The foundational step in drug discovery hinges upon the precise delineation of therapeutic targets. The multifaceted nature of this endeavor necessitates adherence to several imperative criteria. Primarily, targets should exhibit a degree of specificity that highlights disparities between tumor and normal tissues. Furthermore, their involvement in tumor cell survival pathways is paramount. Equally significant is their amenability to small molecule-based interventions. Notably, our research advances the proposition that altered genes, underpinning increased disease aggressiveness and decreased overall survival, hold promise as prime targets. This aligns with the overarching objective of machine learning-driven target identification—an effective prediction of overall survival and relapse-free interval through the discerning selection of indispensable genes while simultaneously accounting for distinctive expression profiles in comparison to normal tissues.

A nuanced challenge arises from genes implicated in pivotal cellular processes that are shared between tumor and normal cells. The intricate balance of targeting such genes mandates a judicious approach. To this end, we have incorporated a refined strategy. Leveraging comprehensive gene expression data from tumor and normal tissues, we discriminate against genes displaying marginal expression differences. Employing deep learning algorithms, we delineate genes that decisively demarcate tumor and normal tissues, thus augmenting the precision of target gene selection.

Beyond the confines of experimental data, deep learning extends its scope to the vast expanse of the published literature. Our research capitalizes on this potential, streamlining decision-making processes by assimilating insights from a meticulously curated repository of scientific articles. A dedicated tabulated summary of findings from a PubMed and PMC search fortifies the arsenal of tools for selecting promising molecular inhibitors.

A pivotal axis of drug discovery revolves around predicting the interaction dynamics between drug molecules and target proteins. We introduce an innovative deep learning model, integrating intricate protein and drug molecule information. This model prognosticates the impact of drug molecules on target proteins, ushering in a refined selection process. This stage inherently winnows the gamut of potential targets, eliminating candidates whose inhibition feasibility or binding efficacy raises concerns.

Transitioning toward preclinical studies demands the emulation of cellular experiments, a crucial precursor to laboratory validation. Deliberations encompassing cell lines that accurately mirror real-world conditions are pivotal. Our investigation extends to prognosticating the likelihood of compounds from prior stages attaining IC₅₀ concentrations within select cell lines, paving a trajectory toward informed preclinical trial design.

In the synthesis of these insights, our study embarks on a journey through the intricate tapestry of machine learning-driven drug discovery in oncology. From the meticulous identification of target genes to the finesse of molecular interaction prediction and the refinement of preclinical trial design, the amalgamation of cutting-edge techniques and comprehensive data analyses presents a formidable paradigm shift in the pursuit of effective therapeutics.

To assess the effect of gene expression on disease prognosis, we used the data on gene expression from the open database TCGA https://gdc.cancer.gov (Cancer Genome Atlas Research Network et al., 2013). In this section, gene expression data (RNA-Seq) were acquired and subsequently subjected to a normalization procedure. Normalization was performed by aligning the expression values to the reference levels represented by the control gene GAPDH, commonly referred to as a “housekeeping” gene. This process was undertaken to facilitate the integration of newly acquired data into the database. The data on overall survival (OS), progression-free interval (PFI), and the same parameters within the follow-up period were derived from this database. The problem of OS prediction was successfully solved in our previous study (Chebanov et al., 2021).

The essence of applying machine learning in this context can be summarized as follows: a dataset is prepared comprising features that include cancer-associated genes and patient medical history data. The target variable is a binary outcome representing whether the patient surpasses the median PFI or OS value for the whole dataset. A model is constructed using a multi-layer perceptron within the Python environment utilizing the Keras library. Upon achieving satisfactory training quality, the most influential features affecting the prediction are extracted from the dataset.

As an example, we selected the diagnosis of lung cancer and extracted a cohort of 514 patients with this diagnosis from the database.

In order to minimize training noise, we specifically curated a dataset containing genes that are integral to tumor-associated signaling pathways, as defined by the KEGG resource. This meticulous curation resulted in the inclusion of a total of 1,821 genes (Kanehisa and Goto, 2000).

To begin, we initially trained the algorithm utilizing a dataset comprising 514 patient records. However, following a rigorous five-fold cross-validation procedure, we observed that the mean ROC-AUC values were 0.69 (0.61–0.74) for overall survival (OS) and 0.61 (0.54–0.69) for progression-free interval (PFI). These results collectively signify a level of predictive performance that falls short of expectations, thus indicating the need for further refinement and enhancement of our predictive model. The reason is that it was a very small dataset for the application of neural networks, even taking into account more than 1,800 features.

To avoid this, we generated additional data comprising 50,000 synthetic patients by applying a generative adversarial network (GAN) to the tabular data on the existing 514 patients. GAN technology has been successfully used in various industries, such as image generation (Goodfellow et al., 2014). To use this methodology, we leveraged the Python SDV library (Patki et al., 2016) with a specific focus on utilizing the CTGAN module (Xu et al., 2019).

To refine the list of potential targets, we trained a deep learning algorithm to classify tissue into healthy and tumor categories based on gene expression. Subsequently, we ranked the features of the original dataset by importance, and genes with the greatest influence on the prediction were identified as more probable candidates for targeting as they contribute more significantly to distinguishing tumors from healthy tissue.

Gene expressions for tumor-normal data were taken from the GENT2 database (Park et al., 2019). This database comprises information on 68,287 samples of patients’ tissues and cell lines for all the diagnoses, of which 58,041 were tumor samples and 10,246 were normal samples.

The dataset for cell experiment emulation was formed using the data on gene expression profiles of the cell lines from the Cancer Cell Line Encyclopedia (CCLE) database and compound sensitivity data for cell lines from the Genomics of Drug Sensitivity in Cancer (GDSC) database (Barretina et al., 2012; Yang et al., 2013). We selected only lung cancer cell lines. A total of 11,330 interactions were obtained for 122 drugs and 32,000 genes in 106 cell lines.

Then, a study was performed for the prediction of these 106 lines’ interactions with potential inhibitors. We substituted each of the 2,921 candidate inhibitors in turn and predicted the success of the IC₅₀ test. After that, we selected all molecules with a probability of above 0.9, achieving IC₅₀ from the results in the A549 and CALU1 cell lines. These cell lines were chosen due to the most frequent use of these lines in various studies of lung cancer.

The study resulted in the obtained data on the structure of 37 molecules with potential toxicity for lung cancer cells. We performed visualization in our own module and found that the resulting molecules were large and consisted of many repeating structural elements of radicals. Therefore, we decided to isolate the active parts of the molecule for further analysis. As a result, another 15 molecules, after their decomposition, were added to the initial set of 37 molecules.

In the next step, we planned to test the selectivity of the obtained molecules in all 1,018 cell lines. We designed a similar experiment to predict the IC₅₀ value.

We generated 50k patients’ data to reach the ROC-AUC value equal to 0.73. We used the Lasso linear regression algorithm with a 5-fold cross-validation to determine the significance of the features. The result of each experiment was obtained as a list of genes ordered by decreasing impact on the effect. We combined the lists of genes obtained in the experiments with OS and with PFI.

Consequently, we identified 36 genes whose expression correlates with compromised survival indicators in individuals diagnosed with lung cancer. Table 1 presents some characteristics of these genes.

At the stage of implementation of the tumor–normal filter, deep learning was performed according to the aforementioned method with the ROC-AUC indicator of 0.83. A total of 4,912 genes were selected in the process of determining the significant features. The genes with the expression associated with distinguishing a tumor tissue from the normal tissue were isolated from the previously found 36 genes with the help of the obtained list of genes. These 12 genes were DKK4, GP6, LEF1, CUL1, KRT10, PIAS4, FSHR, MYLPF, EFNA3, ZFYVE9, GHSR, and MYL10.

As a result of automated literature mining inhibitors, we obtained Table 2, which presents the number of articles published in response to various requests for each of the 12 genes of interest. Such a table will help draw a conclusion about the studying extent of the gene as a target.

However, some references may not mean that there is a direct connection between the name of the gene and the drug used. In other words, they may not be related to it in terms of inhibition but are simply mentioned in a similar context.

As a result of the NLP search, we added the right column to the table, which lists all inhibitors of a certain gene. These data allow a researcher to make a decision on the basis of the available number of inhibitors for each of the genes under consideration.

As a result, a dataset was obtained from 118,379 pairs, including 19,250 pairs describing the compounds bound to proteins and 99,129 precedents describing non-protein-bound compounds.

Deep learning was applied in a similar way, as in the previous approach. ROC analysis of learning quality allowed us to obtain an average area under the curve of 0.86 (Figure 3).

After search for candidate molecules, we received 160,000 pairs with an interaction probability over 0.99 and 2,921 pairs with an algorithm predicted probability of 1.0.

The following distribution by the inhibited genes was obtained for these 2,921 potential inhibitors (Table 3).

During the cell experiment emulation, the algorithm for determining the importance of the features selected 129 genes. The characteristics of the proteins encoded by the revealed genes are presented in Table 4. ROC is shown on Figure 4.

At the cell experiment emulation stage, we chose interactions with a probability of at least 0.9 from the data on the forecast. Molecules were selected that acted on the minimum number of lines with a probability higher than a given one, i.e., those with the highest specificity.

As a result, five small molecules were selected. The certain cell lines used for validation are “A549,” “NCIH23,” “NCIH460,” “NCIH1299,” “HCT116,” “AMO1,” “PC3,” and “CAPAN1.”