewise, experiments with nutrient-enriched water using Alchem Poseidon, a blend of Bacillus subtilis, L. acidophilus, Clostridium butyricum, and Saccharomyces cerevisiae, have shown potential for preventing infections, as the administered bacteria successfully colonized both the host and the aquatic environment.^66,67

Work related to artificial intelligence for the classification of bacteriocin sequences

Among the works carried out using deep learning neural networks to analyze large datasets and achieve accurate classification of bacteriocins is the article by Poorinmohammad et al. (2018).⁶⁸ In this study, peptide sequence analysis is conducted using machine learning alongside feature selection, and a Sequential Minimal Optimization (SMO)-based classifier is developed to predict lantibiotics, achieving precision and specificity values of 88.5% and 94%, respectively. However, this approach was limited to lantibiotics (Class I bacteriocins) and did not address the structural diversity of other bacteriocin classes.

Furthermore, in the work of Yount et al. (2020),⁶⁹ the BACII𝛼 algorithm was created to identify and classify bacteriocin sequences. This algorithm integrates a consensus signature sequence, physicochemical elements, and genomic patterns within a high-dimensional query tool to select peptides resembling bacteriocins. It accurately retrieved and distinguished almost all known class II bacteriocin families, achieving a specificity of 86%. While innovative, BACII’s reliance on predefined class II motifs limits its applicability to novel or atypical bacteriocin families. In the article by Akhter and Miller (2022), a similar approach was taken, where a machine learning-based software tool was developed to extract potential features from bacteriocin and non-bacteriocin sequences, considering their physicochemical and structural properties. Support Vector Machine (SVM) and Random Forest (RF) algorithms were employed. In this article, a precision of 95.54% was achieved.⁷⁰ Notably, this tool used small datasets (<1,000 sequences), which may restrict its generalization to broader bacteriocin diversity.

Various methods have also been used to identify bacteriocins from bacterial genomes based on bacteriocin precursor genes or contextual genes. For instance, BAGEL⁷¹ and BACTIBASE⁷² are online tools that analyze experimentally validated and annotated bacteriocins, similar to the BLASTP protein search tool. These tools rely on methods that facilitate the identification of potential bacteriocin sequences based on the homogeneity of known bacteriocins. However, these similarity-based approaches suffer from two critical limitations. They inherently exclude bacteriocins with low homology to known sequences, and their databases are biased toward well-studied LAB genera (e.g., Lactobacillus), underrepresenting rare producers like Weissella or Vagococcus. This issue led to the development of the BOA software,⁷³ which attempts to address this problem by integrating prediction tools based on the conservation of contextual genes from the bacteriocin operon. Nevertheless, they still rely on genomic searches based on homology.

There are taxonomic bias in existing tools. A recurring challenge in bacteriocin prediction is the overrepresentation of certain LAB genera (e.g., Lactobacillus, Enterococcus) in public databases, which may skew models toward recognizing features specific to these groups. For example, in UniProt, >60% of annotated bacteriocin sequences derive from just three genera, potentially marginalizing structurally unique peptides from less-studied LAB. This bias could lead to false negatives in ecological or industrial applications where microbial diversity is crucial.

In addition, the study by Nguyen et al. (2019) utilized a different technique from the previous methods by applying word embeddings of protein sequences to represent bacteriocins. This approach takes into account the amino acid order in protein sequences to predict new bacteriocins from sequences without relying on sequence similarity. While promising for novel bacteriocin discovery, their model was trained on limited data and did not account for taxonomic imbalances in sequence sources. This method even enables the prediction of potentially unknown bacteriocins with high probability. Overall, representing sequences with word embeddings that preserve information about the sequence order can be applied to peptide and protein classification problems where sequence similarity cannot be used.⁷⁴

Similarly, in the work by Hamid and Friedberg (2019),⁷⁵ word embedding was used to identify bacteriocins, representing protein sequences using Word2vec. These representations were used as inputs for various deep recurrent neural networks (RNNs) to distinguish between bacteriocin and non-bacteriocin sequences. This technique addresses challenges such as diversity among bacteriocin sequences. Though effective, their RNN architecture required manual tuning for different bacteriocin classes, reducing scalability. Meanwhile, Fields et al. (2020) developed a process for designing and testing bacteriocin-derived compounds. They employed machine learning and a filter of biophysical features to generate an algorithm that predicts bacteriocins. This involved generating characteristic sequences of 20-mers.²⁶ A key limitation was their focus on short peptides (≤50 AA), excluding larger bacteriocins like Class III.

Current bacteriocin prediction tools, such as BAGEL⁷¹ and BACII,⁶⁹ fail to address two critical needs in the field. First, they lack taxonomic resolution, making them unable to distinguish bacteriocins produced by lactic acid bacteria (LAB) from those synthesized by other bacterial groups. Second, they depend heavily on sequence homology, which limits their ability to detect structurally novel bacteriocins, particularly those originating from understudied LAB genera. This gap significantly restricts their usefulness in industrial and therapeutic contexts, where the specific taxonomy of the bacteriocin-producing organism is essential for applications such as probiotic development, targeted antimicrobial design, and food safety strategies.

To overcome these limitations, our study uses a balanced dataset spanning all major bacteriocin classes and LAB genera (without genus-specific evaluation), employs k-mer features independent of sequence homology, and validates performance on a generalized LAB group to ensure broad applicability (see Methods).

Additionally, there are other works that use antimicrobial peptide (AMP) sequences. However, it’s important to note that all bacteriocins are antimicrobial peptides, but not all antimicrobial peptides are bacteriocins. For example, in the study by Li et al. (2022),⁷⁶ they present a deep learning model called AMPlify for antimicrobial peptide prediction. The cross-validation results for the model achieve 91.70% accuracy, 91.40% sensitivity, 92.00% specificity, and 91.68% F1 score.

Similarly, in Wang et al. (2023),⁷⁷ they developed a bidirectional short and long-term memory deep learning network called AMP-EBiLSTM with an accuracy of 92.39%. This approach employs a binary profile function and a pseudo-amino acid composition to capture local sequences and extract amino acid information. In another study, a model known as AMP-BERT was developed. This network uses a bidirectional transformer encoder (BERT) architecture to extract structural and functional information from input peptides, categorizing each input as AMP or non-AMP. Notably, this network achieved a correct prediction rate of 76% for external test sequences selected in this research.⁷⁸

Similarly, a system called AMPs-Net was introduced, an algorithm designed to streamline experimentation and improve the efficiency of discovering potent AMPs. It exhibited good prediction of the antibacterial capabilities of numerous peptides, with an average accuracy ranging from 80.98% to 91.2% and precision varying from 75.77% to 94.26%.⁷⁹ In the study by Gull et al. (2019), they achieved 97% accuracy for an algorithm that identifies biologically active and antimicrobial peptides.⁸⁰ Similarly, in the study by Redshaw et al. (2023), a neural network was developed to predict the antimicrobial activity of sequences. It was trained on two different databases, achieving a precision result of 86-92% for one database and 72-77% for the other.⁸¹

In another work, an application used for predicting antimicrobial peptides based on properties achieved an accuracy exceeding 80% and sensitivity above 90%.⁸² In the study by Yan et al. (2020), a method for predicting short-length antimicrobial peptides (≤ 30 aa) is presented. Their convolutional neural network, called Deep-AmPEP30, demonstrated a 77% accuracy rate.⁸³ Additionally, in the study by Veltri et al. (2018), a deep learning neural network using embedding vectors to reduce weights when processing sequences was developed. It was shown that antimicrobial peptides could be constructed using only nine amino acids, achieved through the k-mers method. The network achieved an accuracy of 90.55%.⁸⁴

The primary aim of this study is to develop a deep learning model that accurately distinguishes bacteriocin sequences produced by lactic acid bacteria (LAB) from non-LAB bacteriocins. Unlike existing tools that classify bacteriocins generically, our approach specifically targets the LAB/non-LAB dichotomy, enabling applications in probiotic development and food safety. It uses k-mer signatures and embedding vectors to overcome the limitations of homology-based methods and provides interpretable features (100 characteristic k-mers per length) to guide synthetic peptide design.

Data collection

The AA sequences from both BacLAB and Non-BacLAB were obtained using the publicly accessible UniProt database, downloaded in xlsx format using the Excel option on the platform.¹⁹ The search on this platform was conducted using the keyword “bacteriocin.” The retrieved parameters for each bacteriocin include: Entry, Organism, Length, and Sequence. Additionally, considering the binary classification, a column was added to label the sequences. The BacLAB dataset was labeled as 1, while the Non-BacLAB sequences were labeled as 0.

To classify which sequences correspond to BacLAB and which ones to Non-BacLAB, the parameter “organism” was considered to identify the species that produce the bacteriocin. The LAB genera included for classification encompassed Lactobacillus, Lactococcus, Leuconostoc, Pediococcus, Streptococcus, Aerococcus, Alloiococcus, Carnobacterium, Dolosigranulum, Enterococcus, Oenococcus, Tetragenococcus, Vagococcus, and Weissella.⁸⁵

Sequences with lengths between 50 and 2000 amino acids were selected to ensure consistency. After filtering, the BacLAB dataset contained 24,964 sequences. For the Non-BacLAB dataset, which originally had a larger number of sequences, a random subset of 25,000 sequences was selected to prevent class imbalance in subsequent analyses. Figure 2 illustrates the length of each individual sequence (y-axis) plotted against its position in the ordered dataset (x-axis), allowing a direct comparison of length trends between BacLAB and Non-BacLAB sequences.

Figure 2. Length of each sequence (BacLAB vs. Non-BacLAB) ordered by dataset position.

The curves display the length (in amino acids) of each BacLAB and Non-BacLAB sequence, plotted according to their original position in the dataset. The x-axis represents the sequence index (1 to 25,000), and the y-axis shows the corresponding sequence length. Sequences were filtered to retain lengths between 50 and 2000 amino acids.

Feature extraction

K-mers

In the realm of amino acid sequence processing (or biological sequences in general) using neural networks, a ‘k-mer’ refers to subsequences of length ‘k’.⁸⁶ These subsequences are formed by dividing a longer sequence into specific-sized fragments, where ‘k’ represents the size of each fragment.⁸⁷ For example, a k-mer of size 5 would involve splitting the sequence into all possible subsequences of length 5, as illustrated in Figure 3. The k-mer features of a set of sequences enable the discovery of hidden patterns within that sequence population. Additionally, k-mers are useful for representing sequences in a more manageable way.⁸⁸

Figure 3. Illustration of k-mers generated from an amino acid sequence with different k-values.

On the left side are shown the k-mers that would be obtained from a sequence if k=5 is set. On the right side the same sequence is used, but in this case k=7.

At this stage, a list of the 100 most common k-mers within the BacLAB data set was generated. For this, several values of k were selected (k=3, 5, 7, 15, and 20). The k-mers of each BacLAB sequence were generated. Once all the k-mers were obtained, the frequency of each of them was counted. The 100 k-mers with the highest frequency were selected; this was done for each value of k, resulting in five different lists.

After compiling the lists, feature vectors of ‘0’ and ‘1’ were extracted for each sequence, both for those in the BacLAB and Non-BacLAB groups. The k-mers obtained from each sequence were compared with the list of k-mers. A ‘1’ was assigned if the listed k-mer was present in the analyzed sequence, while a ‘0’ was assigned if the k-mer was not found. This process produced a vector of length 100. Figure 4 illustrates the process.

Figure 4. Feature extraction from an AA sequence.

The list of 100 selected k-mers is compared with the k-mers of the input sequence. If one of the k-mers of the sequence is found in the list, '1' is added; if it is not found, a '0' is added. This process generates a representative vector for the sequence with 100 features in length. In this example, k=5 is used.

Embedding vectors

Word embeddings are numerical representations of amino acids, where each letter denoting an amino acid receives a unique and discrete value.⁸⁹ Each protein is treated as a distinct input token, and the set of 20 amino acids forms a specific dictionary.

For example, for ‘A’ (Alanine), the index 1 is assigned. Consequently, in a sequence, each occurrence of ‘A’ is denoted with the value 1. Figure 5 clarifies the process of generating the index vector. If letters were to appear in the sequence that are not found in the list of amino acids, they will be represented as zero. These indices are used to encode sequences before introducing them into the neural network that generates the embedded vectors.

Figure 5. Encoding of amino acid sequences.

a) The index number corresponding to each aa is assigned. b) Shows how the sequence is encoded with the indices that correspond to each AA. c) Given that there can be letters or numbers in the sequence that do not exist in the aa list, a value of 0 is assigned as an index. This way, errors are avoided when processing the sequence.

Once the index-encoded vectors are obtained, the embedding vectors are extracted. To derive these features, a recurrent neural network (RNN) is applied using the Gated Recurrent Unit (GRU) cell. RNNs with GRUs can handle sequences of varying lengths due to their inherent sequential processing nature and the specific architecture of GRUs. This makes GRU-based RNNs particularly useful in applications where sequence lengths are variable, as they can efficiently handle input length variability without losing learning capacity.^90–92

The embedding layer in the network acts as a lookup table or a weight matrix where each row represents, in our case, a vectorized representation of a specific amino acid.⁹³ The number of rows is equal to the count of unique elements in the vocabulary, which is the number of amino acids plus one, including index zero reserved for a non-existent variable in the amino acid list. The number of columns represents the embedding dimension, a model hyperparameter set to 128 in this case. Consequently, the length of the embedding vector obtained is also 128 for each sequence. Normally, before training begins, the weight matrix is initialized randomly along with all the network parameters. However, for this step, a pre-trained network is used, loading the weights into the model. Figure 6 illustrates the structure of the RNN model.

Figure 6. Flowchart of RNN model using GRU cell.

Input: Amino acid sequence (top) and its integer encoding (middle). Embedding Layer: Converts encoded indices into dense vectors (128-D) via a pre-trained weight matrix. GRU Layer: Processes sequential data (arrows indicate flow direction), capturing contextual relationships between amino acids. Linear Layer: Final transformation (LogSoftmax) for classification.

Concatenated data sets

Different datasets will be used to train the neural network and determine which combination of parameters produces the best results. For the selection of k-mers, values of k=3, k=5, k=7, k=15, and k=20 will be used, as shown in Table 2.

These specific values for k (k = 3, 5, 7, 15, 20) were chosen to align with conserved motifs reported in bacteriocin literature, ensuring coverage of both short functional domains and longer structural regions:

By incorporating this spectrum of k-values, our approach balances granularity (capturing short motifs) with context (preserving structural dependencies), a strategy validated in prior studies on peptide classification.^98,99

Deep neural network

To predict amino acid sequences, a Deep Neural Network (DNN) was employed following the structure described in Jeff et al.’s article.⁹⁹ This type of network was chosen for its ability to learn complex patterns and representations from data. Additionally, they can efficiently handle large datasets.¹⁰⁰ The construction of this neural network used Python 3.10.12 in Google Colab along with several libraries: i) Pandas (RRID:SCR 018214),¹⁰¹ ii) Keras, iii) Scikit-learn (RRID:SCR 002577),¹⁰² iv) NumPy (RRID:SCR 008633),¹⁰³ and v) Matplotlib.

The network architecture consists of four blocks. The input for each sequence is a vector, which corresponds to the concatenation of the results described in the k-mers section and the embedding features. Therefore, the length of the input depends on the number of concatenated features. In Figure 7, a representation is used where the extracted results using k-mers for k=5 and k=7, and the embedding features are concatenated. Since the result in k-mers corresponds to a vector of length 100, while the embedding features provide a vector length of 128, the input corresponds to a vector length of 328 for each sequence. The output of the neural network is the class of each sequence, where 1 denotes BacLAB and 0 represents non-BacLAB.

Figure 7. Flowchart of the deep neural network.

The model established the number of neurons in each defined layer block, with 128 neurons for the first two layers, 64 neurons for the next four layers in the second block, followed by 32 neurons for the five subsequent layers in the third block, and finally, two neurons in the last two layers in the fourth block. The number of neurons was determined based on the input parameters and the DNN architecture.¹⁰⁴ Out of the total thirteen layers in the model (excluding input and output layers), four layers are dense, three layers are activation layers, three layers are dropout layers, two layers are normalization layers, and one layer is a flattening layer. Table 3 provides a summary of the layers in the proposed DNN model.

Additionally, among the hyperparameters used, 75 epochs were set, a batch size of 40, and a learning rate of 2.5×10-5 for the Adam optimizer. “Mean_absolute_error” was used as the loss function. For training and testing the neural network, the k-fold cross-validation technique was employed, with k=30 selected.

For hyperparameter tuning, an iterative approach based on cross-validation (k=30) was employed, where values ​​were progressively optimized through empirical evaluation of key metrics (loss, accuracy, and F1-score). While this method does not follow an automated search (such as grid search), it allowed for flexible adaptation to the dataset characteristics, prioritizing the balance between model stability and computational efficiency. The final hyperparameters (learning rate=2.5×10⁻⁵, batch size=40, epochs=75) were selected when consistent convergence was observed across the evaluation metrics. Given the modular nature of the model (a combination of k-mers and embeddings) and the size of the dataset, manual optimization allowed us to prioritize biologically relevant hyperparameter combinations, reducing the computational cost compared to exhaustive methods (like grid search or random search).

Statistics analysis

In this study, ANOVA test along with Tukey test was used to assess significant differences among multiple groups based on parameters of interest, including accuracy, loss, precision, recall, and F1 score. These parameters are critical for evaluating the performance of the implemented neural network.

A confidence interval of 95% was selected to ensure that the differences identified between the groups are statistically significant, providing greater certainty about the conclusions drawn from the analysis. It is important to note that RStudio Cloud software was used as the statistical analysis tool to conduct these evaluations.

Ethics and consent

Ethical approval and consent were not required.