Machine learning model for predicting therapeutic vulnerabilites in patient cancer tissue samples

Abstract

Machine learning models to predict, based on genomic data and gene expression data associated with a particular patient cancer tissue sample, one or more characteristics for that particular patient cancer tissue sample. The machine learning models are trained using respective feature sets representing one or more topological characteristics of respective protein-protein interaction, "PPI", networks representing models of different cancers, and that indicate reaction pathways associated with the protein interactions.

Description

[0001] 165131 - Pearl A. I. - DependANT+

[0002] MACHINE LEARNING MODEL FOR PREDICTING THERAPEUTIC VULNERABILITES IN PATIENT CANCER TISSUE SAMPLES

[0003] Technical Field

[0004] The technology described herein relates to cancer treatment, and in particular to the training and deployment in computer program products of machine learning models that are operable to analyse genomic data of patient cancer tissue samples to identify a type of cancer and / or therapeutic vulnerabilities in the cancer, to thereby assist oncologists in developing improved personalised treatment plans.

[0005] Background

[0006] Cancer is a leading cause of death globally. For instance, in 2020 there were 19.3 million new cases and 10 million deaths from cancer worldwide.

[0007] Traditional cancer treatment is typically based on the primary site of origin of the cancer. Thus, if cancer appears in the liver, for example, this is typically then treated as liver cancer, e.g. using standard treatment plans for liver cancer.

[0008] However, this approach may not then be effective for cancers having unknown primary or that do not present as typical (e.g. liver) cancer.

[0009] Further, cancer is a complex genetic disease arising from the stepwise accumulation of mutations in a single cell. Each patient’s cancer arises from a distinct combination of mutations interacting with a unique genetic background, resulting in a highly heterogeneous disease that is challenging to treat. For instance, the genetic makeup of cancer cells may be unique to the cancer and unique to the individual, so that any standard treatment plan that is designed for the ‘average’ cancer may or may not therefore be effective in treating a particular cancer, as each individual may respond differently to standard treatments.

[0010] Traditional first-line therapies such as chemotherapy and radiotherapy may therefore be unable to address this genetic diversity. Precision medicine, based on genomic analysis, may offer opportunities for improvements in this regard by shifting focus from one-size-fits-all treatment plans towards more targeted therapeutic strategies that can be tailored to individual patient’s cancers. For example, targeted therapies can exploit specific molecular vulnerabilities in a patient’s cancer while leaving normal cells relatively unharmed, increasing the efficiency of the treatment,and reducing unwanted side effects. Such approaches are, however, currently typically not available to the majority of healthcare providers (e.g. for reasons of time / cost).

[0011] The inventors therefore believe that there remains scope for improvements in this regard.

[0012] Summary

[0013] According to a first aspect of the technology described herein there is provided a method of training one or more machine learning model to predict, based on genomic data and gene expression data associated with a particular patient cancer tissue sample, a set of one or more essential genes for that particular patient cancer tissue sample, the method comprising:

[0014] obtaining, for training data, a set of plural “feature sets”, each feature set representing one or more topological characteristics of a respective protein- protein interaction (PPI) network from which the feature set has been derived, wherein respective PPI networks include a respective plurality of nodes and edges, each edge connecting a respective pair of nodes within the PPI network, wherein respective nodes represent respective proteins and respective edges represent interactions between the respective proteins represented by the nodes that are connected by that edge, and wherein respective edges within the PPI network have an associated directionality indicating a reaction pathway associated with the interaction between the respective proteins represented by the nodes that are connected by that edge,

[0015] wherein the nodes within different PPI networks are configured to reflect a mutational status of the gene encoding the proteins represented by those nodes, different PPI networks thereby representing models of different cancers, and each feature set being labelled with information indicative of which nodes within the corresponding PPI network from which that feature set was derived represent essential genes for the cancer represented by that PPI network; and the method further comprising:

[0016] training the one or more machine learning model using the obtained feature sets derived from the different PPI networks to be able to predict respective sets of essential genes for patient cancer tissue samples.

[0017] Once the machine learning model (or models) has been trained in this way, it can then be suitably deployed and used to predict essential genes for particular patient cancer tissue samples. A user (e.g. an oncologist) may accordingly be ableto obtain from a cancer biopsy suitable genomic data and gene expression data for that particular patient cancer tissue sample and then provide this data, or suitable data derived therefrom, as input to the (trained) machine learning model(s), which then predicts a set of one or more essential genes for that particular patient cancer tissue sample. This information can in turn help to guide improved therapeutic strategies, for example.

[0018] Thus, once the machine learning model(s) has been trained, it can then be (and in embodiments is) deployed, and used, to predict essential genes for particular patient cancer tissue samples. To do this, a suitable “feature set”, i.e. in a corresponding and consistent format to that used for training the one or more machine learning model(s) should be, and therefore is, obtained, which feature set relates to an appropriate PPI network that can be derived from the genomic data and gene expression data for the particular patient cancer tissue sample being analysed. This feature set is then processed by the machine learning model(s) to identify essential genes for the patient cancer tissue sample in question.

[0019] A second aspect of the technology described herein therefore provides a method of predicting essential genes for a particular patient cancer tissue sample, the method comprising:

[0020] obtaining a respective feature set representing one or more topological characteristics of a respective protein-protein interaction (PPI) network model that has been derived from a set of genomic data and gene expression data for the particular patient cancer tissue sample; and

[0021] providing the obtained feature set for the particular patient cancer tissue sample to one or more machine learning model that has been trained in the manner described above for processing,

[0022] the one or more machine learning model outputting a set of one or more predicted essential genes for the particular patient cancer tissue sample.

[0023] Thus, the machine learning model(s) of the technology described herein can be, and in embodiments is, used for guiding therapeutic strategies.

[0024] In this respect, as will be explained further below, the inventors have found that machine learning models trained in this way, as a result of the novel training process, are able to provide more accurate predictions of essential genes for patient cancer tissue samples. Further, the training process can be done using information and data that is easily accessible, and for which corresponding data (that can be input to the machine learning model(s) for processing) can readily be obtained by users from real-life patient cancer tissue samples.The effect of this is therefore to provide a system or methodology for providing more, and more meaningful, information to users (which information may not otherwise be accessible). This can therefore allow the users to make better, e.g. more patient-specific, decisions on therapeutic strategies.

[0025] In general, however, the machine learning model(s) could also be used in other situations, such as in research environments, e.g. for predicting essential genes for (novel) cancer cell lines, in particular where there is little or no experimental data available. Thus, although various examples will be described with respect to analysing patient cancer tissue samples, it will be appreciated that the (same) machine learning model(s) may also be used for analysing cancer cell lines, etc., as appropriate.

[0026] For example, in embodiments, a machine learning model that is produced by the method according to the first aspect of the technology described herein may be deployed, e.g., as part of a web-based system, e.g., and in embodiments, with the machine learning model residing on the ‘cloud’ and with the user (e.g. an oncologist) accessing the machine learning model via a suitable web interface or other local application through which they can upload the patient cancer tissue sample-derived data to be processed.

[0027] In this respect, it will be appreciated that the user could directly upload a feature set derived from the particular patient cancer tissue sample for analysis. That is, the conditioning and processing of the underlying genomic data and gene expression data into the relevant format, i.e. a suitable “feature set”, for processing by the machine learning model could be performed locally, by the user.

[0028] In other embodiments, however, the user may upload the underlying genomic data and gene expression data, with suitable conditioning and / or processing of that data then being performed remotely, e.g., and in embodiments, in an automated manner, on the ‘cloud’. For instance, the underlying genomic data and gene expression data may be processed to generate a respective PPI network, and the feature extraction can then be performed on the generated PPI network, e.g. by suitable software that is operable and configured to do this.

[0029] Thus, the input data provided by the user may already comprise a suitable feature set, i.e. in the desired format for analysis, or could comprise a PPI network that is then processed, in embodiments automatically, e.g. by appropriate software, to extract a feature set in the desired format for analysis. In embodiments, however, the input data provided by the user comprises underlying genomic data and geneexpression that is then used to generate a personalised PPI network model of the individual cancer.

[0030] For instance, in particular embodiments, there are one or more baseline PPI networks, which may, e.g., be derived from existing pathway databases such as the Reactome, Pathway Commons, or Kegg databases, and which baseline PPI networks represent a network of known physical and / or function protein interactions. The genomic data and gene expression derived from a particular patient cancer tissue sample can then be (and in embodiments is) used to perturb (modify) one or more of the baseline PPI networks to generate a respective ‘personalised’ PPI network for the individual cancer.

[0031] Then, in embodiments, (automated) feature extraction is performed on the personalised PPI network generated for the particular patient cancer tissue sample to extract a feature set in the desired format for analysis. The feature extraction may be performed in any suitable and desired manner. For example, the personalised PPI network generated for the particular patient cancer tissue sample may be loaded into suitable network analysis tools, such as the igraph library in Python, where suitable topological modelling and feature extraction is performed.

[0032] Of course, other arrangements would be possible.

[0033] In any case, the feature set that is obtained and analysed is based on the set of genomic data and gene expression data for the particular patient cancer tissue sample being analysed.

[0034] Various arrangements would be possible in this regard.

[0035] The results of the classification are in embodiments then provided, e.g. for display, back to the user via the same web interface and / or application. Various other suitable arrangements may be used in this regard, as desired.

[0036] The technology described herein also extends to machine learning models, and to computer program products containing the same, that have been trained in this way.

[0037] It will be appreciated from the above that the technology described herein thus generally relates to the training and use of machine learning models that are operable to analyse data derived from patient cancer tissue samples (i.e. from cancer biopsies) to (try to) identify certain characteristics of the cancer, in particular in the form of a set of “essential” genes for that particular cancer. This information can thus be used to identify therapeutic vulnerabilities in the cancer and to thereby assist oncologists in developing improved personalised treatment plans.In this respect, it will be appreciated that cells are typically dependent on a subset of genes for survival, and these are referred to as "essential genes". An essential gene is thus one on which a cell depends for cellular survival.

[0038] Genes may be essential across all cells and tissue types (i.e. “pan-essential” genes), or only essential in specific contexts and environments (so-called “context essential” genes). During carcinogenesis, the accumulation of genetic mutations causes cancer cells to develop novel gene dependencies that are not present in normal cells. For example, the cell may become addicted to oncogenes and / or tumour suppressor genes may become inactivated. Consequently, the cell develops new gene dependencies as different genes become transiently essential to that cell as it evolves. Thus, the set of essential genes for a particular cancer cell line or patient sample is typically context specific, depending on the cell type, genetic and epigenetic aberrations, and the environment the cell finds itself in.

[0039] These gene dependences represent potential precision drug targets and have been the focus of intense research in recent years. For instance, the arrival of CRISPR gene editing technology has enabled the exploration of cancer gene dependencies on an unprecedented scale, but it is not currently practical to perform CRISPR screens on every patient.

[0040] More recently, therefore, computational methods have emerged which can identify gene dependencies from easily obtained biological data. Various machine learning algorithms have been proposed in this regard for predicting essential genes, most of which focus on the prediction of pan-essential genes or identifying dependencies in specific cancer types rather than individual tumours and cell lines.

[0041] An exception to this is the so-called “DependANT” algorithm described, e.g., in Benstead-Hume et al. ‘Biological network topology features predict gene dependencies in cancer cell-lines’, Bioinformatics Advances, 2022, 1-8, which uses protein-protein interaction (PPI) data derived from the STRING database, as well as expression and mutation data from cancer cell lines, to train random forest models to predict gene dependencies across different tissue types and in individual cancer cell lines.

[0042] The DependANT algorithm thus uses personalised PPI networks to model the specific transcriptional and genomic changes present in individual cancer cell lines and trains a model on this basis to predict essential genes in previously unlabelled cell lines. Underlying this predictive ability is the centrality-lethality rule which suggests that essential genes tend to be highly expressed, highly connected, and close to the centre of PPI networks. Consequently, topological characteristicsthat measure node centrality (e.g. Degree, Betweenness and Closeness), can be used to successfully predict essential nodes within a PPI network.

[0043] The DependANT algorithm can therefore provide an improved, and “tissuetype agnostic” approach for predicting essential genes. For instance, in practice, many cancers do not behave as might be ‘expected’ due to rewiring of the cancer cell’s regulatory mechanisms by the many mutations and epigenetic changes that accompany the cancer phenotype. The approach used in the DependANT algorithm thus takes these genetic and epigenetic changes into account, thereby allowing a specific prediction of the essential genes in individual cancer cell lines.

[0044] The technology described herein thus takes a similar approach to that used in the DependANT algorithm, i.e. using personalised PPI networks to train a machine learning model to predict genes that individual cancers have become dependent on, and to be able to do this in a tissue-type agnostic manner.

[0045] According to the technology described herein, however, the edges within the PPI networks (each) have an associated “directionality” indicating a particular reaction pathway associated with an interaction between the respective proteins represented by the nodes that are connected by that edge. For example, there will be various physical and / or functional interactions between respective proteins within the PPI networks, and the edge directionality in embodiments reflects the directionality of these physical and / or functional interactions.

[0046] That is, rather than the edges simply connecting nodes between which there is some potential interaction, the edges in the PPI networks that are used in the training process for the machine learning models of the technology described herein are configured to also indicate the (direction of the) reaction pathway associated with the physical and / or functional interaction between the respective proteins represented by the nodes that are connected by that edge.

[0047] Correspondingly, as will be described further below, the one or more topological characteristics in the feature set representing the PPI network, will also reflect this directionality within the PPI network (as, for instance, the topological characteristics will also take into account the directionality within the network). That is, in embodiments, a (and each) PPI network is a directed graph, and at least some of the topological characteristics that are used according to the technology described herein are defined for the directed graph, so that the calculation of the values for these topological characteristics will also reflect the edge directions (that is, the values calculated for such topological characteristics will generally differ for a directed graph compared to those values calculated on its undirected or bidirectionalcounterpart). For instance, the one or more topological characteristics in embodiments include indicators of node centrality, and these topological characteristics can be defined for the directed graph (i.e. PPI network) accordingly. In some embodiments, the topological characteristics that are used may directly capture directionality, but various arrangements would be possible in this regard.

[0048] In embodiments, the edges are further associated with respective edge weights, the edge weight being calculated as a function of gene expression so that the edge weight decreases towards zero with increasing gene expression. In embodiments, each edge will be associated with a pair of edge weights representing the respective gene expression levels for the genes encoding the proteins represented by the pair of nodes that are connected by that edge.

[0049] In this respect, the inventors have found that including such pathway data including the directionality of the interactions, as well as the expression levels of the different proteins, into the PPI networks that are used for training the machine learning model can provide significant improvements in accuracy (e.g. compared to using unweighted edges and / or approaches that do not take account of the directionality of the reaction pathway associated with the physical and / or functional interaction between respective proteins).

[0050] For instance, previous studies have generally assumed a relatively static model, in which genes are either essential or not. However, the inventors recognise that the effect of genetic variations can be transmitted from directly affected proteins to distant gene products through protein interaction pathways, which suggests that a more dynamic model of the PPI network may be appropriate to capture gene dependencies in highly mutated cancer cells. The DependANT algorithm thus represents an improvement compared to more traditional approaches. However, the inventors now recognise that incorporating directional pathway data into the PPI network can provide yet further improvements in this regard.

[0051] For example, in the DependANT algorithm as described in Benstead-Hume et al. ‘Biological network topology features predict gene dependencies in cancer cell-lines’, Bioinformatics Advances, 2022, 1-8, each pair of nodes is connected by two ‘unidirectional’ edges, with each edge having a single edge weight that is indicative of the gene expression of one of the nodes (i.e. such that one of the edges connecting a pair of nodes indicates the expression for the gene encoding the protein represented by one of the nodes and the other edge indicates the expression for the gene encoding the protein represented by the other one of the nodes). The direction of the edges in this case is therefore simply a mechanism forassociating a particular edge with one or other of the nodes that it is connected to, but does not reflect the directionality of the reaction pathways associated with the physical and / or functional interactions between the respective proteins represented by the nodes within the PPI networks (which directionality is now taken into account according to the technology described herein).

[0052] The inventors have also found that using graph neural networks (GNNs) can provide further improvements in this regard. GNNs enable the application of deep learning techniques to graphs, allowing for tasks such as node classification and link prediction by learning features from the graph’s structure and node relationships. The inventors thus recognise that in the context of the technology described herein, where the machine learning model is trained using PPI networks that take into account reaction pathways within the PPI network, the use of GNNs may be particularly effective at capturing longer-distance interactions between proteins, so that knock-on effects further down the cancer pathway can be better captured.

[0053] Thus, whilst the machine learning models of the technology described herein may in general be built upon any suitable and desired machine learning models, in embodiments, the machine learning model(s) comprises one or more GNN (and the training that is performed according to the first aspect of the technology described herein is accordingly to train a corresponding one or more GNN).

[0054] The combination of training the machine learning model(s) using directional networks and the machine learning model(s) comprising one or more GNNs may therefore be particularly effective at capturing longer-distance impact of gene mutations within the PPI networks, and hence can provide improved (e.g. more accurate) predictions of essential genes.

[0055] The machine learning models according to the technology described herein may therefore provide various improvements, e.g. in terms of accuracy of its predictions, compared to the previously described DependANT algorithm (and hence also compared to other approaches).

[0056] This in turn means that the machine learning model of the technology described herein facilitates improved identification of the relevant characteristics of the patient cancer tissue samples being analysed, e.g. the set of essential genes. The machine learning models according to the technology described herein can thus provide a useful, and readily accessible, tool to facilitate improved decision making, and in turn improved clinical outcomes.

[0057] The technology described herein may therefore provide various benefits compared to other possible approaches.The machine learning model(s) of the technology described herein may be any suitable and desired machine learning model.

[0058] In some embodiments of the technology described herein, the machine learning model(s) are machine learning “classifier” models.

[0059] In other embodiments of the technology described herein, the machine learning model(s) are machine learning regression models. Such regression models may allow the models to (directly) predict gene dependency scores as continuous values (rather than predictions within binary categories of dependent versus nondependent which may be the case for classifier models).

[0060] Subject to the particular requirements of the technology described herein the training of the machine learning model(s) may be done in any suitable and desired manner.

[0061] As described above, the training data that is used to train the machine learning model(s) of the technology described herein is derived from a set of plural PPI networks, wherein different PPI networks within the set of plural PPI networks represent different cancer cell lines.

[0062] The set of PPI networks that is used to derive the training data can itself be obtained in any suitable and desired manner. In embodiments, a base set of PPI networks are derived from existing pathway databases such as the Reactome, Pathway Commons, or Kegg databases, as this data will already reflect the directional relationships (pathways) between different proteins. In embodiments, the data is then further conditioned or curated to generate the training data that is used according to the technology described herein.

[0063] For example, as discussed as above, a respective PPI network includes a plurality of nodes and a plurality of edges, each edge connecting a respective pair of nodes, wherein respective nodes represent respective proteins and respective edges represent physical and / or functional interactions between the respective proteins represented by the nodes that are connected by that edge.

[0064] As discussed above in relation to extracting the feature sets for the patient cancer tissue samples that are to be analysed, the nodes within a particular PPI network can thus be (and are) configured to reflect the mutational status of the gene encoding the protein represented by that node for the particular cancer cell line that the PPI network represents.

[0065] For instance, during the process of carcinogenesis, a cell may become addicted to activated oncogenes and / or tumour suppressor genes may become inactivated. Consequently, the cell may develop new gene dependencies asdifferent genes become transiently essential to the cell as it mutates. Nodes in the PPI networks are therefore altered by mutational status of the gene they represent, for example such that mutations that disrupt the reading frame of the encoded protein results in the removal of that node from the network.

[0066] For example, missense mutations (i.e. that result in different amino acids being encoded at a particular position in the resulting protein) may or may not alter the function of the resulting protein. Thus, in embodiments, missense mutations can be assigned as Loss-of-Function (LoF), Gain-of-Function (GoF) or neutral. This can be done by a number of methods, but in embodiments is done using another machine learning model that has been trained on a, in embodiments hand curated proprietary, dataset deriving information from the published literature. Mutations classified as LoF result in the removal of the node representing the mutated gene (and reweighting of associated edges) from the PPI network, while mutations classified as GoF result in an increase in the weighting associated with the corresponding node (and reweighting of associated edges) in the network. Neutral mutations have no effect on the topology or weighting of the PPI network.

[0067] The PPI networks thus represent models of different cancers, i.e. different cancer cell lines or patient samples, wherein the nodes and edges within different PPI networks are configured to reflect the mutational status of the gene encoding the protein represented by a given node, and its interactions represented by the associated edges.

[0068] To train the machine learning model(s) of the technology described herein, appropriate feature extraction is then performed to extract, for respective PPI networks within the obtained set of plural PPI networks, respective "feature sets” representing one or more topological characteristics of the PPI network.

[0069] Thus, the training process will include obtaining, as / for training data, a set of plural “feature sets” that have been derived from different PPI networks in the set of plural PPI networks. These feature sets could be obtained directly (e.g. from storage, if the feature extraction has been performed in advance), or as part of the training process itself.

[0070] That is, in some embodiments, obtaining the feature sets may comprise a (first) step of obtaining a set of plural PPI networks, wherein respective PPI networks within the set of plural PPI networks represent models of different respective cancers, and a (subsequent) step of extracting for respective PPI networks within the obtained set of plural PPI networks respective feature sets representing the one or more topological characteristics of the PPI network. In thiscase, the (subsequent) step of extracting the features may be, and in embodiments is, performed automatically, e.g. by appropriate software and / or neural network processing of the PPI networks, in embodiments in the same manner as discussed above in relation to the extraction of the feature sets for the patient cancer tissue samples that are to be analysed.

[0071] The one or more topological characteristics in embodiments represent indicators of node centrality within a directed graph (since the PPI network in the technology described herein is a directed graph). The one or more topological characteristics may, and in embodiments do, include one or more of: Betweenness; Constraint; Coreness; Degree; Eccentricity; Eigen centrality; Hub score;

[0072] Neighbourhood n size (for n = 1,2,6); and Page rank, similarly to what is described in relation to the “DependANT” algorithm in Benstead-Hume et al. ‘Biological network topology features predict gene dependencies in cancer cell-lines’, Bioinformatics Advances, 2022, 1-8, but with edge directionality taken into account so that the topological characteristics are calculated based on distances from / to nodes (rather than simply distances between nodes). In embodiments, the feature scores are suitably normalised, e.g. so that each feature score is independently scaled between 0 and 1. The one or more topological characteristics may then be used as feature sets for training the machine learning models of the technology described herein.

[0073] For example, topological features of the PPI network, such as Eigen centrality and Betweenness are effective at capturing longer-distance changes in networks, so the inventors have found that the inclusion of genetic alterations into the traditional PPI network model is an appropriate way of predicting altered / novel dependencies.

[0074] In some embodiments, the one or more topological characteristics may also include Closeness. However, the inventors have found that due to PPI networks typically being incomplete, it is often not possible to calculate Closeness (and so those PPI networks would potentially need to be removed from the training set if Closeness were to be included in the feature set). So, in some embodiments, Closeness is not included into the feature set. In that case, Harmonic centrality may be used instead. Thus, in some embodiments, the one or more topological characteristics include Harmonic centrality

[0075] The inventors have found that further improvements may be achieved by also including Transitivity into the feature set. It will be appreciated that Transitivity is a measure of the density of loops of length three (i.e. triangles) within the networkand can be calculated by the ratio between the number of closed triangles and the maximum possible number of closed triangles in the network. Higher Transitivity is thus indicative of higher probability of adjacent nodes being interconnected and this may be better indicative of context-essential genes.

[0076] Various other arrangements would be possible.

[0077] In general, the values of the topological features of a directed graph will be different to the corresponding values calculated on its counterpart undirected or bidirectional graph (and so the topological features will at least indirectly reflect the directionality within the PPI network). In embodiments, some or all of the topological characteristics that are used may directly capture directionality, as alluded to above. For example, rather than considering the total node Degree, the ‘In’ Degree (i.e. the number of edges coming into the node) or the ‘Out’ Degree (i.e. the number of edges coming out of the node) could be used for the directed graph.

[0078] The feature sets, representing the one or more topological characteristics, that are used for the training process are thus in embodiments labelled accordingly with information indicative of which nodes within the corresponding PPI network from which that feature set was extracted represent essential genes for the cancer cell line represented by that PPI network, and these labelled feature sets are then used for training the machine learning models of the technology described herein accordingly, e.g. by a suitable process of “supervised” learning.

[0079] It will be appreciated that the feature sets may also be annotated with any other suitable information, as desired, to facilitate the training process. For example, in addition to mutational signatures and RNA pathway profiles, it may also be desired for the machine learning model to take into account phenotypic data, proteomic data, methylation data, etc., and this can also be done.

[0080] It will also be appreciated here that the processing of the PPI networks to extract the feature sets (i.e. the feature extraction) may be performed in part manually or may be at least partly automated. For example, in some embodiments, as mentioned already above also in relation to extracting the feature sets for the particular patient cancer tissue samples being analysed, a suitable network analysis software program, such as the igraph library in Python, and / or neural network may be used to process the PPI networks to extract the desired respective "feature sets”. Various arrangements would be possible in this regard.

[0081] In this way, the machine learning models can be configured, i.e. through suitable training, to be able to predict tumour-specific essential genes.In embodiments, it is further identified which (predicted) essential genes are targetable by known cancer treatments.

[0082] For example, and in some embodiments, this may involve identifying a potential target, first gene whose dependency is predicted by the machine learning models described above as high in cell lines where the function of a second gene is disrupted (or activated) relative to its dependency in cell lines where it is not disrupted (or activated), e.g. where the first and second genes have a synthetic lethality (SSL) relationship.

[0083] In this regard, calculation of SSL relationships can be performed, for example, by statistical comparison of distributions of dependency scores predicted by a regression version of the machine learning models described above, in cell lines in which the second gene is disrupted compared with cell lines in which the second gene is not disrupted, using a standard metric such as Chi-squared.

[0084] Various other arrangements would however be possible.

[0085] Further, whilst various embodiments are described above in the context of machine learning models that are operable and configured to predict essential genes, it will be appreciated that similar techniques may also be used to predict other characteristics of cancers, including, for example, but not limited to, a (primary) site of origin of the cancer (i.e. a ‘type’ of cancer). In this respect, it will be appreciated that standard cancer treatments are often based on (assumed) site of origin. However, this may not always be apparent from the presentation of the cancer alone, particularly where the cancer has spread rapidly.

[0086] Thus, a further aspect of the technology described herein provides a method of training one or more machine learning model to predict, based on genomic data and gene expression data associated with a particular patient cancer tissue sample, a respective primary site of origin for that particular patient cancer tissue sample, the method comprising:

[0087] obtaining, for training data, a set of plural “feature sets”, each feature set representing one or more topological characteristics of a respective protein- protein interaction (PPI) network from which the feature set has been derived, wherein respective PPI networks include a respective plurality of nodes and a plurality of edges, each edge connecting a respective pair of nodes within the PPI network, wherein respective nodes represent respective proteins and respective edges represent interactions between the respective proteins represented by the nodes that are connected by that edge, and wherein respective edges within the PPI network have an associated directionality indicating a reaction pathway associatedwith the interaction between the respective proteins represented by the nodes that are connected by that edge,

[0088] wherein the nodes within different PPI networks are configured to reflect the mutational status of the genes encoding the proteins represented by those nodes, different PPI networks thereby representing models of different cancers, and each feature set being labelled with information indicative of a primary site of origin for the cancer cell line represented by the PPI network from which the feature set has been derived; and

[0089] the method further comprising:

[0090] training the one or more machine learning model using feature sets that have been derived from a plurality of different PPI networks to be able to predict respective primary sites of origin for patient cancer tissue samples.

[0091] The machine learning model trained in this way can then be used accordingly to predict a type of cancer (i.e. a primary site of origin) for a particular patient cancer tissue sample, e.g., and in particular, for cancers of unknown primary, which would otherwise be difficult to treat.

[0092] Thus, a yet further aspect of the technology described herein provides a method of predicting a type of cancer (e.g. a primary site of origin) for a particular patient cancer tissue sample, the method comprising:

[0093] obtaining a respective feature set representing one or more topological characteristics of a respective protein-protein interaction (PPI) network model that has been derived from a set of genomic data and gene expression data for the particular patient cancer tissue sample; and

[0094] providing the obtained feature set for the particular patient cancer tissue sample to one or more machine learning model that has been trained according to the above aspect,

[0095] the one or more machine learning model outputting a predicted type of cancer (primary site of origin) for the particular patient cancer tissue sample.

[0096] More generally, the techniques of the technology described herein may be used to predict any other characteristics of cancers that may be of interest for guiding therapeutic strategies, including, for example, drug response (i.e. the sensitivity of that sample to a set of drugs used to treat cancer).

[0097] Thus, from a broad aspect, there is provided a method of training one or more machine learning model to predict, based on genomic data and gene expression data associated with a particular patient cancer tissue sample, one ormore characteristics for that particular patient cancer tissue sample, the method comprising:

[0098] obtaining, for training data, a set of plural feature sets, each feature set representing one or more topological characteristics of a respective protein- protein interaction, “PPI”, network from which the feature set has been derived, wherein respective PPI networks includes a respective plurality of nodes and a plurality of edges, each edge connecting a respective pair of nodes within the PPI network, wherein respective nodes represent respective proteins and respective edges represent interactions between the respective proteins represented by the nodes that are connected by that edge, and wherein respective edges within the PPI network have an associated directionality indicating a reaction pathway associated with the interaction between the respective proteins represented by the nodes that are connected by that edge,

[0099] wherein the nodes within different PPI networks are configured to reflect the mutational status of the genes encoding the proteins represented by those nodes, different PPI networks thereby representing models of different cancers, and each feature set being labelled with information indicative of the one or more characteristics for the corresponding PPI network from which that feature set has been derived; and

[0100] the method further comprising:

[0101] training the one or more machine learning model using feature sets that have been derived from a plurality of different PPI networks to be able to predict the one or more characteristics for patient cancer tissue samples.

[0102] In this case, the one or more characteristics for that particular patient cancer tissue sample may thus comprise a set of essential genes, a primary site of origin, a drug response, or any other suitable and desired characteristics for that particular patient cancer that may be useful for guiding improved therapeutic strategies.

[0103] Similarly, the machine learning model once trained can then be used to predict any such characteristics, i.e. the characteristics that it is trained to predict.

[0104] Thus, another broad of the technology described herein provides a method of predicting one or more characteristics for a particular patient cancer tissue sample, the method comprising:

[0105] obtaining a respective feature set representing one or more topological characteristics of a respective protein-protein interaction (PPI) network model that has been derived from a set of genomic data and gene expression data for the particular patient cancer tissue sample; andproviding the obtained feature set for the particular patient cancer tissue sample to one or more machine learning model that has been trained according to the above aspect,

[0106] the one or more machine learning model outputting a predicted one or more characteristics for the particular patient cancer tissue sample.

[0107] It will be appreciated that the machine learning models according to these further aspects may generally be trained and used in the same manner as the machine learning models according to the first and second aspects described above, and may therefore comprise any or all of the features described in relation to those aspects, at least to the extent these are not mutually incompatible. For instance, the machine learning models according to these further aspects will generally be trained using a similar set of PPI networks as described above, but with the feature sets that are used for the training in embodiments then labelled with information indicative of a respective type of cancer (primary site of origin), or another suitable characteristic of the cancer.

[0108] The machine learning models according to the technology described herein can thus be used to predict various characteristics of cancers, including, for example, essential genes for that particular cancer and / or a (primary) site of origin (or ‘type’) of cancer, which characteristics can in turn be used to identify therapeutic opportunities for treating the cancer, or any other suitable and desired characteristics, such as drug response, etc..

[0109] A user (i.e. oncologist) may thus take a cancer biopsy and then perform suitable genomic sequencing of this patient cancer tissue sample to obtain a suitable set of genomic data including the gene expression data specific to that particular cancer. In this respect, it will be appreciated that a particular benefit of the technology described herein is that the machine learning models can be used to predict various characteristics of the specific cancer based on readily available genomic data and gene expression data (e.g. without requiring an expensive experimental CRISPR screen to be performed).

[0110] Subject to the particular requirements of the technology described herein, the genomic data that is used may comprise any suitable and desired genomic data, and this data may be obtained in any suitable way. For example, the genomic data may, and in embodiments does, comprise at least mutation data. This mutation data can be, and in embodiments is, obtained based on genomic / exomic sequencing of the patient cancer tissue sample. The genomic data may, however, and in some embodiments does, also comprise other data, such as copy numbervariation data, methylation data, etc., and any other suitable information that can and may desirably be derived from the genomic / exomic sequencing of the patient cancer tissue sample. In this respect, it will be appreciated that this type of data is readily obtainable (but traditionally the tools for analysing / using such data have not been available).

[0111] In embodiments, the genomic data is generated based on a comparison of the patient cancer tissue sample with DNA sequences derived either from healthy tissue from the same patient, or from comparisons with reference human genomes, e.g. GRCh38. P14. Other arrangements would however be possible in this regard.

[0112] The genomic data that is used may be provided in any suitable and desired format. For example, in embodiments, the mutation data may be provided as a Variant Calling Format data file, e.g. as is commonly used for this purpose.

[0113] A variety of labelled column-based data formats are commonly used for presentation of methylation, copy-number and gene expression data, all of which can be readily converted into a suitable and consistent format for input to the training process and / or to the machine learning model(s) for processing.

[0114] As noted above, the genomic data, in whatever form that takes, is in embodiments also combined with gene expression data, which is in embodiments obtained by gene expression profiling. For example, gene expression data may be derived from ‘read’ counts from Next Generation Sequencing of reverse transcribed RNA from the patient cancer tissue sample. The genomic data and gene expression data for the particular patient cancer tissue sample are then processed to determine respective weights for nodes and edges within a corresponding PPI network model for the patient cancer tissue sample.

[0115] This can be done in any suitable and desired manner, and may be done either manually or automatically, e.g. by software.

[0116] As discussed above, in particular embodiments, there are one or more baseline PPI networks and the genomic data and gene expression data for the particular patient cancer tissue sample is used to perturb (modify) one or more of the baseline PPI networks to generate a corresponding PPI network for the patient cancer tissue sample.

[0117] Once the PPI network has been generated, feature extraction can be performed to extract the desired feature set, e.g. in a consistent and corresponding format to the feature sets that were used as part of the training process.

[0118] Thus, in embodiments, a user may directly input a feature set for processing by the machine learning model(s). Alternatively, as mentioned above, a user mayinput the underlying genomic data and gene expression data, which is then processed (e.g. by software and / or neural network processing) into an appropriate PPI network and from which a suitable feature set is extracted that is then processed by the machine learning model(s). That is, in embodiments, the data that is input is first processed to extract a corresponding feature set that is then processed by the machine learning model(s) to predict the characteristics of the particular patient cancer tissue sample being analysed.

[0119] Various options would be possible in this regard.

[0120] Based on the input data, in whatever form it takes, the machine learning model (or to a set of machine learning models, as appropriate) will then perform the relevant processing to predict the characteristics of the particular patient cancer tissue sample being analysed.

[0121] The machine learning model(s) will then provide an appropriate output, e.g. a set of predicted essential genes and / or a predicted type of cancer (primary site of origin) for the particular patient cancer tissue sample, as above.

[0122] In embodiments, the output of the machine learning model in respect of the patient cancer tissue sample is combined with other information to provide the user (i.e. oncologist) with greater depth of information to guide the treatment plan.

[0123] For example, when the machine learning model is configured to output a predicted set of one or more essential genes for the particular patient cancer tissue sample being analysed, this is in embodiments also output together with personalised treatment data based on the predicted set of one or more essential genes for the particular cancer cell line. For example, this treatment data may include gene-specific ‘druggability’ information as well data as to the use, sideeffects and licensing status of the identified drugs. All of this information is in embodiments provided for output, e.g. for display to the user, together, e.g. in the form of an annotated report detailing the targeted therapeutic opportunities that have been identified.

[0124] Thus, in embodiments, when the set of one or more characteristics includes a set of one or more essential genes for that particular patient cancer tissue sample, it is then determined, based on the set of one or more essential genes that are identified for the particular patient cancer tissue sample identified by the machine learning model, a set of other information including a list of known drugs or biomarkers for targeting one or more of the identified essential genes. This set of other information including the list of known drugs or biomarkers for targeting one or more of the identified essential genes can then be, and in embodiments is, output(e.g. for display to a user) together with a list of the identified essential genes. In embodiments the output comprises a ranked list of identified essential genes, with the ranking either being based on a confidence score that an identified gene is essential, and / or a confidence level that the essential gene can be targeted.

[0125] In embodiments, the set of other information including the list of known drugs or biomarkers for targeting one or more of the identified essential genes is obtained automatically, e.g. by the output of the machine learning models then being further processed and in embodiments used to look up such other information. For example, this can be done using the ChEMBL database. The set of other information may also include an indication of which of these drugs are currently in clinical trials, or are currently for clinical trials, for which the patient may be eligible.

[0126] Thus, in embodiments the output of the machine learning model may be further processed or combined with other sources of information, including, for example, one or more of:

[0127] for a (and in embodiments each) essential gene, a set of licensed drugs and biomarkers that can target that gene;

[0128] for a (and in embodiments each) essential gene, a set of drugs and biomarkers currently in clinical trials.

[0129] As mentioned above, in embodiments, the machine learning models are executed within a cloud-based web server. Thus, the user can upload the relevant genomic data, or suitable data derived therefrom (e.g. a feature set, as discussed above), via a suitable interface to the cloud-based web server, and the cloud-based web server then performs all of the desired processing of that data, i.e. to execute the machine learning models, and to perform any further processing of the output from the machine learning models to return the desired information to the user in the desired format.

[0130] Various arrangements would be possible in this regard.

[0131] It will be appreciated that any and all of the optional features described herein with respect to examples of the present disclosure may be combined in any suitable combination as appropriate.

[0132] The methods in accordance with the technology described herein may be implemented at least partially using software e.g. computer programs. It will thus be seen that when viewed from further embodiments the technology described herein comprises computer software specifically adapted to carry out the methods herein described when installed on a data processor, a computer program element comprising computer software code portions for performing the methods hereindescribed when the program element is run on a data processor, and a computer program comprising code adapted to perform all the steps of a method or of the methods herein described when the program is run on a data processor. The data processor may be a microprocessor system, a programmable FPGA (field programmable gate array), etc..

[0133] The technology described herein also extends to a computer software carrier comprising such software which when used to operate a graphics processor, renderer or microprocessor system comprising a data processor causes in conjunction with the data processor the processor, renderer or system to carry out the steps of the methods of the technology described herein. Such a computer software carrier could be a physical storage medium such as a ROM chip, CD ROM, RAM, flash memory, or disk, or could be a signal such as an electronic signal over wires, an optical signal or a radio signal such as to a satellite or the like.

[0134] It will further be appreciated that not all steps of the methods of the technology described herein need be carried out by computer software and thus from a further broad embodiment the technology described herein comprises computer software and such software installed on a computer software carrier for carrying out at least one of the steps of the methods set out herein.

[0135] The technology described herein may accordingly suitably be embodied as a computer program product for use with a computer system. Such an implementation may comprise a series of computer readable instructions either fixed on a tangible, non-transitory medium, such as a computer readable medium, for example, diskette, CD-ROM, ROM, RAM, flash memory, or hard disk. It could also comprise a series of computer readable instructions transmittable to a computer system, via a modem or other interface device, over either a tangible medium, including but not limited to optical or analogue communications lines, or intangibly using wireless techniques, including but not limited to microwave, infrared or other transmission techniques. The series of computer readable instructions embodies all or part of the functionality previously described herein.

[0136] Those skilled in the art will appreciate that such computer readable instructions can be written in a number of programming languages for use with many computer architectures or operating systems. Further, such instructions may be stored using any memory technology, present or future, including but not limited to, semiconductor, magnetic, or optical, or transmitted using any communications technology, present or future, including but not limited to optical, infrared, or microwave. It is contemplated that such a computer program product may bedistributed as a removable medium with accompanying printed or electronic documentation, for example, shrink-wrapped software, pre-loaded with a computer system, for example, on a system ROM or fixed disk, or distributed from a server or electronic bulletin board over a network, for example, the Internet or World Wide Web.

[0137] Various examples of the technology described herein will now be described with reference to the accompanying drawings, in which:

[0138] Figure 1 shows an example system according to the present embodiments; Figure 2 shows schematically how a user can interact with the system according to the present embodiments;

[0139] Figure 3 shows an example of an output that may be presented to a user; Figure 4 shows schematically how a machine learning model may be trained according to the present embodiments;

[0140] Figure 5, Figure 6 and Figure 7 show various examples of PPI networks; and Figures 8 to 21 show the results of supporting testing that has been performed, as will be described further below.

[0141] The technology described herein relates to techniques for identifying therapeutic vulnerabilities in individual patient cancer tissue samples. In this respect, it will be appreciated that the genetic makeup of every new cancer patient’s tumour is unique, and so individuals may react differently to standard treatments. The ability to determine the genetic makeup of each patient’s tumour at different stages in the disease thus allows more targeted treatments to be applied.

[0142] The present embodiments particularly relate to a web-based system that allows users (e.g. oncologists) to upload genomic and expression data for a particular patient cancer tissue sample for analysis, either directly, or in the form of a “feature set” derived from a protein-protein interaction (PPI) network model of the particular patient cancer tissue sample that encodes the genomic and expression data, with the system then analysing that data and returning to the user certain information that can then help guide therapeutic strategies.

[0143] Figure 1 shows an exemplary system that can be used in the present embodiments that includes a plurality of client computing devices 100 associated with different end users that are operable to communicate over network 102 with server system 104. In embodiments, the server system 104 resides on the cloud such that the network 102 is the internet, but other arrangements would of course be possible. For example, the network 102 could be any suitable wide or local area network that allows the client computing devices 100 to access the server system104. Similarly, the client computing devices 100 may generally represent any suitable and desired forms of processing devices, for example, a personal computer or laptop, a tablet, personal digital assistant (PDA), smart phone, etc., as desired.

[0144] As will be described further below, the server system 104 is operable to execute one or more machine learning models, in this embodiment machine learning classifiers 106, that are particularly trained to be able to predict, from input mutation and expression data for a particular patient cancer tissue sample one or more characteristics of the patient cancer tissue sample. For example, in embodiments, the machine learning classifiers 106 are operable and configured (i.e. through appropriate training) to identify or predict a set of essential genes for the patient cancer tissue sample in question.

[0145] In other embodiments, the machine learning classifiers 106 may (also) be able to identify or predict a ‘type’ of cancer, i.e. a primary site of origin for the patient cancer tissue sample in question, or drug response, i.e. the sensitivity of that sample to a set of drugs used to treat cancer.

[0146] In general, therefore there may be various machine learning models that are trained and operable to predict different characteristics of patient cancer tissue samples, and the system may generally include any suitable such set of machine learning models. Further, as discussed above, the server system 104 may be operable to execute machine learning models other than machine learning classifier models, e.g. machine learning regression models.

[0147] Once suitably trained, the machine learning models, e.g. machine learning classifiers 106 can thus be deployed on the server system 104, and then accessed appropriately via a suitable web interface and / or application executing on the client computing devices 100, in order to allow users to upload suitable input data, i.e. a set of mutation and expression data for a particular patient cancer tissue sample to be analysed to the server system 104, and for the server system 104 to then execute the appropriate machine learning classifiers 106 to predict for the particular patient cancer tissue sample that is to be analysed.

[0148] Figure 2 is a flow chart showing schematically the user’s interaction with this system. Thus, as shown in Figure 2, a user (e.g. oncologist) first uploads a set of mutation and expression data for a particular patient cancer tissue sample to the web interface and / or application executing on their computing device 100 (step 200), and the data is then sent over the network 102 to the server system 104 for processing (step 201). The server system 104 then executes the appropriatemachine learning model, e.g. machine learning classifiers 106, for processing the data (step 202).

[0149] The output of the machine learning models, e.g. machine learning classifiers 106, for the patient cancer tissue sample in question may then be provided via the network 102 back to the client computing device 100 that issued the request, e.g. for display to the user of that client computing device 100. In this respect, the output of the machine learning models, e.g. machine learning classifiers 106, may be, and typically will be, combined with various other information. For example, as shown in Figure 2, the server system 104, prior to returning an output to the user (in step 204), first aggregates data relating to therapeutic opportunities for treating the particular cancer (step 203), and this data is then all returned to the user together.

[0150] For example, this aggregation may comprise, for a given set of essential genes predicted by the machine learning classifiers 106, searching one or more web sources and / or databases for known drugs or biomarkers (either on the market or in clinical trials) that can target those essential genes. This information can then be provided to the user together with the identified set of essential in an appropriate format to help guide the user’s clinical decisions.

[0151] Figure 3 shows an example from the system according to an embodiment. As shown in Figure 3, this may include an identification of acquired essential genes, together with a list of drugs that are associated with those genes (e.g. as determined from the ChEMBL database). The output also includes information on clinical trials using those drugs. The output may further include a list of identified mutations, which in this example includes both gain of function and loss of function mutations.

[0152] As described above, the machine learning classifiers 106 in the present embodiments are thus trained to be able to predict various characteristics of a cancer from genomic data obtained from a particular patient cancer tissue sample. This is done by an appropriate (supervised) training process, which is shown schematically in the flow chart in Figure 4. As will be described further below, this is done based on respective “feature sets” that are derived from a set of PPI networks that represent the mutational effects in different cancer cell lines. This process can be similarly applied to machine learning models other than machine learning classifier models, as appropriate, including regression models, for instance.

[0153] As shown in Figure 4, one or more baseline PPI networks are first obtained (step 400). For example, a baseline human PPI network, representing a network of known physical and / or functional protein interactions can be generated, in whicheach node represents a protein and each edge a known physical and / or functional protein interaction.

[0154] The baseline PPI network is then weighted, or ‘perturbed’, to more accurately represent different cancer cell lines (step 401). For example, mutations such as frame-shift, insertions and deletions (indels) or nonsense substitutions can be labelled as loss-of-function (LOF) mutations. For missense mutations, the pathogenic mutations can be identified and classified as gain-of-function (GOF) or LOF mutations depending on whether they came from oncogenes or tumour suppressor genes. Nodes that represented genes with inactivating mutations are removed from the baseline PPI network.

[0155] As well as modifying the nodes, the edges are weighted appropriately. For example, it is assumed that genes with either GOF mutations and / or high expression levels have greater impact than in the baseline network, and to model this, edges from the corresponding proteins in the PPI network relatively lower weights. This increases the flow of information and so strengthens the impact of the gene.

[0156] For each PPI network, a set of topological features for each node of the PPI is then calculated (step 402), e.g. as shown in Table 1.

[0157]

[0158] >>

[0159]

[0160] Table 1. Definitions of Topological Features

[0161] After normalization, where each feature score is independently scaled between 0 and 1 , these topological features are then used as feature sets fortraining the machine learning models (classifiers) of the present embodiments (step 403). Once the training has been completed, the server system can then return an output indicating that this has been done (step 404).

[0162] It is noted that the approach described so far above generally follows the same approach used to train the DependANT algorithm and which approach is described in Benstead-Hume et al. ‘Biological network topology features predict gene dependencies in cancer cell-lines’, Bioinformatics Advances, 2022, 1-8.

[0163] In the present embodiments, however, the PPI networks are further perturbed so that the edges indicate the directionality of reaction pathways between the proteins that they represent.

[0164] For instance, Figure 5 shows an example of a baseline PPI network including three nodes (representing three proteins), labelled as A, B and C. The edges in this example are undirected and unweighted.

[0165] Figure 6 shows an example of a PPI network that has been modified in the manner used in the previously described DependANT algorithm wherein each pair of proteins is associated with a pair of unidirectional edges, with each edge being weighted based on the expression of the gene encoded by the protein (i.e. node) that the edge extends from. In this example, the edges are therefore directed, but are 100% reciprocal, and so the directionality of the edges therefore simply serves to associate the edge weights with one or other of the nodes (but does not reflect the underlying directionality of the reaction pathway).

[0166] Figure 7 then shows an example of a PPI network that has been modified in the manner of the technology described herein, wherein each edge has an associated directionality representing the reaction pathway within the PPI network, and each edge is further associated with a single edge weight representing the respective expression level of the gene encoded by the protein (i.e. node) that the edge extends from. Other embodiments are however contemplated.

[0167] For example, in some embodiments, each edge could be associated with a pair of edge weights representing the respective expression levels of the genes encoded by the proteins (i.e. nodes) that the edge extends between.

[0168] Correspondingly, the topological features used to characterise the PPI networks will reflect this directionality.

[0169] For example, the following topological features explicitly use edge direction when values are calculated:

[0170] - Closeness

[0171] - Betweenness- Eigenvector centrality

[0172] - PageRank

[0173] - Hub score

[0174] - Degree

[0175] - Closeness / harmonic centrality

[0176] - Coreness

[0177] - Eccentricity

[0178] - Neighbourhood size

[0179] Thus, the values calculated for these topological features will differ significantly for a directed graph compared to those calculated in the DependANT algorithm described in Benstead-Hume et al. ‘Biological network topology features predict gene dependencies in cancer cell-lines’, Bioinformatics Advances, 2022, 1-8, in which the edges were directed but 100% reciprocal, such that the PPI networks did not take into account the directionality of reaction pathways.

[0180] In further embodiments, it is also contemplated that some or all of the topological features are calculated in a manner that directly takes into account the edge directionality. For example, at least the following topological features can additionally be defined in terms of explicit in / out options to differentiate between edges coming into or out of a given node, and in some embodiments, these additional in / out options are used when calculating the topological features:

[0181] - Degree

[0182] - Closeness / harmonic centrality

[0183] - Coreness

[0184] - Eccentricity

[0185] - Neighbourhood size.

[0186] The above thus describes the feature extraction that is performed during the training process shown in Figure 4. It will be appreciated that a similar process may be performed to process the user-uploaded data into a suitable format for analysis by the machine learning classifiers 106 (i.e. the processing in steps 200-201-202 in Figure 2). That is, once the user (e.g. oncologist) has uploaded a suitable set of mutation and expression data for a particular patient cancer tissue sample to be analysed (step 200 in Figure 2), the mutation and expression data for a particular patient cancer tissue sample is then used to generate a personalised PPI network for that particular patient cancer tissue sample, which in the present embodiments is done by perturbing one or more baseline PPI networks in a similar manner as discussed above. Appropriate feature extraction can then be performed on thepersonalised PPI network for the particular patient cancer tissue sample to extract a corresponding feature set including the topological characteristics in the same format that was used for training the machine learning classifiers 106, for analysis (i.e. in step 202 in Figure 2).

[0187] As will be shown further below, the inventors believe that incorporating the reaction pathway directionality into the PPI networks that are then used to train the machine learning models, e.g. the classifiers, can provide significant improvements in accuracy

[0188] The following section entitled ‘Methods’ provides further detailed explanations of testing that has been performed to support the development of the technology described herein according to its various aspects and embodiments.

[0189] Methods

[0190] Key packages and software

[0191] Pandas version 2.2.0, NumPy version 1.23.5 were used for the edge weights pipeline in Jupyter Notebook version 6.5.2 and PyCharm 2023.2.5 (Professional Edition) with Python 3.10.9.

[0192] Caret version 6.0-94, igraph version 1.5.1, pROC version 1.18.4, PRROC version 1.3.1 and dplyr version 1.1.3 were used for the data processing and analytical pipeline in RStudio version 1.1.456 with R version 4.3.1.

[0193] The previously described DependANT classifier was cloned from Bitbucket and installed as per the documentation.

[0194] Data Sources

[0195] Expression, mutation and CRISPR gene dependency data for 1017 cancer cell lines was downloaded from DepMap (Public 18Q3, Public 23Q4) (A. Tsherniak et al., “Defining a Cancer Dependency Map,” Cell, vol. 170, no. 3, pp. 564-576. e16, Jul. 2017). Lists of oncogene and tumour suppressor genes were obtained from the Cancer Gene Census database (Z. Sondka et al., “COSMIC: a curated database of somatic variants and clinical data for cancer,” Nucleic Acids Research, vol. 52, no. D1, pp. D1210-D1217, Jan. 2024).

[0196] Undirected human PPI data was downloaded from STRING (D. Szklarczyk et al., “The STRING database in 2023: protein-protein association networks and functional enrichment analyses for any sequenced genome of interest,” Nucleic Acids Res, vol. 51, no. D1, pp. D638-D646, Nov. 2022) (detailed version 10, fullversion 12). Directed PPI data from the Reactome database (M. Milacic et al., “The Reactome Pathway Knowledgebase 2024,” Nucleic Acids Research, vol. 52, no. D1, pp. D672-D678, Jan. 2024) was downloaded from Pathway Commons (version 12, HGNC SIF format) (I. Rodchenkov et al., “Pathway Commons 2019 Update: integration, analysis and exploration of pathway data,” Nucleic Acids Research, vol.

[0197] 48, no. D1, pp. D489-D497, Jan. 2020).

[0198] Replication of original DependANT results

[0199] To confirm successful installation and establish baseline results, the previously described DependANT algorithm was run with the original source data (DepMap Public 18Q3, STRING v10) and selection of 39 cell lines (19 breast, 11 kidney and 9 pancreas). Genes with DepMap dependency scores > 0.65 were considered essential in a given cell line. Around 500 genes were essential in all 39 cell lines (pan-essential genes).

[0200] STRING v10 PPI data was filtered to remove interactions with an experimental score < 80. The base PPI network comprised 7,262 nodes and -60,000 edges (Table 2).

[0201]

[0202] Table 2. Nodes and edges in each of the base PPI networks used.

[0203] To map gene dependency scores to PPI network nodes, Ensembl protein identifiers in the STRING dataset were converted to Ensembl gene identifiers (see, F. J. Martin et al., “Ensembl 2023,” Nucleic Acids Research, vol. 51, no. D1, pp. D933-D941, Jan. 2023).

[0204] A base PPI model was created for each cell line using the igraph package in RStudio and twelve topological features calculated for each node (as shown in Table 1, above). Genes with dependency scores > 0.65 were labelled as dependent and genes below this threshold, non-dependent. Feature data was partitioned into training and testing datasets in an 80:20 ratio. Under-sampling was used to balance both sets with an equal number of dependent / non-dependent nodes.An Adaboost classifier was trained for each cell line. Hyperparameters were tuned using the default Caret settings. Models were optimised on Area Under the Curve (AUC) of the Receiver Operating Characteristic (ROC) and performance was estimated with 10-fold cross validation. Each model was validated against the unseen holdout test data for an unbiased performance evaluation.

[0205] To determine whether classifiers trained on individual cell lines could predict dependencies in other cell lines and tissue types, a second round of validation was performed. Each model was validated again on test data from each of the other cell lines. To prevent bias, genes were removed from the active test set if they were also present in the model’s training data.

[0206] To establish how well models classify pan-essential genes and less common gene dependencies, cross-cell-line validation was repeated. This time, holdout test sets were manipulated to include gene dependencies found in a specified number of cell lines, ranging from those found in one cell line to those present in all 39 cell lines.

[0207] Perturbed PPI network models

[0208] Classifiers were retrained on PPI network models which had been customised to reflect the unique mutation and expression profile of each cell line (the DependANT classifier). As described above in relation to Figure 6 and Figure 7, edge weights were calculated as a function of gene expression and scaled between 0 and 1 with:

[0209] w = 0.5 - 0.5 tanh (log(g + 1e - 10)),

[0210] so that edge weight decreases towards zero with increasing gene expression and increases towards one with decreasing expression. Genes with frameshift, nonsense and non-stop mutations were categorized as loss of function (LOF) and excluded from the network. Missense mutations were evaluated with the FATHMM algorithm (see H. A. Shihab et al., “Predicting the functional, molecular, and phenotypic consequences of amino acid substitutions using hidden Markov models,” Hum Mutat, vol. 34, no. 1, pp. 57-65, Jan. 2013).

[0211] Damaging mutations were labelled as gain-of-function (GOF) if they occurred in oncogenes and LOF if they occurred in tumour suppressor genes. LOF-labelled genes were removed from the network for the respective cell line. It was assumed that genes with GOF mutations exert greater influence within the PPInetwork. To model this, expression values for GOF-labelled genes were multiplied by ten before edge weights were calculated.

[0212] Care was taken to only remove LOF-labelled genes (and not to remove nodes that are not categorized as LOF but have zero expression). That is, in embodiments, unexpressed nodes remain within the PPI network.

[0213] Pan-cancer model

[0214] Training data from all cell lines was combined into a single pan-cancer training set and feature values normalized using the Caret package scale and center functions. An Adaboost classifier was trained on the pan-cancer data and validated against the holdout test sets of each individual cell line. This was repeated with filtered test sets containing genes that are essential in a specified number of cell lines.

[0215] Feature importance analysis

[0216] The class distribution of training feature values was plotted and the Caret package varlmp function was utilised to rank the importance of training features.

[0217] Data update and integration

[0218] Data-processing pipelines were rebuilt to accommodate directed Reactome PPI data and ensure compatibility with the latest versions of DepMap and STRING, as detailed below.

[0219] Updated identifiers- DepMap has transitioned from using Cancer Cell Line Encyclopedia (CCLE) cell line names and Ensembl gene IDs to Broad cell line IDs and HGNC gene symbols. To align with this, STRING PPI nodes were converted from Ensembl protein IDs to HGNC symbols using Ensembl Biomart data. As CCLE names include the tissue of origin, Broad cell line IDs in DepMap 23Q4 were converted back to CCLE names to facilitate tissue-specific analyses.

[0220] Genome assembly alignment- Gene variants in DepMap 23Q4 are aligned to the hg38 genome assembly whereas FATHMM requires hg37 aligned sequences. AlphaMissense is an alternative mutation assessment tool that utilises deep learning to predict the pathogenicity of missense mutations. AlphaMissense predictions covering allpotential single amino acid variations across 19,233 canonical human proteins and 60,000 alternative transcripts have been published, and all are aligned to hg38. To circumvent the need to map variants between assemblies, the AlphaMissense prediction datasets were integrated with the weights pipeline in place of FATHMM.

[0221] Compatibility update for directed PPI data- Feature-processing scripts that were originally coded for bi-directional STRING data were updated for compatibility with directed PPI data from the Reactome database.

[0222] Performance metrics- The output of performance metrics was increased to include not only ROC AUC, sensitivity but also specificity metrics precision, negative predictive value, and F1 score (See Table 3 for definitions).

[0223]

[0224]

[0225] Table 3. Performance metrics used for model evaluation.

[0226] Pipeline optimisation- The Python-based weights pipeline was optimised to provide a clearer view of data transformations at each stage. Scripts within the primary R-based analytical pipelines were optimised for parallel execution to reduce processing times.

[0227] Updating the STRING PPI network

[0228] PPI databases often have a bias towards well-studied proteins, resulting in some areas being densely populated while others are relatively sparse. To mitigate this imbalance a modified STRING (v12) dataset was created, aiming to maximize node coverage while also reducing the edge count on over-represented nodes (defined as nodes with > 1000 edges). STRING (v12) data was filtered to include interactions with a human experimental evidence score > 80.

[0229] This step ensured no node exceeded 1000 edges, but many nodes were eliminated. The STRING data was filtered again to reselect nodes with a lower experimental score of > 60 if they had previously been excluded by the stricterfiltering criteria. Combining these filtered datasets resulted in a base PPI network with 10,009 nodes and 59,569 edges (Table 2).

[0230] The whole pipeline was rerun with the refreshed STRING and DepMap datasets. Two of the original 39 cell lines were absent from DepMap 23Q4 so the analysis continued with the remaining 37 cell lines (19 breast, 9 kidney, and 9 pancreas cell lines).

[0231] Integrating directed Reactome PPI data

[0232] As described above in relation to Figure 7, the PPI networks that are used for training the machine learning classifiers in the present embodiments incorporate directional data indicative of reaction pathways within the PPI network.

[0233] Directed PPI network data was downloaded in SIF format from the Reactome database via Pathway Commons. Non-protein interactions from the Chemical Entities of Biological Interest (ChEBI) database were removed.

[0234] Personalised PPI networks were generated for the 37 cell lines and network topology features calculated for each node. Adaboost classifiers were trained on features for each individual cell line and a combined pan-cancer training set, as previously described.

[0235] Reducing noise in the dependency data

[0236] The distribution of gene dependencies described in Benstead-Hume et al. ‘Biological network topology features predict gene dependencies in cancer celllines’, Bioinformatics Advances, 2022, 1-8 suggests that -800 genes were essential in just one cell line. CRISPR-Cas9 knockout screens are known to produce false positives in highly amplified regions and the large number of isolated dependencies may reflect this. Removing this noise from the training data may improve the accuracy of predictions. To achieve this, an additional data cleaning step may be applied to the pre-processing pipeline to remove genes from the dependency data if they were essential in < 5% of DepMap samples.

[0237] Inverting selected edge weights

[0238] The method for calculating edge weights produces high weights for genes with low expression and vice versa. This is logical for features such as closeness and betweenness which interpret edge weight as a distance, as shorter distances between nodes implies a closer relationship. Other topological features such as page rank, eigen centrality, and hub score interpret edge weight as the strength ofconnection between nodes. The weight calculation is counterintuitive for these features, as it suggests a gene’s impact increases as expression diminishes. Where features interpret edge weight as a strength rather than a cost, inverting the weight may offer a truer depiction of the PPI model and lead to stronger predictive performance. Thus, in embodiments, weights may be inverted (using 1 -weight) for Constraint, Hub score, Eigen centrality and Page rank calculations.

[0239] Novel feature testing

[0240] Examination of the feature sets across all models revealed that some topological features were not always successfully calculated. In particular, a large number of genes had invalid NaN values for closeness across all models. Genes with invalid features values were filtered from the feature sets reducing the amount the training data available. The igraph documentation states that closeness can only be calculated for connected graphs and recommends harmonic centrality as an alternative for disconnected graphs. As a PPI network is a disconnected graph consisting of multiple components, Closeness may be replaced with Harmonic centrality.

[0241] Gene essentiality is known to be associated with measures of node centrality in PPI networks. However, the centrality-lethality rule may only apply to panessential genes and demonstrated that measures of localised clustering such as transitivity (also known as the clustering co-efficient) are better predictors of contextessential genes. Transitivity may therefore be incorporated into the training data as an additional feature.

[0242] Increasing the sample size

[0243] The best performing model was retrained on a new sample of 100 cell lines from 7 cancers. The three original cancer types were retained, and four additional cancer types were selected at random from the DepMap model data. A python script was used to randomly select up to 15 cell lines from each of the seven cancers, up to a total of 100 cell lines. The new sample comprised 14 breast, 12 kidney, 14 pancreas, 13 colon, 14 liver, 13 ovary and 14 skin samples. As before, validation was performed on class-balanced holdout test sets. To evaluate model performance on unbalanced real-world data, validation was repeated on an independent selection of 35 cell lines from the same seven cancer types, in which the natural balance of dependent and non-dependent genes remained intact.Results

[0244] Replication of original findings

[0245] The performance of the original DependANT algorithm as described in Benstead-Hume et al. ‘Biological network topology features predict gene dependencies in cancer cell-lines’, Bioinformatics Advances, 2022, 1-8 was successfully replicated. The baseline PPI model classified cell-line-specific dependencies with an average ROC AUC of 0.761 (SD 0.012) compared to 0.758 in the original paper (Table 4). The DependANT classifier trained on individual cell-line data achieved an average ROC AUC of 0.801 (SD 0.016), a performance improvement of 5.3% over the baseline model.

[0246]

[0247] Table 4. Average ROC AUC scores (and standard deviation) for baseline and DependANT models. Single cell line models were validated on test data from the same cell line and all other cell lines. The pan-cancer trained DependANT model was tested on data from all cell lines.

[0248] Cross-cell line classification

[0249] DependANT was trained on individual cell lines and used to predict dependent genes in other cell lines of the same and different tissue types. As per the original study, performance dropped slightly to an average ROC AUC of between 0.788 and 0.800 when classifying dependencies in cell lines of the same tissue type, and between 0.786 and 0.795 in different tissue types (Table 5).

[0250]

[0251]

[0252] Table 5. Replication of the original study findings. Average ROC AUC results (and standard deviation) for individual cell line models tested on validation data from the same and different tissue types.

[0253] Pan-tissue DependANT classifier

[0254] The DependANT model trained on concatenated pan-cancer training data classified dependencies with an average ROC AUC of 0.838 (SD 0.013) across all cell lines, a performance improvement of 5.7 percent compared to models trained on single cell line data (Table 4, Figure 8).

[0255] Figure 8 shows the ROC AUC score distributions for baseline, DependANT and DependANT pan-cancer classifiers trained using STRING v10 PPI data, i.e. replicating the results described in Benstead-Hume et al. ‘Biological network topology features predict gene dependencies in cancer cell-lines’, Bioinformatics Advances, 2022, 1-8.

[0256] Dependencies subset by frequency

[0257] Across all of the above models (baseline, DependANT, and DependANT trained on pan-cancer data), classification performance was lowest for genes that are essential in only one cell line, and highest for genes that are essential in all 39 cell lines (pan-essential genes) (see Figure 9). DependANT consistently outperformed the baseline model by up to 10%. DependANT trained on pan-cancer data provided an additional performance improvement of between 2% and 5% across all rarity intervals.

[0258] Figure 9 thus shows the average model performance (ROC AUC) on test sets filtered by dependency frequency when replicating the results described in Benstead-Hume et al. ‘Biological network topology features predict gene dependencies in cancer cell-lines’, Bioinformatics Advances, 2022, 1-8. The bars show the frequency of gene dependencies in the 39 cell-lines (e.g. -800 genes were essential in a one cell line and -500 were essential in all 39 cell lines). Lines showaverage ROC AUC scores for classification of genes that are reported to be essential in 1, 10, 20, 30 and 39 cell lines.

[0259] Feature importance and distribution

[0260] Feature importance and distribution were as reported in the original paper. Constraint, Coreness, Eigen centrality and Hub score showed distribution differences between dependent and non-dependent classes indicating potential predictive power. The trained pan-cancer model ranked Degree, Coreness, Neighborhood 1 size and Page rank as the most important.

[0261] Figure 10 shows the distributions of topological network features for dependent and non-dependent gene classes in the original STRING v10 pan-cancer training dataset. Features which show a different distribution for dependent and non-dependent classes may have predictive power.

[0262] Figure 11 then shows a heatmap of feature importance scores for 4 pancancer DependANT model iterations built on different base PPI networks, including models according to the present embodiments that include directional pathway data based on the Reactome dataset.

[0263] Updating the STRING PPI network

[0264] The curated STRING v12 PPI dataset contained 10,008 nodes and 56,487 edges. No node had more than 1000 edges.

[0265] The DepMap 23Q4 release included expression, mutation, and dependency data for 1013 cell lines. 619 genes had dependency scores > 0.65 in all 37 cell lines and were considered pan essential.

[0266] Classifier performance

[0267] The STRING v12 model achieved improved performance in all measures compared to the STRING v10 version (Table 6, Figure 12). The baseline model classified within-cell-line dependencies with an average AUC ROC of 0.824 (SD 0.018) - comparable to the performance of the pan-tissue classifier in the published study. The average AUC ROC for DependANT models trained on individual cell lines was 0.853 (SD 0.012). Cross cell-line dependency classification was also improved when using a classifier trained on one cell line to predict dependencies in another (average ROC AUC 0.820 SD 0.010). The pan-cancer trained DependANT model achieved an average ROC AUC of 0.874 (SD 0.011) across all cell lines (Figure 12).

[0268]

[0269] Table 6. Average ROC AUC scores (and standard deviation) for baseline and DependANT models trained on STRING v10, STRING v12 and Reactome PPI networks. Single cell line models were validated on test data from the same cell line and all other cell lines. The pan-cancer trained DependANT model was tested on data from all cell lines.

[0270] Figure 12 shows a comparison of performance metrics for DependANT pancancer models based on STRING v10, STRING v12 and for models that are trained according to the present embodiments based on Reactome PPI data. In the legend, auc = receiving operator characteristic area under the curve, sens = sensitivity, spec = specificity, prec = precision, f1 = the harmonic mean of specificity and sensitivity, npv = negative predictive value. See Table 3 for definitions.

[0271] Subset analysis

[0272] As with the original model, classification performance was weakest when predicting dependencies that are present in a single cell line and strongest when classifying pan-essential genes. Figure 13 thus shows the average model performance (ROC AUC) on test sets filtered by dependency frequency. Results from the original study (‘Baseline model., ‘DependANT model’, ‘DependANT pancancer STRING v10’) are shown alongside results from the new STRING v12 and Reactome-based model iterations. Histogram bars depict the frequency of genedependencies in the DepMap 18Q3 data release.

[0273] Feature importance and distribution

[0274] Feature distribution for the STRING v12 dataset was similar to the STRING v10 feature distribution. Constraint, Coreness, Degree and Neighborhood 1 size all showed differences between dependent and non-dependent classes. Figure 14 thus shows the class distribution of values for topological network features in the STRING v12 pan-cancer training dataset. The trained pan-cancer model ranked Degree, Coreness and Neighborhood 1 size as the most important features (as shown in Figure 11, above).

[0275] Integrating directed PPI network data from Reactome

[0276] The Reactome PPI data downloaded from Pathway Commons contained 10,015 nodes and 300,720 edges.

[0277] Classifiers based on directed PPI data from the Reactome database achieved a further increase in performance compared to both the original model and the updated STRING v12 version (see Figure 12, Table 7 above).

[0278] Individual cell line models classified dependencies with an average AUC ROC of 0.885 (SD 0.018) within cell lines, an increase of around 10% on the original model and 5% on the STRING v12 model.

[0279] The pan-cancer Reactome model achieved an average ROC AUC of 0.900 (SD 0.015) (Figure 12, Table 6) and showed consistently higher performance at all rarity intervals compared to the STRING based models (Figure 13).

[0280] Feature importance and distribution

[0281] Feature distribution for the Reactome model did not differ significantly from the STRING-based feature sets. Constraint, Coreness, Degree and Neighborhood 1 size all showed differences between dependent and non-dependent classes. Figure 15 thus shows the distribution of values for topological network features in the Reactome pan-cancer training dataset. The most important features as determined by the trained pan-cancer model, were Closeness, Coreness, Neighborhood 1 size and Degree (as shown in Figure 11, above).Data pre-processing and feature modifications

[0282] Thus, various further data pre-processing and feature modifications are contemplated, as discussed above, and the impact of each of the following modifications was tested for the pan-cancer Reactome model:

[0283] 1. To reduce noise in the dependency data, genes that were almost never essential (dependency score > 0.65 in less than 5% of DepMap samples) were excluded when labelling the training data.

[0284] 2. Unexpressed nodes, which had previously been removed along with LOF nodes, were kept in the PPI network.

[0285] 3. Edge weights were inverted for topological features that interpret it as a strength rather than a distance.

[0286] 4. The Closeness feature was replaced with Harmonic centrality.

[0287] 5. Transitivity was trialled as an additional topological feature.

[0288] 6. Transitivity and Harmonic centrality were tested together.

[0289] Models were evaluated with 10-fold cross validation and optimised for ROC AUC. As the differences in ROC AUC values were relatively small, scores for each model iteration were compared to the standard Reactome model cross validation scores with DeLong’s test, using the pROC package. Benjamin Hochberg correction was applied to p-values with an FDR of 0.05 and p <= 0.05 was considered statistically significant.

[0290] Figure 16 shows the average model performance (ROC AUC) of standard and optimized Reactome models on test sets filtered by dependency frequency.

[0291] Figure 17 further shows test set validation performance metrics for iterations of the Reactome-based pan-cancer model with various data pre-processing and feature adjustments. Wl = weights inverted for selected topological features, DC = DepMap data cleaned, Standard = standard Reactome model, TR = addition of transitivity feature, Zl = unexpressed nodes kept in the network, HC = addition of harmonic centrality feature, TR+HC = addition of transitivity and harmonic centrality features. In the legend: auc = receiving operator characteristic area under the curve, sens = sensitivity, spec = specificity, prec = precision, f1 = the harmonic mean of specificity and sensitivity, npv = negative predictive value.

[0292] Figure 18 is an ROC curve for optimized Reactome model classifying panessential genes.

[0293] <

[0294] <

[0295]

[0296] Table 7. Average cross-validation ROC AUC scores and adjusted DeLong’s test p-values for models with data pre-processing and feature modifications.

[0297] DeLong’s test is a test of difference to compare the ROC curves between two conditions. Modified models were compared to the standard Reactome-based pancancer DependANT model.

[0298] Inverting edge weights for selected features and cleaning the DepMap data did not significantly alter the average cross-validation AUC ROC score compared to the standard Reactome model (Table 7, Figure 17).

[0299] On the other hand, adding Transitivty to the feature set produced a small but statistically significant increase from 0.897 to 0.900 (p=0.014).

[0300] Keeping unexpressed nodes in the network and replacing Closeness with Harmonic centrality increased the average ROC AUC score to 0.908 (p=1.36 x 10'12) and 0.914 (p < 2.2 x 10'15) respectively.

[0301] Combining Transitivity and Harmonic centrality in the feature set increased the average ROC AUC to 0.916 (p < 2.2 x 10’15).

[0302] The small increase in performace compared to the model with Harmonic centrality alone was also statistically significant (p = 2.2 x 10'4). This model also achieved the highest average scores for sensitivity, specifcity, precision, F1 score and negative predictive value (NPV) (Figure 17).

[0303] The optimised Reactome model consistencly performed around 5% higher than the standard Reactome model when evaluated on test sets filtered bydependency frequency (Figure 16). The optimised model predicted pan-essential genes with an AUC ROC of 0.947 (Figure 18).

[0304] Feature importance and distibution

[0305] Feature distribution was largely the same for the the standard and optimised Reactome models, with the exception of the newly introduced features which showed marked differences in class distribution (see Figure 19 showing the distribution of topological network features in the optimised Reactome pan-cancer training dataset)). The optimsed model ranked Harmonic centralitry as the most important feature, followed by Coreness, Neighbourhood 1 size, Degree and Transitivity (Figure 11).

[0306] To assess performance on unbalanced real-world data, the model was validated again with unbalanced data from 35 previsouly unseen cell lines.

[0307] Figure 20 shows the distribution of performance metrics for DependANT pan-cancer model iterations. STRING v10 is the original model. STRING v12 and Reactome are iterations based on different baseline PPI networks. Reactome optimised is the Reactome model with additional feature and data pre-processing optimisations (see method for details). The optimised Reactome model was subsequently trained in an independent set of 100 cell lines from 7 different cancers.

[0308] Figure 21 shows the distribution of validation metrics for the Reactome pancancer DependANT model trained on 100 cell lines.

[0309] Balanced validation set = results of validation with the class-balanced holdout set.

[0310] Unbalanced test set = results from tests on unbalanced data from an independent selection of 35 new cell lines.

[0311] Balanced test set = results from the same test set of 35 cell lines after classbalancing the data.

[0312] ROC AUC scores were consistent with those obtained from cross validation and holdout set tests, but sensitivity dropped signifciantly to an average of 0.420 (SD 0.049), resulting in a corresponding uplift in speciificity and drop in F1 score. To test whether this change in performnace was due to the test set class imbalance, undersampling was used to balance the new test set of 35 cell lines with an equal number of depdendent and and non-dependent nodes. When retested on balanced test data, all performance metrics recovered and were in line with the holdout validation results (Figure 21).The results published in Benstead-Hume et al. ‘Biological network topology features predict gene dependencies in cancer cell-lines’, Bioinformatics Advances, 2022, 1-8 were thus successfully replicated and demonstrate that the DependANT algorithm provides enhanced predictive ability compared to models built on baseline PPI networks. Performance gains were largely maintained when DependANT was retrained on a more diverse set of 100 cell lines, and independent validation on a separate set of 35 cell lines confirms its ability to predict gene dependencies in previously unseen samples.

[0313] The present inventors have now achieved further performance improvements by refreshing the source data and revising the base PPI models, in particular to incorporate directional reaction pathway data (e.g. based on the Reactome database), as shown above. This is consistent with the inventors’ hypothesis that capturing directionality in the PPI network would boost performance.

[0314] Performance was measured on subsets defined by the rarity of gene dependencies. All model iterations performed best when predicting pan-essential genes, and this may reflect the focus on traditional measures of centrality in the training data. Adding Transitivity as a training feature marginally improved performance across all rarity intervals which may be due to better detection of context-essential genes.

[0315] A performance uplift was also achieved by replacing Closeness with Harmonic centrality underlining the importance of selecting meaningful training features. Closeness could not be calculated for many nodes because the PPI network is disconnected, and the calculation of eigen centrality may also be affected by this issue. One solution is to unify the network by connecting individual components using additional PPI data sources. Alternatively, smaller disconnected subnetworks could be pruned to leave only the largest connected network component, but this is likely to result in a significant loss of nodes. Additional topological features that capture the nature of inbound and outbound connections in directed graphs could also be explored.

[0316] Thus, the foregoing detailed description has been presented for the purposes of illustration and description. It is not intended to be exhaustive or to limit the technology described herein to the precise form disclosed. Many modifications and variations are possible in the light of the above teaching. The described embodiments were chosen in order to best explain the principles of the technology described herein and its practical applications, to thereby enable others skilled in the art to best utilise the technology described herein, in various embodiments and withvarious modifications as are suited to the particular use contemplated. It is intended that the scope be defined by the claims appended hereto.

Claims

1. 165131 - Pearl A. I. - DependANT+Claims;1. A method of training one or more machine learning model to predict, based on genomic data and gene expression data associated with a particular patient cancer tissue sample, a set of one or more essential genes for that particular patient cancer tissue sample, the method comprising:obtaining, for training data, a set of plural feature sets, each feature set representing one or more topological characteristics of a respective protein- protein interaction, “PPI”, network from which the feature set has been derived, wherein respective PPI networks includes a respective plurality of nodes and a plurality of edges, each edge connecting a respective pair of nodes within the PPI network, wherein respective nodes represent respective proteins and respective edges represent interactions between the respective proteins represented by the nodes that are connected by that edge, and wherein respective edges within the PPI network have an associated directionality indicating a reaction pathway associated with the interaction between the respective proteins represented by the nodes that are connected by that edge,wherein the nodes within different PPI networks are configured to reflect the mutational status of the genes encoding the proteins represented by those nodes, different PPI networks thereby representing models of different cancers, and each feature set being labelled with information indicative of which nodes within the corresponding PPI network from which that feature set has been derived represent essential genes for the cancer represented by that PPI network; and the method further comprising:training the one or more machine learning model using feature sets that have been derived from a plurality of different PPI networks to be able to predict respective sets of essential genes for patient cancer tissue samples.

2. A method of training one or more machine learning model to predict, based on genomic data and gene expression data associated with a particular patient cancer tissue sample, a respective primary site of origin for that particular patient cancer tissue sample, the method comprising:obtaining, for training data, a set of plural feature sets, each feature set representing one or more topological characteristics of a respective protein- protein interaction, “PPI”, network from which the feature set has been derived, wherein respective PPI networks include a respective plurality of nodes and a plurality of edges, each edge connecting a respective pair of nodes within the PPI network, wherein respective nodes represent respective proteins and respective edges represent interactions between the respective proteins represented by the nodes that are connected by that edge, and wherein respective edges within the PPI network have an associated directionality indicating a reaction pathway associated with the interaction between the respective proteins represented by the nodes that are connected by that edge,wherein the nodes within different PPI networks are configured to reflect the mutational status of the genes encoding the proteins represented by those nodes, different PPI networks thereby representing models of different cancers, and each feature set being labelled with information indicative of a primary site of origin for the cancer cell line represented by the PPI network from which the feature set has been derived; andthe method further comprising:training the one or more machine learning model using feature sets that have been derived from a plurality of different PPI networks to be able to predict respective primary sites of origin for patient cancer tissue samples.

3. A method of training one or more machine learning model to predict, based on genomic data and gene expression data associated with a particular patient cancer tissue sample, one or more characteristics for that particular patient cancer tissue sample, the method comprising:obtaining, for training data, a set of plural feature sets, each feature set representing one or more topological characteristics of a respective protein- protein interaction, “PPI”, network from which the feature set has been derived, wherein respective PPI networks includes a respective plurality of nodes and a plurality of edges, each edge connecting a respective pair of nodes within the PPI network, wherein respective nodes represent respective proteins and respective edges represent interactions between the respective proteins represented by the nodes that are connected by that edge, and wherein respective edges within the PPI network have an associated directionality indicating a reaction pathway associatedwith the interaction between the respective proteins represented by the nodes that are connected by that edge,wherein the nodes within different PPI networks are configured to reflect the mutational status of the genes encoding the proteins represented by those nodes, different PPI networks thereby representing models of different cancers, and each feature set being labelled with information indicative of the one or more characteristics for the corresponding PPI network from which that feature set has been derived; andthe method further comprising:training the one or more machine learning model using feature sets that have been derived from a plurality of different PPI networks to be able to predict the one or more characteristics for patient cancer tissue samples.

4. The method of claim 1, 2 or 3, wherein respective edges are further associated with an edge weight that is indicative of the expression status of the gene encoding the protein represented by the node from which that respective edge extends.

5. The method of claim 4, wherein the edge weights are calculated as a function of gene expression so that the edge weight decreases towards zero with increasing gene expression.

6. The method of any preceding claim, wherein the one or more topological characteristics of the PPI network that are included into the feature sets comprise one or more indicators of node centrality.

7. The method of claim 6, wherein the topological characteristics comprise one or more of: Transitivity; Betweenness; Constraint; Coreness; Closeness; Harmonic centrality; Degree; Eccentricity; Eigen centrality; Hub score; Neighbourhood n size (for n = 1,2,6); and PageRank.

8. The method of any preceding claim, wherein the one or more machine learning model comprises one or more graph neural network(s).

9. The method of any preceding claim, wherein the one or more machine learning model is a machine learning classifier model or a machine learning regression model.

10. The method of any preceding claim, wherein obtaining the set of plural feature sets comprises:obtaining a set of plural PPI networks, wherein respective PPI networks within the set of plural PPI networks represent models of different respective cancers; andextracting for respective PPI networks within the obtained set of plural PPI networks respective feature sets representing the one or more topological characteristics of the PPI network.

11. A machine learning model that has been trained by the method of any of claims 1 to 10.

12. A method of identifying one or more characteristics of a particular patient cancer tissue sample, the method comprising:obtaining a respective feature set representing one or more topological characteristics of a respective protein-protein interaction (PPI) network model that has been derived from a set of genomic data and gene expression data for the particular patient cancer tissue sample; andproviding the obtained feature set for the particular patient cancer tissue sample to one or more machine learning model, wherein the one or more machine learning model has been trained by the method of any of claims 1 to 10, or comprises the machine learning model of claim 11.

13. The method of claim 12, wherein:when dependent on claim 1, the set of one or more characteristics includes a set of one or more essential genes for that particular patient cancer tissue sample; andwhen dependent on claim 2, the set of one or more characteristics includes a respective type of cancer for that particular patient cancer tissue sample.

14. The method of claim 12 or 13, wherein obtaining the respective feature set for the particular patient cancer tissue sample to be analysed comprises:obtaining a set of genomic data and gene expression data for the particular patient cancer tissue sample; andprocessing the obtained set of genomic data and gene expression data to generate the respective feature set.

15. The method of any of claims 12, 13 or 14, wherein the set of one or more characteristics includes a set of one or more essential genes for that particular patient cancer tissue sample, and wherein the method comprises:determining, based on the set of one or more essential genes that are identified for the particular patient cancer tissue sample identified by the machine learning model, a set of other information including a list of known drugs or biomarkers for targeting one or more of the identified essential genes; and outputting the set of other information including the list of known drugs or biomarkers for targeting one or more of the identified essential genes together with a list of the identified essential genes.

16. A system for identifying a set of one or more characteristics of particular patient cancer tissue samples, the system comprising:a local interface via which a user can input data associated with a particular patient cancer tissue sample to be analysed; andat least one processor that is operable and configured to execute one or more machine learning model, wherein the one or more machine learning model comprises a machine learning model that has been trained according to the method of any of claims 1 to 10, or comprises the machine learning model of claim 11 , and wherein the one or more machine learning model is configured to identify, from an input set of data associated with a particular patient cancer tissue sample or cancer cell line, one or more characteristics of the particular patient cancer tissue sample.

17. The system of claim 16, wherein:when the machine learning model has been trained according to the method of claim 1, the one or more characteristics include a set of one or more essential genes for that particular patient cancer tissue sample; andwhen the machine learning model has been trained according to the method of claim 2, the one or more characteristics a respective primary site of origin for that particular patient cancer tissue sample.

18. The system of claim 16 or 17, wherein when the set of one or more characteristics includes a set of one or more essential genes for a particular patient cancer tissue sample, the at least one processor is further operable and configured to:determine, based on the set of one or more essential genes that are identified for the particular patient cancer tissue sample identified by the machine learning model, a set of other information including a list of known drugs or biomarkers for targeting one or more of the identified essential genes; and output the set of other information including the list of known drugs or biomarkers for targeting one or more of the identified essential genes together with a list of the identified essential genes.

19. A computer program product comprising instructions that when executed by one or more processor performs a method of training a machine learning model as claimed in any of claims 1 to 10 or a method of identifying one or more characteristics of a particular patient cancer tissue sample as claimed in any of claims 12 to 15.

Images