Page 67 - TD-4-1

Tumor Discovery Drug repurposing for pancreatic cancer via AI

Finally, we used backpropagation and the Adam on these criteria. Specifically, we assessed the regulatory
optimal learning algorithm with a learning rate of 0.001 ability of drugs by referring to the LINCS L1000 level
to train the DNN-based DTI model. We set the number 5 dataset, which allowed us to identify drugs capable of
of epochs to 100 and the batch size to 100. The gradient regulating gene expression to normal levels. A regulatory
update algorithm is given as follows: ability value >0 indicates upregulation of gene expression,
while a value <0 suggests downregulation. We also selected
w

(XLI) compounds with small absolute sensitivity values from the
b
Pharmaceutical Regulatory Information System (PRISM)
*
arg minL (XLII) database to minimize interference with normal cells. Most
importantly, in considering drug toxicity, we used tools
f
f 1
f 1 L (XLIII) from the ADMETlab 2.0 website and focused on LC50
values. LC50 represents the concentration at which 50%
of organisms are lethally affected, and higher LC50 values
L f 1 indicate lower toxicity. Therefore, we selected compounds
with higher LC50 values.
with L f 1 w f 1 In conclusion, we followed three drug design
L specifications — toxicity, regulatory ability, and sensitivity
b — to screen potential molecular drugs for the biomarkers
where f represents the f-th iteration of the DNN training of PDAC, ultimately identifying two potential molecular
process, η is the learning rate, and ∇ denotes the gradient drugs. These drugs were then combined into multi-target
operator. therapy for treating PDAC, as shown in Table 2. Ultimately,
To evaluate the model, we employed five-fold cross- we successfully identified a suitable combination of
validation and used receiver operating characteristic molecular drugs for the treatment of PDAC.
(ROC) curves for the binary classification problem. 3. Results
The area under the ROC curve (AUC) is an important
evaluation metric, where a higher AUC value indicates a 3.1. Overview of the systems biology approach to
better prediction of drug-target interaction. The formulas PDAC mechanisms and systematic drug repurposing
for calculating the ROC curves and AUC are as follows: and design

TPRTruePositiveRate TP (XLIV) In this study, a systems biology approach was employed to
TP FN investigate the carcinogenic mechanism of PDAC, utilizing

TNRTrueNegativeRate TN (XLV) big data mining and genome-wide microarray data. This
approach led to the identification of crucial biomarkers of
FP TN PDAC carcinogenesis, which were subsequently targeted

FPRFalse Positive Rate FP (XLVI) for drug repurposing. A DNN-based DTI model was
trained using DTI databases to predict potential drugs
TN FP
targeting these biomarkers. These molecular drugs were

FNRFalse Negative Rate FN (XLVII) designed based on drug design specifications and their
FN TP ability to restore the cellular functions of pancreatic
cancer cells. Finally, the selected molecular drugs were
where TP means the judgment is true and it is indeed considered as a multi-molecular therapeutic strategy
true; TN means the judgment is false and it is indeed false; of PDAC. The flowchart outlining the systems biology
FP means the judgment is true, but it is actually false; FN approach and drug repurposing process for PDAC is
means the judgment is false, but it is actually true. depicted in Figure 1.
Utilizing predictions from the DNN-based DTI model, Initially, to understand the carcinogenic mechanisms
we obtained three candidate molecular drugs for each key underlying PDAC and identify important biomarkers
biomarker, as shown in Table 1. These drugs were screened for therapeutic targeting, candidate GWGENs were
based on some drug design specifications to identify constructed using a big database mining approach.
potential molecular targets for PDAC. In this study, we The following databases were used: starBase, DIP,
27
18
considered pharmacological properties such as regulatory CircuitDB, BioGRID, IntAct, HTRIdatabase, ITFP,
20
22
19
23
25
ability, sensitivity, and toxicity as key design specifications, MINT, TRANSFAC, and TargetScanHuman. The
26
21
24
and selected suitable candidate drugs from Table 1 based candidate GWGENs are represented as a Boolean matrix,
Volume 4 Issue 1 (2025) 59 doi: 10.36922/td.4709