Systems and methods for predicting bacteriophage cocktail - bacteria interaction

Abstract

In some embodiments, the present disclosure provides a computer-implemented system for predicting the inhibitory effect of a bacteriophage cocktail on the growth of bacterial isolates, wherein the system comprises a training module, configured to train machine learning models to identify patterns and data structures from interaction data between different bacteriophage cocktails and bacterial isolates at times t and t+1 to fit prediction parameters; and, a prediction module, configured to predict the interaction between a bacteriophage cocktail and a bacterium at time t+1 from the interaction data at time t using the trained learning models. In some embodiments, the present disclosure further comprises a pre-processing module, configured to pre-process interaction data between a bacteriophage cocktail and a bacterial isolate for normalization using filters and / or transformations, and a post-processing module configured to produce rankings and comparisons between the different bacteriophage cocktails and bacterial isolates analyzed. Another embodiment of the present disclosure provides a computer-implemented method for predicting the inhibitory effect of a bacteriophage cocktail on the growth of bacterial isolates. Another embodiment of the present disclosure provides a use of the computer- implemented system for predicting the inhibitory effect of a bacteriophage cocktail on the growth of bacterial isolates.

Description

SYSTEMS AND METHODS FOR PREDICTING BACTERIOPHAGE COCKTAIL - BACTERIA INTERACTIONFIELD

[0001] The present disclosure relates to the field of machine learning systems and methods for prediction of bacteriophage cocktail-bacteria interaction.BACKGROUND

[0002] In recent years, the proliferation of antibiotic-resistant bacteria has become one of the most challenging and serious problems for public health. The use of antibiotics has been frequently used by modern medicine to ameliorate or cure life-threatening bacterial infections. However, the ability of bacteria to mutate in a short time may allow the development of antibiotic-resistant bacterial strains. The increase of treatments with these drugs and the inadequate use of them to treat non-bacterial diseases may cause significant challenges arising from the proliferation of antibiotic-resistant bacteria.

[0003] Faced with the potential threat of the rapid emergence and spread of multidrug resistant microorganisms, alternative treatments against bacterial infections are being developed. Bacteriophage viruses may be used as a natural alternative to fight various bacterial infections when classical treatment using antibiotics is inefficient. Bacteriophages may be natural antagonists of bacteria, therefore, they may regulate the ecosystem by limiting the abundance of their hosts through lytic infections. The use of bacteriophages or phages to inhibit bacterial growth, named phage therapy, may be a tool of growing interest as an alternative treatment to antibiotics, especially in the food industry.

[0004] In phage therapy, the use of phage cocktails as an antibacterial agent may be a common practice to enhance bactericidal activity while decreasing the likelihood of hosts developing resistance to multiple phages simultaneously. Rational cocktail design may be one of the key elements for successful phage therapy.

[0005] Knowing the bacteriophage cocktail-host interaction may be vital to determine a suitable treatment to different bacterial populations. However, searching for and characterizing bacteriophages that infect specific hosts may be a laborious and time-consuming task, carried out by one-to-one experiments and manual annotation of the interactions between samples. Moreover, only 1% of bacterialhosts may have been successfully cultured in the laboratory, thereby limiting the detection and characterization of the phage cocktail-host relationship.

[0006] The effectiveness of phage-based therapies may be confirmed empirically, and therefore, there is a growing demand for systems and methods for modeling and predicting bacteriophages-host interactions without the need for extensive experimentation.SUMMARY

[0007] The use of computational approaches to predict these interactions has the advantage of being faster, automatable and less resource intensive. Therefore, there is a growing demand for systems that allow to predict bacteriophage cocktail-bacteria interactions without the need for extensive experimentation.

[0008] The present disclosure relates to the field of machine learning systems and methods for the prediction of the bacteriophage cocktail-bacteria interaction.

[0009] In some aspects, the present disclosure provides a computer-implemented system for predicting the inhibitory effect of a bacteriophage cocktail on the growth of bacterial isolates.

[0010] In some aspects, the system comprises a training module, configured to train machine learning models to identify patterns and data structures from interaction data between different bacteriophage cocktails and bacterial isolates at times t and t+1 to fit prediction parameters; and, a prediction module, configured to predict the interaction between a bacteriophage cocktail and a bacterium at time t+1 from the interaction data at time t using the trained learning models.

[0011] Another aspect of the present disclosure provides a computer-implemented system for predicting the inhibitory effect of a bacteriophage cocktail on the growth of bacterial isolates, wherein the system comprises a pre-processing module, configured to pre-process interaction data between a bacteriophage cocktail and a bacterial isolate for normalization using filters and / or transformations; a training module, configured to train machine learning models to identify patterns and data structures from interaction data between different bacteriophage cocktails and bacterial isolates at times t and t+1 to adjust prediction parameters; a prediction module, configured to predict the interaction between a bacteriophage cocktail and a bacterium at time t+1 from the interaction data at time t using the trained learning models; and, a post-processing moduleconfigured to produce rankings and comparisons between the different bacteriophage cocktails and bacterial isolates analyzed.

[0012] Another aspect of the present disclosure provides a computer-implemented method for predicting the inhibitory effect of a bacteriophage cocktail on the growth of bacterial isolates, wherein the method comprises stages of train machine learning models to identify patterns and data structures from interaction data between different bacteriophage cocktails and bacterial isolates at times t and t+1 to fit prediction parameters; and, predict the interaction between a bacteriophage cocktail and a bacterium at time t+1 from the interaction data at time t using the trained learning models.

[0013] Another aspect of the present disclosure provides a computer-implemented method for predicting the inhibitory effect of a bacteriophage cocktail on the growth of bacterial isolates, wherein the method comprises stages of preprocessing interaction data between a bacteriophage cocktail and a bacterial isolate for normalization by means of filters and / or transformations; training machine learning models to identify patterns and data structures from interaction data between different bacteriophage cocktails and bacterial isolates at times t and t+1 to fit prediction parameters; predict the interaction between a bacteriophage cocktail and a bacterium at time t+1 from the interaction data at time t using the trained learning models; and, to elaborate rankings and comparisons between the different bacteriophage cocktails and bacterial isolates analyzed.

[0014] Another aspect of the present disclosure provides the use of the computer- implemented system and method for predicting the inhibitory effect of a bacteriophage cocktail on the growth of a bacterial isolate.

[0015] Additional aspects and advantages of the present disclosure will become readily apparent to those skilled in this art from the following detailed description, wherein only illustrative embodiments of the present disclosure are shown and described. As will be realized, the present disclosure is capable of other and different embodiments, and its several details are capable of modifications in various obvious respects, all without departing from the disclosure. Accordingly, the drawings and description are to be regarded as illustrative in nature, and not as restrictive.INCORPORATION BY REFERENCE

[0016] All publications, patents, and patent applications mentioned in this specification are herein incorporated by reference to the same extent as if each individual publication, patent, or patent application was specifically and individually indicated to be incorporated by reference. To the extent publications and patents or patent applications incorporated by reference contradict the disclosure contained in the specification, the specification is intended to supersede and / or take precedence over any such contradictory material.BRIEF DESCRIPTION OF THE DRAWINGS

[0017] The novel features of the invention are set forth with particularity in the appended claims. A better understanding of the features and advantages of the present invention will be obtained by reference to the following detailed description that sets forth illustrative embodiments, in which the principles of the invention are utilized, and the accompanying drawings (also “figure” and “FIG.” herein), of which:

[0018] Figure 1 illustrates distributions of training time series values. Fig. 1 A illustrates a time series of the distribution observed in a set of growth curves of different Salmonella isolates. Fig. IB illustrates a time series of the distribution observed in a set of growth curves of different E. coll isolates.

[0019] Figure 2 illustrates an example of predicted E. coll interaction curves, where the segmented lines represent the treatment curves, the solid lines represent the control curves and the shaded area represents the 95% confidence interval considering 500 samples for each time instant.

[0020] Figure 3 illustrates an example of prediction of Salmonella interaction curves, where the segmented lines represent the treatment curves, the continuous lines represent the control curves and the shaded area represents the 95% confidence interval considering 500 samples for each time instant.

[0021] Figure 4 illustrates an example of prediction of Salmonella interaction curves at variable bacteriophage multiplicity of infection (MOI), which corresponds to MOI 2.0, MOI 0.2 and MOI 0.02. The segmented lines represent the treatment curves, the continuous lines represent the control curves and the shaded area represents the 95% confidence interval considering 500 samples for each time instant.DETAILED DESCRIPTION OF THE INVENTION

[0022] While various embodiments of the invention have been shown and described herein, it will be obvious to those skilled in the art that such embodiments are provided by way of example only. Numerous variations, changes, and substitutions may occur to those skilled in the art without departing from the invention. It should be understood that various alternatives to the embodiments of the invention described herein may be employed.

[0023] The present disclosure provides a system and method of advanced analytics and methods based on machine learning for the automatic prediction of bacteriophage cocktail-bacteria interaction. Such automatic prediction may comprise optimally determining and predicting in silico the interaction between a bacteriophage cocktail and a bacterium of interest.

[0024] In some embodiments, the present disclosure provides a system and a method for predicting the inhibitory effect of a bacteriophage cocktail on bacterial growth of an isolate of interest, wherein said system and method takes as input data of the interaction between said bacteriophage cocktail and said bacterium.

[0025] In an aspect, the present disclosure provides a computer-implemented system for predicting the inhibitory effect of a bacteriophage cocktail on the growth of bacterial isolates. This system aims to accelerate the study of bacteriophage cocktail-host interactions and the optimization of the selection of bacteriophages for use in a formulation.

[0026] The interaction data received by the system preferably comprise interaction curves between different bacteriophage cocktails and bacterial isolates at a time t and t+1, allowing to predict the interaction between a bacteriophage cocktail and a bacterial isolate at a time t+1 from the interaction information between said bacteriophage cocktail and said bacterial isolate at a time t, based on the patterns and structure identified in the training of the system.

[0027] In some embodiments, the system considers data related to in vitro experimental results to assess bacteriophage cocktail-bacteria interaction. The data used may be interaction curves between a bacteriophage cocktail and a bacteria. In some embodiments, the interaction curves may comprise results of in vitro assays for measuring the interaction between a bacteriophage cocktail and a bacteria. In a preferred embodiment, the in vitro experimental data refer to results of challenge between a bacteria and a bacteriophage cocktail obtained from experiments inliquid or semi-solid culture media. More specifically, the challenges may comprise interaction assays in spectrophotometer, bacterial growth curves at OD600, or any other related data.

[0028] The interaction predicted by the system may comprises growth curves of the bacterial isolate in the presence of the bacteriophage cocktail. Said interaction can be transformed to a percentage inhibition of bacterial growth by the bacteriophage cocktail.

[0029] Preferably, time t corresponds to a time between 1 and 10 hours and time t+1 corresponds to a time between 10 to 100 hours. More preferably, time t corresponds to 6 hours and time t+1 corresponds to 18 hours.

[0030] In some embodiments, the system is capable of predicting the inhibitory capacity of more than one bacteriophage cocktail based on a bacterial isolate of interest.

[0031] In some embodiment, the present disclosure provides a computer-implemented system for predicting the inhibitory effect of a bacteriophage cocktail on the growth of bacterial isolates, wherein the system comprises: a training module, configured to train machine learning models to identify patterns and data structures from interaction data between different bacteriophage cocktails and bacterial isolates at times t and t+1 to fit prediction parameters; and, a prediction module, configured to predict the interaction between a bacteriophage cocktail and a bacterium at time t+1 from the interaction data at time t using the trained learning models.

[0032] In another embodiment, the system further comprises: a pre-processing module, configured to pre-process interaction data between a bacteriophage cocktail and a bacterial isolate for normalization using filters and / or transformations, wherein said pre-processing module is upstream of the training module.

[0033] In another embodiment, the system further comprises: a post-processing module configured to produce rankings and comparisons between the different bacteriophage cocktails and bacterial isolates analyzed, wherein said post-processing module is downstream from the prediction module.

[0034] In some embodiments, a training module is provided. The training module comprises at least the following steps: a) separate a normalized set of sequential individual interaction data between different bacteriophage cocktails and different bacterial isolates intotraining, validation, and test data subsets; b) select appropriate machine learning models; c) receive, in each selected model, the first subset of sequential interaction data, which corresponds to the training data subset; d) perform, on each model, transformations, decompositions or a composition of several transformations and compositions of such data to compact feature vectors or compact representations; e) identify patterns or structures of the data in the subset of training data; f) predict the interaction between a bacteriophage cocktail and a bacterial isolate at time t+1 from data of such interaction at time t; g) compare said interaction prediction at time t+1 with the actual value of said interaction by means of performance metrics; h) receive, for each selected model, the second subset of sequential individual interaction data, which corresponds to the validation data subset; i) validate the results using the validation subset under the same performance metrics used with the training subset for each selected model.

[0035] In some embodiments, selecting appropriate machine learning models comprises selecting from 1 to 10 learning models with different levels of complexity. Preferably, selecting models comprises selecting 3 to 5 learning models.

[0036] In some embodiments, the learning models may be selected from single block models or deep learning models, such as Long Short-Term Memory (LSTM), Neural Basis Expansion Analysis for Time Series Forecasting (N-BEATS), Neural Hierarchical Interpolation for Time Series Forecasting (N-HiTS), Wavenet-type networks, Temporal Convolutional Networks (TCN), some types of Transformers, among other related models. As an expert in the technique will be able to identify, the module is capable of using any learning model that has the ability to model temporal patterns.

[0037] In some embodiments, the learning models comprise deep learning models designed to detect temporal patterns at different scales, combining them into a single prediction.

[0038] In some embodiments, identifying patterns or structures of the data comprises modeling patterns using function blocks, wherein the function blocks comprise operations such as matrix multiplications, one-dimensional convolutions, or combinations thereof, configured to encode the time series into compact feature vectors, storing information from the past to be able to predict one point in thefuture at a time.

[0039] In another embodiments, identifying patterns or structures of the data comprises decompositions of the data at different frequencies of its power spectrum. This ensures that the learning system detects specific patterns occurring at different time scales such as seasonality and trend. Prediction of the next point on the curve is made by combining the patterns associated with the different scales considered. In one embodiment of the invention, the selected learning models are capable of predicting into the past to improve predictive power.

[0040] In some embodiments, the learning models are selected from supervised learning models. In some embodiments, the training module uses supervised learning models, which minimize a loss function that measures the difference between the point predicted by the system and the experimentally measured point. Subsequent points on the curve are predicted by applying the model as a sliding window, incorporating a dropout strategy in order to force the model to learn with missing information.

[0041] In another embodiments, the learning models are selected from unsupervised learning models.

[0042] In some embodiments, the learning models selected are chosen based on the nature of the data received by the pre-processing module and the training module, as well as on performance, considering execution times, data characteristics, generation capacity, among other criteria.

[0043] In some embodiments, the interactions predicted by the training module are used to calculate the percentage inhibition of bacterial growth predicted by the system. For this purpose, the predicted interactions may be reconverted to the original units in which they were entered into the system. The original units may comprise optical density.

[0044] In some embodiments, the percentage inhibition of bacterial growth predicted by the system is converted to categorical variables. In some embodiments, the categorical variables comprise 3 performance categories: null, comprising a percentage of inhibition less than or equal to 15%; partial, comprising a percentage of inhibition greater than 15% and less than 85%; and total, comprising a percentage of inhibition greater than or equal to 85%. In another embodiments, the categorical variables comprise continuous variables, as percentage ranges of 30%, of 10%, or of 1%.

[0045] According to the present disclosure, the categorical variables are used formeasuring performance metrics. In some embodiments, the measuring performance metrics comprise accuracy, precision, recall, f-score, among other related. Accuracy measures the proportion of correct classifications with respect to the total number of samples. Accuray = Correct predictions / Total predictions. Precision measures the proportion of correct classifications of the positive class with respect to the total positive classifications. Recall measures how well all instances of the positive class are identified. It is the fraction of correctly classified positive instances (TP) over the total number of positive instances, which includes the false negatives (TP+FN). The F-score corresponds to the harmonic mean of precision and recall.

[0046] Optionally, if the results of the performance metrics are not satisfactory, a fine tuning can be done to improve selected metrics.

[0047] In some embodiments, the stages of the training module are iterated. In preferent embodiments, the stages of the training module are iterated between 10 and 10000 times. In some embodiments, the number of iterations can be set to an amount of interest by an operator.

[0048] In some embodiments, a process termination criterion is set. In preferent embodiments, such criterion for completion of the iterations comprising stabilization of the performance metrics being analyzed. In some embodiments, stabilizing of the performance metric is understood as the performance metric not exhibiting variations greater than a percentage of interest over 5 to 50 iterations.

[0049] In some embodiments, the performance metrics comprise a loss function, such as mean square deviation, mean absolute error, coefficient of variation (homogeneity), MAPE, SMAPE, MASE, among other related ones.

[0050] In some embodiments, in the training module, the free parameters of the models are adjusted such that a defined loss function is minimized. When this function has converged, training stops and the parameters are fixed. The trained and converged models are stored for later use in the prediction module.

[0051] In some embodiments, a prediction module is provided. The prediction module comprises at least the following steps: a) use the trained models with fixed parameters set in the training module; b) receive a set of data not seen by the models in the training module; c) for each input data, generate n predictions for each time t;d) based on the predictions for each time instant t, estimate confidence intervals; e) pair the predicted interaction predictions with the corresponding control interactions to estimate the percentage inhibition at time t+1.

[0052] In case the data are new, they must be normalized.

[0053] In some embodiments, the number of predictions (n) generated by the module comprises from 100 to 1000 predictions. In a preferred embodiment, n is 500 predictions.

[0054] In some embodiments, the interaction predictions comprise bacterial growth curves from calculation of the optical density of the bacterial culture in the presence of a cocktail of bacteriophage. According to the present disclosure, the bacterial growth curves allow calculation of the percentage inhibition of the bacteriophage cocktail on the growth of the bacterial isolate as a function of the ratio of the areas under the curve between the interaction curve and the control curve, which comprises the optical density of the bacterial culture in the absence of bacteriophage cocktail.

[0055] In some embodiments, the percentage inhibition of bacterial growth predicted by the system is converted to categorical variables. In some embodiments, the categorical variables comprise 3 performance categories: null, comprising a percentage of inhibition less than or equal to 15%; partial, comprising a percentage of inhibition greater than 15% and less than 85%; and total, comprising a percentage of inhibition greater than or equal to 85%. In another embodiments, the categorical variables comprise continuous variables, as percentage ranges of 30%, of 10%, or of 1%.

[0056] According to the present disclosure, the categorical variables are used for the measurement of performance metrics, wherein the performance metrics comprise accuracy, precision, recall, f-score, among other related.

[0057] In some embodiments, a pre-processing module is provided. The pre-processing module comprises at least the following steps: a) receive sequential interaction data between a bacteriophage cocktail and a bacterial isolate of interest with n samples; b) optionally, filter said sequential interaction data to reduce noise associated with said data; c) optionally, transform the filtered sequential interaction data to a range appropriate for use by the following modules.

[0058] In some embodiments, the sequential interaction data comprises a plurality of data from different interactions between different bacteriophage cocktails and different bacterial isolates at a time t and at a time t+1.

[0059] In some embodiments, the step of filtering the sequential interaction data comprises using low-pass filters, high-pass filters, band-pass filters, or other related filters. In preferred embodiments, the filters are selected based on the quality or type of sequential interaction data used.

[0060] In some embodiments, transforming the sequential interaction data comprises rescaling the data to a range [0,1], rescaling the data to a range [-1,1], standardization, normalization, quantile transformation, or another related transformations.

[0061] In some embodiments, the filters used in such module can be implemented from scratch or with the help of available open source libraries, such as scipy, numpy, filterpy, or other related open source libraries. The filters used can be selected from a group comprising Butterworth low-pass filter, among other related filters.

[0062] In some embodiments, data transformation can be implemented from scratch or with the help of available open source libraries, such as scikit-learn or other related open source libraries.

[0063] In some embodiments, a post-processing module is provided. The pre-processing module comprises at least the following steps: a) in case the results obtained by the prediction module contain results from more than one bacteriophage cocktail for the same set of bacterial isolates, produce a ranking based on the median bacterial growth inhibition; b) compare the performance of the cocktails in pairs in order to discriminate those that have a greater inhibitory effect; c) optionally, generate a report including interaction curves predicted by the system, their transformation to categorical variables, rankings and comparisons between cocktails performed by the system; d) optionally, visualize these results on a platform.

[0064] In some embodiments, the comparison between pairs of bacteriophage cocktails are based on the results of their median inhibition predicted by the system. In some embodiments, the comparison is performed by paired tests for comparison, comprising Wilcoxon test, Hodges-Lehmann test.

[0065] In some embodiments, the post-processing module comprises an additional feedback function, wherein an operator analyzes the results reported by thesystem and can manually correct parameters according to experimental results.

[0066] In some embodiments, the platform to visualize the result comprise a Streamlit platform.

[0067] In some embodiments, the present disclosure provides a computer-implemented system for predicting the inhibitory effect of a bacteriophage cocktail on the growth of bacterial isolates, wherein the system comprises: a pre-processing module, configured to pre-process interaction data between a bacteriophage cocktail and a bacterial isolate for normalization using filters and / or transformations; a training module, configured to train machine learning models to identify patterns and data structures from interaction data between different bacteriophage cocktails and bacterial isolates at times t and t+1 to adjust prediction parameters; a prediction module, configured to predict the interaction between a bacteriophage cocktail and a bacterium at time t+1 from the interaction data at time t using the trained learning models; and, a post-processing module configured to produce rankings and comparisons between the different bacteriophage cocktails and bacterial isolates analyzed.

[0068] In some embodiments, the system according to the present disclosure provides the prediction of the interaction between a bacteriophage cocktail and a bacterial isolate of different species and under various conditions, such as varying bacteriophage multiplicity of infection (MOI), different culture temperatures, different oxygen concentrations in culture, among others.

[0069] In some embodiments, the present disclosure provides the use of the system for predicting the inhibitory effect of a bacteriophage cocktail on the growth of bacterial isolates.

[0070] In another aspect, the present disclosure provides a computer-implemented method for predicting the inhibitory effect of a bacteriophage cocktail on the growth of bacterial isolates. This method aims to accelerate the study of bacteriophage cocktail-host interactions and the optimization of the selection of bacteriophages for use in a formulation.

[0071] The interaction data received by the method preferably comprise interaction curves between different bacteriophage cocktails and bacterial isolates at a time t and t+1, allowing to predict the interaction between a bacteriophage cocktail and a bacterial isolate at a time t+1 from the interaction information between saidbacteriophage cocktail and said bacterial isolate at a time t, based on the patterns and structure identified in the training of the method.

[0072] In some embodiments, the method considers data related to in vitro experimental results to assess bacteriophage cocktail-bacteria interaction. The data used may be interaction curves between a bacteriophage cocktail and a bacteria. In some embodiments, the interaction curves may comprise results of in vitro assays for measuring the interaction between a bacteriophage cocktail and a bacteria. In a preferred embodiment, the in vitro experimental data refer to results of challenge between a bacteria and a bacteriophage cocktail obtained from experiments in liquid or semi-solid culture media. More specifically, the challenges may comprise interaction assays in spectrophotometer, bacterial growth curves at OD600, or any other related data.

[0073] The interaction predicted by the method may comprises growth curves of the bacterial isolate in the presence of the bacteriophage cocktail. Said interaction can be transformed to a percentage inhibition of bacterial growth by the bacteriophage cocktail.

[0074] Preferably, time t corresponds to a time between 1 and 10 hours and a time t+1 corresponds to a time between 10 to 100 hours. More preferably, time t corresponds to 6 hours and time t+1 corresponds to 18 hours.

[0075] In some embodiments, the method is capable of predicting the inhibitory capacity of more than one bacteriophage cocktail based on a bacterial isolate of interest.

[0076] In some embodiment, the present disclosure provides a computer-implemented method for predicting the inhibitory effect of a bacteriophage cocktail on the growth of bacterial isolates, wherein the method comprises stages: i) train machine learning models to identify patterns and data structures from interaction data between different bacteriophage cocktails and bacterial isolates at times t and t+1 to fit prediction parameters; and, ii) predict the interaction between a bacteriophage cocktail and a bacterium at time t+1 from the interaction data at time t using the trained learning models.

[0077] In another embodiment, the method comprises further stage: iii) pre-process interaction data between a bacteriophage cocktail and a bacterial isolate for normalization using filters and / or transformations, wherein said pre-processing is prior to the stages identified above.

[0078] In another embodiment, the method comprises further stage:iv) produce rankings and comparisons between the different bacteriophage cocktails and bacterial isolates analyzed, wherein said post-processing is subsequent to the stages identified above.

[0079] In some embodiments, the stage i) of the method comprises at least: a) separate a normalized set of sequential individual interaction data between different bacteriophage cocktails and different bacterial isolates into training, validation, and test data subsets; b) select appropriate machine learning models, c) receive, in each selected model, the first subset of sequential interaction data, which corresponds to the training data subset; d) perform, on each model, transformations, decompositions or a composition of several transformations and compositions of such data to compact feature vectors or compact representations; e) identify patterns or structures of the data in the subset of training data; f) predict the interaction between a bacteriophage cocktail and a bacterial isolate at time t+1 from data of such interaction at time t; g) compare said interaction prediction at time t+1 with the actual value of said interaction by means of performance metrics; h) receive, for each selected model, the second subset of sequential individual interaction, which corresponds to the validation data subset; i) validate the results using the validation subset under the same performance metrics used with the training subset for each selected model.

[0080] In some embodiments, selecting appropriate machine learning models comprises selecting from 1 to 10 learning models with different levels of complexity. Preferably, selecting models comprises selecting 3 to 5 learning models.

[0081] In some embodiments, the learning models may be selected from single block models or deep learning models, such as Long Short-Term Memory (LSTM), Neural Basis Expansion Analysis for Time Series Forecasting (N-BEATS), Neural Hierarchical Interpolation for Time Series Forecasting (N-HiTS), Wavenet-type networks, Temporal Convolutional Networks (TCN), some types of Transformers, among other related models. As an expert in the technique will be able to identify, the stage is capable of using any learning model that has the ability to model temporal patterns.

[0082] In some embodiments, the learning models comprise deep learning models designed to detect temporal patterns at different scales, combining them into asingle prediction.

[0083] In some embodiments, identifying patterns or structures of the data comprises modeling patterns using function blocks, wherein the function blocks comprise operations such as matrix multiplications, one-dimensional convolutions, or combinations thereof, configured to encode the time series into compact feature vectors, storing information from the past to be able to predict one point in the future at a time.

[0084] In another embodiments, identifying patterns or structures of the data comprises decompositions of the data at different frequencies of its power spectrum. This ensures that the learning method detects specific patterns occurring at different time scales such as seasonality and trend. Prediction of the next point on the curve is made by combining the patterns associated with the different scales considered. In one embodiment of the invention, the selected learning models are capable of predicting into the past to improve predictive power.

[0085] In some embodiments, the learning models are selected from supervised learning models. In some embodiments, stage i) uses supervised learning models, which minimize a loss function that measures the difference between the point predicted by the method and the experimentally measured point. Subsequent points on the curve are predicted by applying the model as a sliding window, incorporating a dropout strategy in order to force the model to learn with missing information.

[0086] In another embodiments, the learning models are selected from unsupervised learning models.

[0087] In some embodiments, the learning models selected are chosen based on the nature of the data received by stage iii) and stage i), as well as on performance, considering execution times, data characteristics, generation capacity, among other criteria.

[0088] In some embodiments, the interactions predicted by the stage i) are used to calculate the percentage inhibition of bacterial growth predicted by the method. For this purpose, the predicted interactions may be reconverted to the original units in which they were entered into the method. The original units may comprise optical density.

[0089] In some embodiments, the percentage inhibition of bacterial growth predicted by the stage i) are converted to categorical variables. In some embodiments, the categorical variables comprise 3 performance categories: null, comprising apercentage of inhibition less than or equal to 15%; partial, comprising a percentage of inhibition greater than 15% and less than 85%; and total, comprising a percentage of inhibition greater than or equal to 85%. In another embodiments, the categorical variables comprise continuous variables, as percentage ranges of 30%, of 10%, or of 1%.

[0090] According to the present disclosure, the categorical variables are used for measuring performance metrics. In some embodiments, the measuring performance metrics comprise accuracy, precision, recall, f-score, among other related.

[0091] Optionally, if the results of the performance metrics are not satisfactory, a fine tuning can be done to improve selected metrics.

[0092] In some embodiments, the steps of stage i) are iterated. In preferent embodiments, the steps of stage i) are iterated between 10 and 10000 times. In some embodiments, the number of iterations can be set to an amount of interest by an operator.

[0093] In some embodiments, a process termination criterion is set. In preferent embodiments, such criterion for completion of the iterations comprising stabilization of the performance metrics being analyzed. In some embodiments, stabilizing of the performance metric is understood as the performance metric not exhibiting variations greater than a percentage of interest over 5 a 50 iterations.

[0094] In some embodiments, the performance metrics comprise a loss function, such as mean square deviation, mean absolute error, coefficient of variation (homogeneity), MAPE, SMAPE, MASE, among other related ones.

[0095] In some embodiments, in the stage i), the free parameters of the models are adjusted such that a defined loss function is minimized. When this function has converged, training stops and the parameters are fixed. The trained and converged models are stored for later use in stage ii).

[0096] In some embodiments, the stage ii) comprises at least: a) use the trained models with fixed parameters set in stage i); b) receive a set of data not seen by the models in the stage i); c) for each input data, generate n predictions for each time t; d) based on the predictions for each time instant t, estimate confidence intervals;e) pair the predicted interaction predictions with the corresponding control interactions to estimate the percentage inhibition at time t+1.

[0097] In case the data are new, they must be normalized.

[0098] In some embodiments, the number of predictions (n) generated by stage i) comprises from 100 to 1000 predictions. In a preferred embodiment, n is 500 predictions.

[0099] In some embodiments, the number of predictions (n) generated by stage ii) comprises from 100 to 1000 predictions.

[0100] In some embodiments, the interaction predictions comprise bacterial growth curves from calculation of the optical density of the bacterial culture in the presence of a cocktail of bacteriophage. According to the present disclosure, the bacterial growth curves allow calculation of the percentage inhibition of the bacteriophage cocktail on the growth of the bacterial isolate as a function of the ratio of the areas under the curve between the interaction curve and the control curve, which comprises the optical density of the bacterial culture in the absence of bacteriophage cocktail.

[0101] In some embodiments, the percentage inhibition of bacterial growth predicted by stage ii) is converted to categorical variables. In some embodiments, the categorical variables comprise 3 performance categories: null, comprising a percentage of inhibition less than or equal to 15%; partial, comprising a percentage of inhibition greater than 15% and less than 85%; and total, comprising a percentage of inhibition greater than or equal to 85%. In another embodiments, the categorical variables comprise continuous variables, as percentage ranges of 30%, of 10%, or of 1%.

[0102] According to the present disclosure, the categorical variables are used for the measurement of performance metrics, wherein the performance metrics comprise accuracy, precision, recall, f-score, among other related.

[0103] In some embodiments, stage iii) comprises at least: a) receive sequential interaction data between a bacteriophage cocktail and a bacterial isolate of interest with n samples; b) optionally filter said sequential interaction data to reduce noise associated with said data; c) optionally, transform the filtered sequential interaction data to a range appropriate for use by stages i) and ii).

[0104] In some embodiments, the sequential interaction data comprises a pluralityof data from different interactions between different bacteriophage cocktails and different bacterial isolates at a time t and at a time t+1.

[0105] In some embodiments, filtering the sequential interaction data comprises using low-pass filters, high-pass filters, band-pass filters, or other related filters. In preferred embodiments, the filters are selected based on the quality or type of sequential interaction data used.

[0106] In some embodiments, transforming the sequential interaction data comprises rescaling the data to a range [0,1], rescaling the data to a range [-1,1], standardization, normalization, quantile transformation, or another related transformations.

[0107] In some embodiments, the filters used in such stage can be implemented from scratch or with the help of available open source libraries, such as scipy, numpy, filterpy, or other related open source libraries. The filters used can be selected from a group comprising Butterworth low-pass filter, among other related filters.

[0108] In some embodiments, data transformation can be implemented from scratch or with the help of available open source libraries, such as scikit-learn or other related open source libraries.

[0109] In some embodiments, stage iv) comprises at least: a) in case the results obtained by stage ii) contain results from more than one bacteriophage cocktail for the same set of bacterial isolates, produce a ranking based on the median bacterial growth inhibition; b) compare the performance of the cocktails in pairs in order to discriminate those that have a greater inhibitory effect; c) optionally, generate a report including interaction curves predicted by the method, their transformation to categorical variables, rankings and comparisons between cocktails performed by the method; d) optionally, visualize these results on a platform.

[0110] In some embodiments, the comparison between pairs of bacteriophage cocktails are based on the results of their median inhibition predicted by the method. In some embodiments, the comparison is performed by paired tests for comparison, comprising Wilcoxon test, Hodges-Lehmann test.

[0111] In some embodiments, stage iv) comprises an additional feedback function, wherein an operator analyzes the results reported by the method and can manually correct parameters according to experimental results.

[0112] In some embodiments, the platform to visualize the result comprise a Streamlit platform.

[0113] In some embodiment, the present disclosure provides a computer- implemented method for predicting the inhibitory effect of a bacteriophage cocktail on the growth of bacterial isolates, wherein the method comprises: pre-process interaction data between a bacteriophage cocktail and a bacterial isolate for normalization using filters and / or transformations; train machine learning models to identify patterns and data structures from interaction data between different bacteriophage cocktails and bacterial isolates at times t and t+1 to adjust prediction parameters; predict the interaction between a bacteriophage cocktail and a bacterium at time t+1 from the interaction data at time t using the trained learning models; and, produce rankings and comparisons between the different bacteriophage cocktails and bacterial isolates analyzed.

[0114] In some embodiments, the method according to the present disclosure provides the prediction of the interaction between a bacteriophage cocktail and a bacterial isolate of different species and under various conditions, such as varying MOI, different culture temperatures, different oxygen concentrations in culture, among others.

[0115] In some embodiments, the present disclosure provides the use of the method for predicting the inhibitory effect of a bacteriophage cocktail on the growth of bacterial isolates.

[0116] While preferred embodiments of the present invention have been shown and described herein, it will be obvious to those skilled in the art that such embodiments are provided by way of example only. It is not intended that the invention be limited by the specific examples provided within the specification. While the invention has been described with reference to the aforementioned specification, the descriptions and illustrations of the embodiments herein are not meant to be construed in a limiting sense. Numerous variations, changes, and substitutions will now occur to those skilled in the art without departing from the invention. Furthermore, it shall be understood that all aspects of the invention are not limited to the specific depictions, configurations or relative proportions set forth herein which depend upon a variety of conditions and variables. It should be understood that various alternatives to the embodiments of theinvention described herein may be employed in practicing the invention. It is therefore contemplated that the invention shall also cover any such alternatives, modifications, variations or equivalents. It is intended that the following claims define the scope of the invention and that methods and structures within the scope of these claims and their equivalents be covered thereby.EXAMPLES

[0117] Example 1. Construction of a pre-processing module

[0118] Using systems and methods of the present disclosure, a pre-processing module was constructed as follows. For the pre-processing module, bacteriophage cocktail - bacteria interaction information was obtained from experimental data of bacteriophage cocktail challenges with bacterial hosts. A total of 1101 data were included in the analysis, from two different bacterial species: Salmonella enterica and Escherichia coli.

[0119] The data were pre-processed as follows: a) receive sequential interaction data between a bacteriophage cocktail and a bacterial isolate of interest with n samples; b) optionally, filter said sequential interaction data to reduce noise associated with said data; c) optionally, transform the filtered sequential interaction data to a range appropriate for use by stages i) and ii).

[0120] The output of the module corresponds to the normalized and filtered data, which can then be used as input for the training module and the prediction module.

[0121] Example 2. Construction of a training module.

[0122] Using systems and methods of the present disclosure, a training module was constructed as follows. The training module analyzed as input data the bacteriophage cocktail - bacteria interaction previously normalized (by transformation, scaling or other technique).

[0123] In order to evaluate a large number of bacteriophage cocktail - bacteria interaction data, the module considered the following procedure: a) separate a normalized set of sequential individual interaction data between different bacteriophage cocktails and different bacterial isolates into training, validation, and test data subsets; b) select appropriate machine learning models, c) receive, in each selected model, the first subset of sequential interactiondata, which corresponds to the training data subset; d) perform, on each model, transformations, decompositions or a composition of several transformations and compositions of such data to compact feature vectors or compact representations; e) identify patterns or structures of the data in the subset of training data; f) predict the interaction between a bacteriophage cocktail and a bacterial isolate at time t+1 from data of such interaction at time t; g) compare said interaction prediction at time t+1 with the actual value of said interaction by means of performance metrics; h) receive, for each selected model, the second subset of sequential individual interaction data, which corresponds to the validation data subset; i) validate the results using the validation subset under the same performance metrics used with the training subset for each selected model.

[0124] If the results of the performance metrics are not satisfactory, a fine is performed tuning to improve selected metrics.

[0125] The free parameters of the models are adjusted such that a defined loss function is minimized. When this function has converged, training stops and the parameters are fixed. The trained and converged models are stored for later use in the prediction module.

[0126] Example 3. Construction of a prediction module.

[0127] Using systems and methods of the present disclosure, a prediction module was constructed as follows.

[0128] Using the trained and converged models developed in the training module, the prediction module performed the prediction of the interaction between a bacteriophage cocktail and a bacterial isolate. The module considered the following procedure: a) use the trained models with fixed parameters set in the training module; b) receive a set of data not seen by the models in the training module; c) for each input data, generate n predictions for each time t; d) based on the predictions for each time instant t, estimate confidence intervals; e) pair the predicted interaction predictions with the corresponding control interactions to estimate the percentage inhibition at time t+1.

[0129] When the data are new, they must be normalized.

[0130] The interactions predicted by the system are post-processed by the post-processing module.

[0131] Example 4. Construction of a post-processing module.

[0132] Using systems and methods of the present disclosure, a post-processing was constructed as follows. The set of predicted bacteriophage cocktail - bacterial isolate interaction generated by the prediction module was post-processed in the post-processing module. For this, the module performed the following procedure: a) produce a ranking based on the median bacterial growth inhibition of different bacteriophage cocktail against the same bacterial isolate; b) compare the performance of the cocktails in pairs in order to discriminate those that have a greater inhibitory effect; c) generate a report including interaction curves predicted by the system, their transformation to categorical variables, rankings and comparisons between cocktails performed by the system; d) visualize these results on a platform.

[0133] The results of the predictions made by the prediction module are finally presented on a platform, which allows a thorough analysis of the interaction results by operators and laboratory researchers.

[0134] Example 5. Automatic prediction of Salmonella bacteriophage cocktails - Salmonella isolates interaction.

[0135] Using systems and methods of the present disclosure, the prediction of the inhibitory effect of bacteriophage cocktails on the growth of Salmonella isolates was performed. For analysis purposes, a total of 9665 control and treatment curves were used for bacteriophage cocktail-bacterial isolate challenges. The set was divided into 10 training / test subsets for cross-validation.

[0136] The system performed the following procedure:1. Preprocessing is applied to each training and test subset: scaling to the interval [le-3, 1]; application of noise reduction filter;2. For each training subset, a machine learning model designed to decompose the time sequences into patterns useful for predicting a time step in the future is trained;3. Each trained model is evaluated in the following way: a. measurement of the root mean square error of predictions;b. the subset test data are paired so that each treatment curve is associated with its control; c. predictions are scaled to the interval [le-3, 1]; d. interaction points are predicted at t+1; e. areas under the curve are determined and the percentage of inhibition is calculated; f. inhibition is classified into two categories: Null, for inhibition less than 15%; Interaction, for inhibition greater than 15%; g. accuracy, precision, recall, and fl -score metrics are calculated;4. The final performance is obtained by averaging the metrics of all data subsets.

[0137] Performance metrics were calculated for the interaction curves. The results are illustrated in Table 1.

[0138] Table 1. Performance metrics for prediction of interaction prediction of bacteriophage cocktails with Salmonella isolates.

[0139] It is observed that the inhibition categories derived from the predicted curves are classified with sufficient accuracy.

[0140] Figure 1A illustrates the subset of Salmonella training time series, where each time instant is represented by the distribution of values of the curves in the set. By way of example, Figure 3 illustrates an interaction curve between a bacteriophage cocktail and a Salmonella bacterial isolate predicted by the system according to one embodiment of the invention.

[0141] Example 6. Automatic prediction of E. coli bacteriophage cocktail - E. coli isolate interaction

[0142] Using systems and methods of the present disclosure, the prediction of the inhibitory effect of bacteriophage cocktails on the growth of E. coli isolates was performed. For analysis purposes, a total of 6838 control and treatment curves were used for bacteriophage cocktail-bacterial isolate challenges. The set was divided into 10 training / test subsets for cross-validation.

[0143] The same procedure was performed as indicated in Example 5.

[0144] With the data obtained, performance metrics for the interaction curves were calculated. The results are illustrated in Table 2.

[0145] Table 2. Performance metrics for interaction prediction of bacteriophage cocktails with E. coli isolates.

[0146] It is observed that the inhibition categories derived from the predicted curves are classified with high accuracy.

[0147] Figure IB illustrates the subset of E. coli training time series, where each time instant is represented by the distribution of values of the curves in the set. By way of example, Figure 3 illustrates an interaction curve between a bacteriophage cocktail and an E. coli bacterial isolate predicted by the system according to one embodiment of the invention

[0148] Example 7. Automatic prediction of Salmonella bacteriophage cocktail - Salmonella isolate interaction considering different MOIs.

[0149] Using systems and methods of the present disclosure, the prediction of the inhibitory effect of bacteriophage cocktails on the growth of Salmonella isolates was performed considering 3 different bacteriophage multiplicities of infection (MOI 2.0, MOI 0.2 and MOI 0.02). For analysis purposes, a total of 2202 control and treatment curves were used for bacteriophage-bacterial isolate cocktail challenges. The set was divided into 10 training / test subsets for cross- validation.

[0150] The same procedure was performed as indicated in Example 5.

[0151] With the data obtained, performance metrics for the interaction curves were calculated. The results are illustrated in Table 3.

[0152] Table 3. Performance metrics for interaction prediction of bacteriophage cocktails with Salmonella isolates at different MOI.

[0153] It is observed that the inhibition categories derived from the predicted curves are classified with high accuracy.

[0154] By way of example, Figure 4 illustrates an interaction curve between a bacteriophage cocktail and a Salmonella bacterial isolate at variable MOI predicted by the system according to one embodiment of the invention.

[0155] The results demonstrated that the system was able to automatically predict the interaction between a bacteriophage cocktail and a bacterial isolate having high concordance between predicted and observed in vitro bacterial growth inhibition, and indicates that the system may be used to contribute to the design of bacteriophage cocktails by reducing the analysis, search and selection time of the compositions.

Claims

CLAIMSWHAT IS CLAIMED IS:

1. A computer-implemented system for predicting the inhibitory effect of a bacteriophage cocktail on the growth of bacterial isolates, wherein the system comprises: a training module, configured to train machine learning models to identify patterns and data structures from interaction data between different bacteriophage cocktails and bacterial isolates at times t and t+1 to fit prediction parameters; and, a prediction module, configured to predict the interaction between a bacteriophage cocktail and a bacterium at time t+1 from the interaction data at time t using the trained learning models.

2. The system according to claim 1, wherein the system further comprises: a pre-processing module, configured to pre-process interaction data between a bacteriophage cocktail and a bacterial isolate for normalization using filters and / or transformations, wherein said pre-processing module is upstream of the training module.

3. The system according to the preceding claims, wherein the system further comprises: a post-processing module configured to produce rankings and comparisons between the different bacteriophage cocktails and bacterial isolates analyzed, wherein said post-processing module is downstream of the prediction module.

4. The system according to the preceding claims, wherein the training module is configured to at least: a) separate a normalized set of sequential individual interaction data between different bacteriophage cocktails and different bacterial isolates into training, validation, and test data subsets; b) select appropriate machine learning models, c) receive, in each selected model, the first subset of sequential interaction data, which corresponds to the training data subset; d) perform, on each model, transformations, decompositions or a composition of several transformations and compositions of such data to compact feature vectors or compact representations; e) identify patterns or structures of the data in the subset of training data; f) predict the interaction between a bacteriophage cocktail and a bacterial isolate at time t+1 from data of such interaction at time t;g) compare said interaction prediction at time t+1 with the actual value of said interaction by means of performance metrics; h) receive, for each selected model, the second subset of sequential individual interaction data, which corresponds to the validation data subset; i) validate the results using the validation subset under the same performance metrics used with the training subset for each selected model.

5. The system according to claim 4, wherein selecting suitable learning models comprises selecting from 1 to 10 learning models with different levels of complexity.

6. The system according to claim 5, wherein the selecting models comprises selecting 3 to 5 learning models.

7. The system according to claims 5 to 6, wherein the learning models are selected from supervised learning models.

8. The system according to claim 5 to 6, wherein the learning models are selected from unsupervised learning models.

9. The system according to claim 4, wherein identifying patterns or structures of the data comprises modeling patterns using function blocks.

10. The system according to claim 9, wherein the function blocks comprise operations such as matrix multiplications, one-dimensional convolutions, or combinations thereof, configured to encode the time series into compact feature vectors.

11. The system according to claim 4, wherein identifying patterns or structures of the data comprises decompositions of the data at different frequencies of its power spectrum.

12. The system according to claim 4, wherein the stages of the training module are iterated between 10 and 10000 times.

13. The system according to claim 12, wherein a criterion for completion of the iterations is established, comprising stabilization of the performance metrics being analyzed.

14. The system according to claim 4, wherein the performance metrics comprise a loss function.

15. The system according to claim 4, wherein the interactions predicted by the training module are used to calculate the percentage inhibition of bacterial growth predicted by the system.

16. The system according to claim 15, wherein the percent inhibition of bacterial growth predicted by the system is converted to categorical variables.

17. The system according to claim 16, wherein the categorical variables comprise 3 performance categories: null, comprising a percentage of inhibition less than or equal to 15%; partial, comprising a percentage of inhibition greater than 15% and less than 85%; and total, comprising a percentage of inhibition greater than or equal to 85%.

18. The system according to claim 16, wherein the categorical variables comprise percentage ranges of 30%, of 10%, or of 1%.

19. The system according to claims 16 to 18, wherein the categorical variables are used for measuring performance metrics, wherein the performance metrics comprise accuracy, precision, recall, f-score, among other related.

20. The system according to claim 1, wherein the prediction module is configured to at least: a) use the trained models with fixed parameters set in the training module; b) receive a set of data not seen by the models in the training module; c) for each input data, generate n predictions for each time t; d) based on the predictions for each time instant, estimate confidence intervals; e) pair the predicted interaction predictions with the corresponding control interactions to estimate the percentage inhibition at time t+1.

21. The system according to claim 20, wherein the number of predictions (n) generated by the module comprises from 100 to 1000 predictions.

22. The system according to claim 20, wherein the interaction predictions comprise bacterial growth curves from calculation of the optical density of the bacterial culture in the presence of a cocktail of bacteriophage.

23. The system according to claim 22, wherein the bacterial growth curves allow calculation of the percentage inhibition of the bacteriophage cocktail on the growth of the bacterial isolate as a function of the ratio of the areas under the curve between the interaction curve and the control curve.

24. The system according to claim 23, wherein the percentage inhibition of bacterial growth predicted by the system is converted to categorical variables.

25. The system according to claim 24, wherein the categorical variables comprise null performance categories, comprising a percentage of inhibition less than or equal to 15%; partial, comprising a percentage of inhibition greater than 15% and less than 85%; and total, comprising a percentage of inhibition greater than or equal to 85%.

26. The system according to claim 24, wherein the categorical variables comprise percentage ranges of 30%, of 10%, or of 1%.

27. The system according to claims 24 to 26, wherein the categorical variables are used for the measurement of performance metrics, wherein the performance metrics comprise accuracy, precision, recall, f-score, among other related.

28. The system according to claim 2, wherein the pre-processing module is configured to at least: a) receive sequential interaction data between a bacteriophage cocktail and a bacterial isolate of interest with n samples; b) optionally filter said sequential interaction data to reduce noise associated with said data; c) optionally, transform the filtered sequential interaction data to a range appropriate for use by the following modules.

29. The system according to claim 28, wherein the sequential interaction data comprises a plurality of data from different interactions between different bacteriophage cocktails and different bacterial isolates at a time t and at a time t+1.

30. The system according to claim 28, wherein filtering the sequential interaction data comprises using low-pass filters, high-pass filters, band-pass filters, or other related filters.

31. The system according to claim 30, wherein the filters are selected based on the quality or type of sequential interaction data used.

32. The system according to claim 30, wherein transforming the sequential interaction data comprises rescaling the data to a range [0,1], rescaling the data to a range [-1,1], standardization, normalization, quantile transformation.

33. The system according to claim 3, wherein the post-processing module is configured to at least: a) in case the results obtained by the prediction module contain results from more than one bacteriophage cocktail for the same set of bacterial isolates, produce a ranking based on the median bacterial growth inhibition; b) compare the performance of the cocktails in pairs; c) optionally, generate a report including interaction curves predicted by the system, their transformation to categorical variables, rankings and comparisons between cocktails performed by the system; d) optionally, visualize these results on a platform.

34. The system according to claim 33, wherein the comparison between pairs of bacteriophage cocktails is performed by paired tests for comparison, comprising Wilcoxon test, Hodges-Lehmann test.

35. The system according to claim 33, wherein the module comprises an additional feedback function, wherein an operator analyzes the results reported by the system and can manually correct parameters according to the experimental results.

36. A computer-implemented system for predicting the inhibitory effect of a bacteriophage cocktail on the growth of bacterial isolates, wherein the system comprises: a pre-processing module, configured to pre-process interaction data between a bacteriophage cocktail and a bacterial isolate for normalization using filters and / or transformations; a training module, configured to train machine learning models to identify patterns and data structures from interaction data between different bacteriophage cocktails and bacterial isolates at times t and t+1 to adjust prediction parameters; a prediction module, configured to predict the interaction between a bacteriophage cocktail and a bacterium at time t+1 from the interaction data at time t using the trained learning models; and,a post-processing module configured to produce rankings and comparisons between the different bacteriophage cocktails and bacterial isolates analyzed.

37. A computer-implemented method for predicting the inhibitory effect of a bacteriophage cocktail on the growth of bacterial isolates, wherein the method comprises the stages: i) train machine learning models to identify patterns and data structures from interaction data between different bacteriophage cocktails and bacterial isolates at times t and t+1 to fit prediction parameters; ii) predict the interaction between a bacteriophage cocktail and a bacterium at time t+1 from the interaction data at time t using the trained learning models.

38. The method according to claim 37, further comprising the stage: iii) pre-process interaction data between a bacteriophage cocktail and a bacterial isolate for normalization using filters and / or transformations, wherein said stage is upstream of the stages described in claim 37.

39. The method according to claim 37 and 38, further comprising the stage: iv) produce rankings and comparisons between the different bacteriophage cocktails and bacterial isolates tested, wherein said stage is downstream of the stages described in claim 37.

40. The method according to claims 37 to 39, wherein stage (i) comprises at least: a) separate a normalized set of sequential individual interaction data between different bacteriophage cocktails and different bacterial isolates into training, validation, and test data subsets; b) select appropriate machine learning models, c) receive, in each selected model, the first subset of sequential interaction data, which corresponds to the training data subset; d) perform, on each model, transformations, decompositions or a composition of several transformations and compositions of such data to compact feature vectors or compact representations; e) identify patterns or structures of the data in the subset of training data; f) predict the interaction between a bacteriophage cocktail and a bacterial isolate at time t+1 from data of such interaction at time t;g) compare said interaction prediction at time t+1 with the actual value of said interaction by means of performance metrics; h) receive, for each selected model, the second subset of sequential individual interaction data, which corresponds to the validation data subset; i) validate the results using the validation subset under the same performance metrics used with the training subset for each selected model.

41. The method according to claim 40, wherein selecting suitable learning models comprises selecting 1 to 10 learning models with different levels of complexity.

42. The method according to claim 41, wherein the selecting models comprises selecting 3 to 5 learning models.

43. The method according to claim 41, wherein the learning models are selected from supervised learning models.

44. The method according to claim 41 , wherein the learning models are selected from unsupervised learning models.

45. The method according to claim 40, wherein identifying patterns or structures of the data comprises modeling patterns by function blocks.

46. The method according to claim 45, wherein the function blocks comprise operations such as matrix multiplications, one-dimensional convolutions, or combinations thereof, configured to encode the time series into compact feature vectors.

47. The method according to claim 40, wherein identifying patterns or structures of the data comprises decompositions of the data at different frequencies of their power spectrum.

48. The method according to claim 40, wherein the steps of stage i) are iterated from 10 to 10000 times.

49. The method according to claim 48, wherein a criterion for completion of the iterations is established, comprising stabilizing the performance metrics being analyzed.

50. The method according to claim 40, wherein the performance metrics comprise a loss function.

51. The method according to claim 40, wherein the interactions predicted by stage i) are used to calculate the percent inhibition of bacterial growth predicted by the method.

52. The method according to claim 51, wherein the percentage inhibition of bacterial growth predicted by the method is converted to categorical variables.

53. The method according to claim 52, wherein the categorical variables comprise 3 performance categories: null, comprising a percentage inhibition less than or equal to 15%; partial, comprising a percentage inhibition greater than 15% and less than 85%; and total, comprising a percentage inhibition greater than or equal to 85%.

54. The method according to claim 52, wherein the categorical variables comprise percentage ranges of 30%, of 10%, or of 1%.

55. The method according to claims 52 to 54, wherein the categorical variables are used for measuring performance metrics, wherein the performance metrics comprise accuracy, precision, recall, f-score, among other related metrics.

56. The method according to claim 37, wherein stage ii) comprises at least: a) use the trained models with fixed parameters set in stage i); b) receive a set of data not seen by the models in stage i); c) for each input data, generating n predictions for each time t; d) based on the predictions for each time instant t, estimate confidence intervals; e) pair the predicted interaction predictions with the corresponding control interactions to estimate the percentage inhibition at time t+1.

57. The method according to claim 56, wherein the number of predictions (n) generated by stage ii) comprises 100 to 1000 predictions.

58. The method according to claim 56, wherein the interaction predictions comprise bacterial growth curves from calculation of the optical density of the bacterial culture in the presence of a bacteriophage cocktail.

59. The method according to claim 58, wherein the bacterial growth curves enable calculation of the percentage inhibition of the bacteriophage cocktail on the growth of the bacterial isolate as a function of the ratio of the areas under the curve between the interaction curve and the control curve.

60. The method according to claim 59, wherein the percentage inhibition of bacterial growth predicted by the method is converted to categorical variables.

61. The method according to claim 60, wherein the categorical variables comprise null performance categories, comprising a percentage of inhibition less than or equal to 15%; partial, comprising a percentage of inhibition greater than 15% and less than 85%; and total, comprising a percentage of inhibition greater than or equal to 85%.

62. The method according to claim 60, wherein the categorical variables comprise percentage ranges of 30%, of 10%, or of 1%.

63. The method according to claims 60 to 62, wherein the categorical variables are used for measuring performance metrics, wherein the performance metrics comprise accuracy, precision, recall, f-score, among other related metrics.

64. The method according to claim 38, wherein stage (iii) at least: a) receive sequential interaction data between a bacteriophage cocktail and a bacterial isolate of interest with n samples; b) optionally, filter said sequential interaction data to reduce noise associated with said data; c) optionally, transform the filtered sequential interaction data to a range suitable for use by stages i) and ii).

65. The method according to claim 64, wherein the sequential interaction data comprises a plurality of data of different interactions between different bacteriophage cocktails and different bacterial isolates at a time t and / or a time t+1.

66. The method according to claim 64, wherein filtering the sequential interaction data comprises using low-pass filters, high-pass filters, band-pass filters, or other related filters.

67. The method according to claim 66, wherein the filters are selected based on the quality or type of sequential interaction data used.

68. The method according to claim 64, wherein transforming the sequential interaction data comprises rescaling the data to a range [0,1], rescaling the data to a range [-1,1], standardization, normalization, quantile transformation.

69. The method according to claim 39, wherein stage (iv) comprises at least: a) in case the results obtained by stage ii) contain results from more than one bacteriophage cocktail for the same set of bacterial isolates, produce a ranking based on the median bacterial growth inhibition; b) compare the performance of the cocktails in pairs; c) optionally, generate a report including interaction curves predicted by the method, their transformation to categorical variables, rankings and comparisons between cocktails performed by the method; d) optionally, visualize these results on a platform.

70. The method according to claim 69, wherein the comparison between pairs of bacteriophage cocktails is performed by paired tests for comparison, comprising Wilcoxon, Hodges-Lehmann test.

71. The method according to claim 69, wherein stage iv) comprises an additional feedback function, wherein an operator analyzes the results reported by the method and can manually correct parameters according to the experimental results.

72. A computer-implemented method for predicting the inhibitory effect of a bacteriophage cocktail on the growth of bacterial isolates, wherein the method comprises the stages: pre-process interaction data between a bacteriophage cocktail and a bacterial isolate for normalization by means of filters and / or transformations; train machine learning models to identify patterns and data structures from interaction data between different bacteriophage cocktails and bacterial isolates at times t and t+1 to fit prediction parameters; predict the interaction between a bacteriophage cocktail and a bacterium at time t+1 from the interaction data at time t using the trained learning models; and,elaborate rankings and comparisons between the different bacteriophage cocktails and bacterial isolates analyzed.

73. Use of the system and method according to claims 1 to 72 for predicting the inhibitory effect of a bacteriophage cocktail on the growth of bacterial isolates.

Images