Phage therapy (PT) is a therapeutic approach to treat patients with bacterial infections. It is based on the use of viruses, called bacteriophages (or phages), to infect and kill pathogenic bacteria throughout their lifecycle (Matsuzaki et al., 2005). Present in all ecosystems, bacteriophages are viruses that naturally and specifically infect bacteria and are, therefore, unable to infect eukaryotic cells. This therapy was developed more than a century ago with the discovery of phages by researchers Frederick Twort (Twort, 1915) and Félix d’Herelle (1917) (D'Herelle, 2011). After promising successes, it was abandoned in favor of antibiotic therapy. However, in recent years, there has been a renewed interest in PT due to the emergence of nosocomial infections with antimicrobial-resistant (AMR) bacteria and the lack of new effective antibiotics. AMR is now considered by the World Health Organization as one of the greatest threats to global health, food security, and development [antibiotic resistance (WHO, 2021)]. This problem has been concretely measured both by the Centers for Disease Control (CDC), which reported more than 2.8 million antibiotic-resistant infections each year in the United States, causing more than 35,000 deaths (CDC, 2019) and by the United Kingdom, which commissioned a report on AMR in 2016 projecting to cause 10 million deaths per year by 2050. The concept of PT is to correctly match a bacterium with one or more phages capable of infecting and killing it.

After adsorption of the phage onto the target bacterium, the phage transfers its genome (viral nucleic acid, either DNA or RNA) into the bacterial cytoplasm. During the lytic cycle, the DNA is then transcribed, translated, and copied, to be assembled into viral particles (hijacking the bacterial replication machinery). In general, when a critical mass of viral particles is reached, bacterial lysis is actively triggered via lytic proteins that disrupt the bacterial wall, allowing the release of new viruses and leading to the death of the bacterium (Delbrück, 1940; Weinbauer, 2004). The advantages of PT are numerous and linked to the very nature of phages. These viruses, used as therapeutic agents, are able to regulate themselves, at sites of infection and once the bacteria are killed, phages do not replicate and can be rapidly eliminated by the immune system or other mechanisms.

It is estimated that there are about 10³¹ phages on Earth (Weitz et al., 2013; Mann, 2005) and they are in constant co-evolution with bacteria, which makes them a potentially inexhaustible source in nature. It is therefore theoretically possible to isolate new phages for most types of bacteria. This is important because new variants of pathogenic bacteria are appearing, leading to complicated therapeutic situations, especially with the emergence of multi-antibiotic–resistant bacteria. Moreover, the mechanisms of action of phages seem to be independent of those of antibiotics, and they do not provide selective pressure likely to increase antibiotic resistance. Finally, their narrow specificity against bacterial strains allows them to have a negligible impact on the patient’s microbiota (commensal flora). The host range of a phage is the spectrum or number of strains of bacterial species that a given specific phage can infect. Phages exhibit a narrow host range, and each phage can only infect a small number of bacteria, and therefore the use of a single phage has a low probability of being able to treat infections caused by several bacteria (Matsuzaki et al., 2014; Nilsson, 2014; Mapes et al., 2016). It is therefore essential to precisely identify the bacteria responsible for the infection before implementing the PT and to use a combination of several phages (phage cocktail) (Matsuzaki et al., 2005), which increases the number of targeted bacterial strains, that is, the spectrum of action of the treatment and reduces the rate of evolution of resistance to phages (Filippov et al., 2011; Gu et al., 2019; Ramirez et al., 2020).

However, the selection of adequate phages (discovery, isolation, and characterization) is time-consuming and requires laborious regulatory approval (Socher et al., 2011), which makes it one of the main limiting steps of PT. One forward-thinking modernization of phage therapy involves genetically modifying phages to overcome the limited efficiency of natural phages. In the last few years, the main genetic modifications applied to phages generally include i) mutations in genes (Yehl et al., 2019), ii) the partial or full replacement of genes (Mahichi et al., 2009; Lin et al., 2012; Dunne et al., 2019), and iii) the insertion of foreign genes (Bikard et al., 2014; Pei and Lamas-Samanamud, 2014; Yosef et al., 2015; Lam et al., 2021) using molecular techniques such as homologous recombination (HR) and genome rebooting with the aim of mainly improving the host range or enhancing the antibacterial effect of phages. For more details on the methodologies used by genetic engineering, we suggest recent reviews (Chen et al., 2019; Guo et al., 2021; Lenneman et al., 2021). All this research led to the first success in phage genetic engineering in 2019. In the context of the treatment of a 15-year-old lung transplant recipient (Dedrick et al., 2019), some genes were removed from phages to increase their activity. Such genetically engineered (GE) phages can therefore provide substantial advantages over natural phages in terms of host range, immune system recognition, and environmental stability. Phage engineering could provide a rapid strategy to generate phages with unique properties, and thus accelerate the development of PT, provided that a sound methodology is developed to suggest appropriate modifications to be applied to phage genomes. Currently, there is no such systematic method to guide the design of genetically engineered phages. Phages exhibit unparalleled genetic diversity, which makes it extremely complicated to factor in all possible variables for creating GE phages exhibiting expected properties. New technologies are needed to accelerate the design–build–test cycle for engineering phages and to make it possible to translate proof-of-concept academic work more efficiently into real-world use.

In this context, we explore herein the application of artificial intelligence (AI) which has the potential to both guide and accelerate genome modifications to generate phage variants with unique properties that overcome the limitations of natural phages. Until now, the application of AI to phage biology mainly concerns automated recovery, prediction, and classification of bacteriophages (Ren et al., 2017; Amgarten et al., 2018; Chibani et al., 2019; Kieft et al., 2020; Shang et al., 2021) predicting phage–bacteria interactions or host prediction at the genus, species, and/or strain levels (Leite et al., 2018; Ataee et al., 2020; Boeckaerts et al., 2021; Li et al., 2021; Zhou et al., 2022), bacteriophage lifecycle (McNair et al., 2012; Tynecki et al., 2020), or the identification of viral sequences (Seguritan et al., 2012; Manavalan et al., 2018; Cantu et al., 2020; Meng et al., 2020).

We propose a novel approach, dubbed PERPHECT (for Deep Generative Networks for Bacteriophage Genetic Edition), aiming to genetically engineer bacteriophages to enhance the activity of resulting phages (Ataee et al., 2020) (Ataee et al., 2020). The case study will focus on increasing the host range of phages. To achieve this, we combine state-of-the-art techniques from deep learning: a phagi–bacterium interaction predictor, the PERPHECT predictor, and a phage genome-sequence generator, the PERPHECT generator. Therefore, to evaluate the ability of our generator, we compare the host range of phages determined experimentally with their predicted host range computed by counting the number of bacterial strains predicted as positive by the PERPHECT predictor after rewriting their genomes. A key point of the PERPHECT architecture is that its two fundamental components are loosely coupled. With this property, different methods and approaches could be used to implement either or both the predictor and the generator.

In this article, we explore and evaluate the adequacy of a deep learning model to implement the generator component. For this purpose, we take advantage of LSTMs, a special type of recurrent neural network (Servan-Schreiber et al., 1988; Servan-Schreiber et al., 1991), to process long genomic sequences to capture high-level structures contained within them. LSTMs (Hochreiter and Schmidhuber, 1997; Sherstinsky, 2020) are extremely powerful deep learning models used to capture long-range dependencies since they are made of memory units allowing to save important features. LSTMs are particularly used in natural language processing (NLP), in applications such as paraphrase detection (Socher et al., 2011), speech recognition (Li et al., 2015), language modeling (Siami-Namini et al., 2018), text generation (Santhanam, 2018), but also for genome modeling (Li, 2019), or temporal data analysis (Abdel-Nasser and Mahmoud, 2017; Deng et al., 2020). The PERPHECT architecture paves the way for the use of deep learning methodologies to genetically modify phages and extend, for example, their host range, thus confirming the potential of these approaches to guide bacteriophage engineering.

As already mentioned, the PERPHECT architecture is formed of two fundamental components (schematically represented in Figure 1):

• The PERPHECT predictor, used to predict potential interactions between bacteria and phages based solely on their genomic information, is a deep learning model, composed of a multi-branch 1-D CNN (Brownlee, 2018) (Figure 2). The proposed predictor was trained, validated, and tested on a dataset composed of 7,720 interactions between 227 bacteria and 3,208 phages. Its evaluation results showed performance figures as follows: 85% accuracy, 85% recall, 72% precision, and 78% f1-score on the test set. This predictor is explained in more detail (Ataee et al., 2020).

The results of the phage genomic modification using the generator models A (single LSTM) and B (two-stacked LSTMs) are shown in Table 1. For both models and for the four seed lengths used, the following results are presented: the number and ID of the phages used to train the model together with the training accuracy, the number of phages whose host range has been extended, the percentage of host range improvement, the percentage of the phage genome that is modified and, finally, the maximum host range obtained with the corresponding generator. The proposed generators A and B are trained with an average accuracy of 94.3 and 98.7%, respectively, and are able to improve the host range for, respectively, an average of 18 and 17 phages among the 42 under study. The average host range of phages modified with model A is improved by 73%, while with model B, it is improved by 103.7%.

A more detailed view of the effect of the PERPHECT generator on the host range of all the 42 phages is illustrated in Figures 7–9 for three configurations of interest (Note: The Supplementary Figures S1–S8 present these results for all the configurations explored). In these graphics, green dots represent phages whose host range is improved after modification. Figure 7 shows the configuration exhibiting a maximum host range of 24, the highest for model A. It is obtained with n = 200. Figure 8 presents the behavior of model A with a seed of 1,000 nucleotides. This configuration allows model A to improve the host range for 22 of the 42 phages under study, although the maximum host range is only 22. Finally, Figure 9 illustrates the behavior of the best configuration overall: model B (Two stacked LSTMs) with a seed of 1000 nucleotides which was trained on sequences from three phages. As shown in Table 1, it improves the host range of 26 of the 42 phages under study. This generator is also able to improve the maximum host range from 21 to 29 (a 38% improvement) with six of the phages exceeding 20 predicted interactions.

In general, one can observe from these results that there is no, or little, correlation between the host ranges before and after the modification of the phages. Indeed, the largest host ranges after modification are not necessarily obtained with phages exhibiting the largest host range before modification. From the results, one can associate a higher model complexity (i.e., a stacked LSTM model, trained on several phages, and using longer seeds) with better performance.

In this article, we proposed an innovative approach composed of two fundamental deep learning components: a phage–bacterium interaction predictor and a phage genome-sequence generator. For the latter component, we developed two different artificial RNN models composed of either one or two stacked LSTM layers with 256 neurons each, while the genome sequence of either one or several phages participated, respectively, in the process of training the models. The proposed generators are used to engineer the genome sequences of 42 selected phages. The phage–bacterium interaction predictor, composed of a multi-branch 1D-CNN, is then used to estimate the host range of the modified bacteriophages across 46 strains of the Pseudomonas aeruginosa species under study.

As shown by the evaluation results presented in Table 1 and Figures 7–9, both generator architectures are able to significantly improve the host range of the phage panel under study. For model A, based on a single LSTM hidden layer, the seed length used does not seem to have a strong effect on its performance and the most limiting issue appears to be its relatively limited learning capabilities. For model B, with two stacked LSTM hidden layers and trained on several phage genomes, the effect of the seed length is more apparent as the host range improvement for seeds of 700 and 1,000 nucleotides is markedly better with respect to the number of improved phages and the average improvement. Nevertheless, the effect on the maximum host range is substantially better only for the longest seed attaining 29. Exploring even the longest seeds could not provide additional advantages as the training accuracy of this model is already 99.5%. These encouraging performances give a green light to further usage of deep learning models, especially LSTM models, in guiding genetic editing of phages to improve their antibacterial power.

Our proposed approach is the first step toward a systematic method to guide the search for genetically engineered phages. Nevertheless, our approach is currently addressing one optimization goal, that is, maximizing phage interaction with target bacteria, while other fundamental criteria, such as minimizing noise, repeats, and non-informative code and maintaining the biological coherence of the organism are yet to be addressed. Some of them, such as noises, repeats, and even unnecessary code, could be tackled by including them in the loss function during the optimization process. Nevertheless, reinforcing the biological coherence of the resulting sequence would require more elaborated strategies. For example, the use of XAI (explainable AI) methods could allow to systematically detect phage-relevant hotspots/motifs in the predictor model. On this base, only essential code segments, selected on the basis of sequence similarity and alignment, will be allowed to integrate existing phage genomes.