Page 70 - TD-4-1

Tumor Discovery Drug repurposing for pancreatic cancer via AI

512, 256, 128, and 64 nodes, respectively, each using the resection, radiotherapy, chemotherapy, immunotherapy,
ReLU activation function. The output layer consists of a and targeted therapy. However, not all patients are suitable
single node with sigmoid activation function. To prevent candidates for surgery, and standard chemotherapy
overfitting during training, we added a dropout layer. regimens may be ineffective due to the complex interactions
After training the DNN-DTI model to predict candidate between healthy pancreatic cells, cancer cells, and the
molecular drugs for these drug targets, we evaluated the tumor microenvironment, leading to drug resistance and
model’s learning effectiveness. Figures S3 and S4 show the suboptimal therapeutic outcomes. Therefore, there is a
accuracy and loss during the training process, respectively. pressing need to explore new treatment approaches.
The five-fold cross-validation was employed to assess the Given the substantial financial costs and time required
model’s performance, achieving an average accuracy of for the discovery of new drugs, repurposing existing drugs
98.3% and a standard deviation of 0.138%, as shown in for new therapeutic indications has become an attractive
Table S1. Additionally, the AUC was used to evaluate the alternative. Drug repurposing involves identifying new
model’s classification performance, as shown in Figure S5. uses for already-approved drugs beyond their original
An AUC of 0.5 represents random guessing, while an AUC 17
of 1 indicates perfect classification. The DNN-DTI model medical indications.
achieved an AUC of 0.980, indicating that its predictive In this study, we employed database mining and
ability is much superior to random guessing and close to genome-wide microarray data from PDAC and healthy
perfect classification. This highly efficient DNN-based DTI controls, utilizing systems biology approaches to identify
model enables us to accurately predict the probability of their core GWGENs. These networks were annotated
interactions between drugs and the selected biomarkers. using the KEGG database to establish the core signaling
pathways of PDAC and the associated downstream cellular
To identify suitable potential drugs, we considered
three drug design specifications to ensure the rationality dysfunctions, as shown in Figure 2. After investigating
and effectiveness of the candidate multi-molecular drugs the oncogenic mechanisms of PDAC and identifying key
predicted by the DNN-DTI model. These specifications biomarkers suitable for drug targeting, we trained a DNN-
include regulatory ability, sensitivity, and toxicity, among based DTI model using data from DTI databases to predict
other pharmacological properties. The regulatory ability the probability of interaction between these biomarkers
of the drugs was assessed using the LINCS L1000 level 5 and candidate molecular drugs. The model was validated
dataset, guiding the selection of drugs that could restore using five-fold cross-validation, as shown in Table S1, and
key biomarkers to their normal expression levels. the drug repurposing flowchart is presented in Figure 3.
58
A regulatory ability value >0 indicates an increase in We subsequently screened potential drugs as a multi-
gene expression, while a value <0 indicates a decrease. molecular therapy based on drug design specifications,
Additionally, the sensitivity of the drugs was assessed using focusing on sensitivity, toxicity, and regulatory capability.
the PRISM dataset, and we selected drugs with small Ultimately, we predicted a combination of potential
59
absolute sensitivity values to avoid excessive chemical molecular drugs, MK-2206 and gemcitabine, to modulate
perturbation from potential drugs. Finally, we evaluated the overexpression of c-MYC and FOXO3, as well as the
the drug toxicity using tools from the ADMETlab 2.0 mutation of TP53, as shown in Table 2.
website, which calculates LC50. A higher LC50 value Gemcitabine is one of the most widely used treatments
60
indicates lower toxicity, which helps avoid life-threatening for PDAC. As a deoxycytidine nucleoside analog, it
10
from low drug doses. Table 1 presents the candidate inhibits DNA chain elongation by phosphorylation after
molecular drugs predicted by the DNN-DTI model for the entering the cells, leading to cell apoptosis and death.
selected biomarkers, listing their relevant pharmacological However, the therapeutic effect of gemcitabine is limited
properties, such as regulatory ability, sensitivity, and by its unstable metabolism and the potential for drug
toxicity. Based on these drug design specifications, we resistance. As a result, it is often used in combination
61
selected two potential molecular drugs, MK-2206 and with other drugs, such as Fluorouracil and paclitaxel, 62,63
gemcitabine, which have adequate regulatory ability, to enhance treatment efficacy. Studies have pointed out
normal sensitivity (small absolute value), and weaker that gemcitabine can induce the activation of mutated
toxicity, as shown in Table 2. These drugs were combined p53, leading to cancer cell death, as the drug causes the
64
into a multi-molecular therapeutic strategy for PDAC. accumulation of Bax downstream of TP53, which induces
apoptosis. 65,66 In our research, we referenced the LINCS
4. Discussion L1000 level 5 dataset and found that gemcitabine can reduce

Currently, treatment options for exocrine ductal the expression c-MYC. However, the high expression of
67
adenocarcinoma, also known as PDAC, include surgical c-MYC may induce resistance to gemcitabine. Therefore,
68
Volume 4 Issue 1 (2025) 62 doi: 10.36922/td.4709