An algorithm for drug discovery based on deep learning with an example of developing a drug for the treatment of lung cancer

In this study, we present an algorithmic framework integrated within the created software platform tailored for the discovery of novel small-molecule anti-tumor agents. Our approach was exemplified in the context of combatting lung cancer. In the initial phase, target identification for therapeutic intervention was accomplished. Leveraging deep learning, we scrutinized gene expression profiles, focusing on those associated with adverse clinical outcomes in lung cancer patients. Augmenting this, generative adversarial neural (GAN) networks were employed to amass additional patient data. This effort yielded a subset of genes definitively linked to unfavorable prognoses. We further employed deep learning to delineate genes capable of discriminating between normal and tumor tissues based on expression patterns. The remaining genes were earmarked as potential targets for precision lung cancer therapy. Subsequently, a dedicated module was formulated to predict the interactions between inhibitors and proteins. To achieve this, protein amino acid sequences and chemical compound formulations engaged in protein interactions were encoded into vectorized representations. Additionally, a deep learning-based component was developed to forecast IC50 values through experimentation on cell lines. Virtual pre-clinical trials employing these inhibitors facilitated the selection of pertinent cell lines for subsequent laboratory assays. In summary, our study culminated in the derivation of several small-molecule formulas projected to bind selectively to specific proteins. This algorithmic platform holds promise in accelerating the identification and design of anti-tumor compounds, a critical pursuit in advancing targeted cancer therapies.

Abstract PO-045: Machine learning for predicting overall survival using whole exome DNA and gene expression data and analyzing the significance of features

The aggressiveness of a tumor depends on its genomic profile. Accordingly, it should be expected that the overall survival of cancer patients also depends on this, in particular, on the number and nature of mutations and the degree of gene activity. In this work, we try to predict overall survival by the genomic profile of the tumor, both by primary DNA and by RNA activity. One of the objectives of the study is to compare which of the presented baseline data better predict overall survival. The data were taken from the pan-cancer TCGA database (33 types of cancer) on DNA and gene expression. They were split into 2 datasets: DNA data only and expression only. In the DNA data, we select only pathogenic and likely pathogenic variants. The total number of genes containing these mutations was 1806, they are accepted as features. In the expression data, we selected only those genes that belong to the cancer-related pathways in the KEGG database (1821 genes). As a prediction effect for both datasets, a 3-year OS was chosen. Accordingly, if a patient crossed the three-year line of OS, he was considered a positive example, otherwise - a negative one. The DNA dataset contained 2159 positive examples and 1687 negative examples. The expression dataset contained 3363 positive and 2212 negative ones. Machine learning algorithms have been implemented using python 3. To determine the significance of the features, we used the Lasso linear regression algorithm with 5-fold cross validation. The result was obtained in the form of list of genes ordered by decreasing importance on the effect. In the DNA dataset, the algorithm selected 64 significant genes, including a sign (plus or minus) indicating an influence on a positive or negative effect, and a coefficient indicating the relative strength of an influence. For example, age 81-90 and EGFR mutations were at the negative end of the scale, while stage I and HRAS mutations were at the positive end. In the RNA dataset, the algorithm selected 75 of such important genes. At the negative end of the scale there were age 81-90 and changes in CDK6 expression, at the positive end - stage I and changes in RPS6 expression. Only 11 of significant features were shared across the two datasets. To predict the effect, we used a logistic regression algorithm with 5-fold cross-validation. Receiver characteristic curves (ROC), reflecting the sensitivity and specificity of the classification, were evaluated by the area under the curve (AUC). For the DNA dataset, the mean ROC-AUC for the 5 predictions was 0.72 (0.64-0.77), for the RNA dataset 0.74 (0.69-0.77). Predicting overall survival is essential for planning treatment strategies and selecting patients for clinical trials. Sufficiently high indicators of the classification quality show that this approach makes sense for further development. Further tuning of the algorithms will make it possible to predict the effect more accurately. Combinations of different input data must be tested. The list of important genes can be helpful in detecting molecular targets in drug discovery.

Citation Format: Dmitrii K. Chebanov, Nadezhda S. Tatevosova, Irina N. Mikhaylova. Machine learning for predicting overall survival using whole exome DNA and gene expression data and analyzing the significance of features [abstract]. In: Proceedings of the AACR Virtual Special Conference on Artificial Intelligence, Diagnosis, and Imaging; 2021 Jan 13-14. Philadelphia (PA): AACR; Clin Cancer Res 2021;27(5_Suppl):Abstract nr PO-045.

Abstract PO086: Method for predicting the effectiveness of the developed immune dendritic cell vaccine in melanoma patients based on cell surface antigens and machine learning with non-classical logic

Patients data: We had 39 patients of 2 categories: with an objective response on dendritic cell vaccine therapy (20) and with disease progression without a response (19). All of them had 21 biomarkers (antigen concentration) as features. The positive effect means that the patient responded to therapy. The features data has quantitative (continuous) values, but we made it categorical by determining the 6 intervals, so each of the biomarkers was replaced with 6 encoding (‘dummy’) variables with possible values 1 or 0, depending on if the patient’s biomarker value belongs to this interval.

Methods: The machine learning algorithm for response prediction is called JSM method for automatic support of scientific research (JSM method ASSR). It allows conducting a plausible reasoning that is realized in hypotheses generating and keeping only those that remain after each database enlargement. The reasoning is based on the similarity of the objects, that can be obtained with patients’ (objects’) features intersection using the statements from the set theory. According to it, the object is representing by a set of features, and hypotheses about its belongings to a class are also sets of features, that are specific for the current class. So, for each class there is a separate amount of hypotheses is generating. On the prediction stage each object given for the prediction is being checked for how many hypotheses are entering into it, or, in other words – is a subset of this object. Based on this information prediction is making: it depends on which hypotheses (of which class) are prevailing in entering in the object. This kind of machine learning approach also allows us to get the reasons why the particular object is classified into his class. So it can be used not only for the classification problem but also for the knowledge discovery about effects’ reasons. We divided the database into 2 batches: source base (18 objects) and first enlargement (17 objects) for the learning, and the rest 4 objects were left for testing. The source base and its enlargement are being permutated during the learning process for more reliability and robustness. We applied a cross-validation, according to which each object was at least 1 time in the test group. So it was 10 learning launches with predictions: 9 with 4 test examples and the rest 1 - with 3 test examples.

Results: On all 10 cross-validation launches, there were 26 correct predictions. Also were 5 cases with a failure, 5 false-positive predictions, and 3 false-negative ones. Recall of the model was 85%, and precision is 77%, F1 score = 0.81. We also obtained reasons, which were common for all the database permutation. It meant that patients who will not respond to therapy should have CD8 value at interval 39.9-54.1 and IRI at interval 0.29-0.7.

Discussion: Actually 39 samples are a small amount of data even for the JSM method ASSR, but we showed the suitability of described approach for the quantity data predicting and the reasons extracting. With the enlargement of the source database, it will be possible to get higher results.

Citation Format: Dmitrii K. Chebanov, Irina N. Mikhaylova, Nadezhda S. Tatevosova. Method for predicting the effectiveness of the developed immune dendritic cell vaccine in melanoma patients based on cell surface antigens and machine learning with non-classical logic [abstract]. In: Abstracts: AACR Virtual Special Conference: Tumor Immunology and Immunotherapy; 2020 Oct 19-20. Philadelphia (PA): AACR; Cancer Immunol Res 2021;9(2 Suppl):Abstract nr PO086.

Immunological monitoring for prediction of clinical response to antitumor vaccine therapy

Immunotherapy has shown promising results in a variety of cancers, including melanoma. However, the responses to therapy are usually heterogeneous, and understanding the factors affecting clinical outcome is still not achieved. Here, we show that immunological monitoring of the vaccine therapy for melanoma patients may help to predict the clinical course of the disease.

We studied cytokine profile of cellular Th1 (IL-2, IL-12, IFN-γ) and humoral Th2 (IL-4, IL-10) immune response, vascular endothelial growth factor (VEGFA), transforming growth factor-β 2 (TGF-β 2), S100 protein (S100A1B and S100BB), adhesion molecule CD44 and serum cytokines β2-microglobulin to analyze different peripheral blood mononuclear cell subpopuations of patients treated with dendritic vaccines and/or cyclophosphamide in melanoma patients in the course of adjuvant treatment.

The obtained data indicate predominance of cellular immunity in the first adjuvant group of patients with durable time to progression and shift to humoral with low cellular immunity in patients with short-term period to progression (increased levels of IL-4 and IL- 10). Beta-2 microglobulin was differentially expressed in adjuvant subgroups: its higher levels correlated with shorter progression-free survival and the total follow-up time. Immunoregulatory index was overall higher in patients with disease progression compared to the group of patients with no signs of disease progression.