Advancing Drug Repurposing in Neuroimmunological Research_adera4.pdf

Advancing Drug Repurposing in Neuroimmunological Research. we show a complete pipeline for

1. Advancing Drug Repurposing in
NeuroimmunologicalResearch: Introducing
Adera2
Michel Edwar Mickael
Dr.Hab.
Instituite of Genetics and Animal Biotechnology
Warsaw
Poland

2. Traditional drug discovery workflows suffer from high cost
and long delays
•The process of discovering a new drug can cost over
2.5 billion dollars
•It can take more than 15 yearsfrom initial research
to market​.
•Most of this time and cost are absorbed in the early
development stages, such as compound
identification, which is where most drugs fail.
•Over 90%of new drugs fail to progress beyond the
early stages of development, making the process
highly inefficient and risky​.

3. What is Drug repurposing and why it is good
•Repurposinga new function for an old
drug
•Repurposing known drugs circumvents
much of this initial risk and cost.
•Since these drugs have already been
approved for human use,
•the expensive and lengthy stages of
safety testing can be bypassed​.
•This approach offers a shortcut to bringing
new treatments to patients,
•especially for diseases where urgent
solutions are needed,
•such as neuroinflammatoryand
neurodegenerative conditions like
depressionand Alzheimer’s disease

4. Bottle neck in drug repurposing of brain
metastasis that we donot know the origin
•Metastatic cancer cells spreading to
distant sites significantly affect
patients’ prognosis and limit
treatment effectiveness
•Identifying the origin of metastatic
cancer cells is essential to prevent
further spreading of said cancerous
cells.
•Failure to correctly pinpoint the
origin of cancer can significantly
reduce survival rates, as seen in
cases of cancer with unknown
primary sites
•Drug effectiveness depend on
accurately discovering the origin of
the cancer

5. •We built an ensemble pipeline that first predict the origin of
metastasis and then find an appropriated new drug for it.
•A) Network 1: finding the metatsis origin using CAN
•B) Network 2: finding old drugs that can hit this specific tumor
How can we use AI use to discover the brain
cancer origin and then repurpose drugs for it?

6. WorkflowoftheAImodelling.
(A)Thedatasetincluded20typesof
cancer.
(B)Weemployedtwoworkflows:
generalizeddeepnetworksand
specializedshallownetworks.Inthefirst
workflow,weexploredtheperformance
oftwodifferentmodels,namelyaCNN-
basednetworkandaReLU-based
network.Inthespecializedshallow
networksworkflow,webuiltshallow
modelsthatdifferentiatebetweencancer
typeswithinthesamesystemorcontext.
Forexample,thedigestivesystemmodel
aimstodistinguishbetweentwocancers,
namely cholangiocarcinomaand
colorectaladenocarcinoma.
Lazarczyk, M., Duda, K., Mickael, M. E., AK, O., Paszkiewicz, J., Kowalczyk, A., Horbańczuk, J. O., & Sacharczuk,
M. (2022). Adera2.0: A Drug Repurposing Workflow for Neuroimmunological Investigations Using Neural
Networks. Molecules, 27(19), 6453. https://doi.org/10.3390/molecules27196453
Mickael, M. E., Kubick, N., Atanasov, A. G., Martinek, P., Horbańczuk, J. O., Floretes, N., Michal, M., Vanecek, T., Paszkiewicz, J., Sacharczuk,
M., & Religa, P. (2024). Using Copy Number Variation Data and Neural Networks to Predict Cancer Metastasis Origin Achieves High Area
under the Curve Value with a Trade-Off in Precision. Curr Issues MolBiol, 46(8), 8301–8319. https://doi.org/10.3390/cimb46080490

7. For the first model, we experimented with several architectures
and optimizers, usingcross-entropy as the loss function. We
employed Softmax as the final layer to generate probability
predictions for the 20 types of cancer. Optimization was carried
out using either the ADAMorADAMAX optimizer with the
categorical cross-entropy option. The main
architecturesexploredwere the following: (a) convolution +
ReLUand (b) SELU + ReLU+ ELU. The SELUmodelutilized
multiple sequential layers within the Keras framework (Table 2).
Mickael, M. E., Kubick, N., Atanasov, A. G., Martinek, P., Horbańczuk, J. O., Floretes, N., Michal, M., Vanecek, T., Paszkiewicz, J., Sacharczuk,
M., & Religa, P. (2024). Using Copy Number Variation Data and Neural Networks to Predict Cancer Metastasis Origin Achieves High Area
under the Curve Value with a Trade-Off in Precision. Curr Issues MolBiol, 46(8), 8301–8319. https://doi.org/10.3390/cimb46080490

8. For the specialized models, we
employed binary classification
either with a singlelayer
classifier or a classifier
combined with a ReLUlayer.
Subsequently, we
calculatedaccuracyand loss.
Internal validation was
conducted using the train–test
split method,dividingthe
database in a 3:2 ratio
Mickael, M. E., Kubick, N., Atanasov, A. G., Martinek, P., Horbańczuk, J. O., Floretes, N., Michal, M., Vanecek, T., Paszkiewicz, J., Sacharczuk,
M., & Religa, P. (2024). Using Copy Number Variation Data and Neural Networks to Predict Cancer Metastasis Origin Achieves High Area
under the Curve Value with a Trade-Off in Precision. Curr Issues MolBiol, 46(8), 8301–8319. https://doi.org/10.3390/cimb46080490

9. Comparison of AUC values
between the two generalized
models. Both models, (A)
CNNand(B) SELU, achieved
significant AUC values.
CNN
SELU
MALE BRAIN
Mickael, M. E., Kubick, N., Atanasov, A. G., Martinek, P., Horbańczuk, J. O., Floretes, N., Michal, M., Vanecek, T., Paszkiewicz, J., Sacharczuk,
M., & Religa, P. (2024). Using Copy Number Variation Data and Neural Networks to Predict Cancer Metastasis Origin Achieves High Area
under the Curve Value with a Trade-Off in Precision. Curr Issues MolBiol, 46(8), 8301–8319. https://doi.org/10.3390/cimb46080490

10. Figure 8. Comparison of the models’
performance. We calculated accuracy, precision,
recall, and F1score for each of the models
studied. While all the models achieved an
accuracy higher than 80%,none of them
achieved a precision higher than 60%. However,
the brain cancer model achieved an80% recall
Mickael, M. E., Kubick, N., Atanasov, A. G., Martinek, P., Horbańczuk, J. O., Floretes, N., Michal, M., Vanecek, T., Paszkiewicz, J., Sacharczuk,
M., & Religa, P. (2024). Using Copy Number Variation Data and Neural Networks to Predict Cancer Metastasis Origin Achieves High Area
under the Curve Value with a Trade-Off in Precision. Curr Issues MolBiol, 46(8), 8301–8319. https://doi.org/10.3390/cimb46080490

11. Figure 9. Effects of parameter fine-tuning. We explored the
impact of tuning parameters using a gridsearchand Keras
tuner. (A) Regarding batch size, our findings suggest that
reducing the batch sizecanenhance accuracy. (B) Analysis of
the hyperparameterinvolving learning rate, clip norm,
andvalidationsplit indicates that each parameter can
profoundly affect accuracy.
Mickael, M. E., Kubick, N., Atanasov, A. G., Martinek, P., Horbańczuk, J. O., Floretes, N., Michal, M., Vanecek, T., Paszkiewicz, J., Sacharczuk,
M., & Religa, P. (2024). Using Copy Number Variation Data and Neural Networks to Predict Cancer Metastasis Origin Achieves High Area
under the Curve Value with a Trade-Off in Precision. Curr Issues MolBiol, 46(8), 8301–8319. https://doi.org/10.3390/cimb46080490

12. Figure 11. Biological explainability
analysis of the generalized model
performance. Our resultsindicatethat
the gene locus parameters have a
larger impact on model accuracy in
comparison toCNAtype. Model
accuracy is more dependent on the
dense layer with the ReLUactivation
functionthanon the convolution
network layers (solid line represents
ReLUlayers and dashed line
representstheconvolution layer)
Mickael, M. E., Kubick, N., Atanasov, A. G., Martinek, P., Horbańczuk, J. O., Floretes, N., Michal, M., Vanecek, T., Paszkiewicz, J., Sacharczuk,
M., & Religa, P. (2024). Using Copy Number Variation Data and Neural Networks to Predict Cancer Metastasis Origin Achieves High Area
under the Curve Value with a Trade-Off in Precision. Curr Issues MolBiol, 46(8), 8301–8319. https://doi.org/10.3390/cimb46080490

13. Figure 12. Box plot of the features
based on their respective SHAP
values. The chromosome
numberparameteris the most
important parameter in determining
the model’s performance. The
resultsindicatethat the chromosome
parameter is the most influential in
the model’s decision-makingprocess,
as evidenced by its larger variation
and numerous outliers. The start and
end parametersalsocontribute to the
model’s predictions to a lesser
extent. Amplification, deletion, and
strand havetheleast impact, with
their SHAP values being consistently
close to zero.
Mickael, M. E., Kubick, N., Atanasov, A. G., Martinek, P., Horbańczuk, J. O., Floretes, N., Michal, M., Vanecek, T., Paszkiewicz, J., Sacharczuk,
M., & Religa, P. (2024). Using Copy Number Variation Data and Neural Networks to Predict Cancer Metastasis Origin Achieves High Area
under the Curve Value with a Trade-Off in Precision. Curr Issues MolBiol, 46(8), 8301–8319. https://doi.org/10.3390/cimb46080490

14. PHASE II
Artificial intelligence identifies 9 putative compounds as
NRF2 activators. (2a) Deep neural network text mining
workflow. (2b) Embedding is calculated using a deep
averaging network (DAN) network (2c) Sentence similarity
is calculated between the imposed question and each
sentence and each paragraph of each document. (2d) We
have chosen the inner product to score similarity because it
has higher accuracy. (2e) We used a differential convolution
network (DCN) network to extract features between each
answer of the previous phase and the question. The DCN
consist of a 2D layer and dense layer implemented in Keras
library in python. After that, we used an mean squared
error (MSE) approximation to calculate the distance
between features of each answer and the imposed
question.
Kubick, N., Pajares, M., Enache, I., Manda, G., & Mickael, M. E. (2020). Repurposing Zileuton as a
Depression Drug Using an AI and In Vitro Approach. Molecules, 25(9), 2155.
https://doi.org/10.3390/molecules25092155

15. Physiochemical comparison of the
resulting drugs.
(3a) clogPcalculation showing TCDD
having high lipophilicity(logP> 5),
which often contributes to high
metabolic turnover, low solubility,
and poor oral absorption while
lisinoprilis significantly soluble in
aqueous solutions. (3b) Overall
score of the drugs. (3c) Doxorubicin
probability of crossing the BBB is
very low. (3d) Overall comparison of
filtered drugs showing zileutonand
caffeicacid to be the most suitable
compounds for further
investigations. (3e)and (3f) Zileuton
structure showing its biosuitability. Kubick, N., Pajares, M., Enache, I., Manda, G., & Mickael, M. E. (2020). Repurposing Zileuton as a
Depression Drug Using an AI and In Vitro Approach. Molecules, 25(9), 2155.
https://doi.org/10.3390/molecules25092155

16. EXPERIMENTAL VALIDATION
Zileuton activates Nrf2. A, Representative immunoblots for the indicated
proteins in cell lysates from RAW264.7 cells treated with zileuton(10 µM, 16
h) in the absence of serum. (4b) Densitometricquantification of
representative immunoblots from (4a) relative to ACTB protein levels. Data
are mean ±SEM (n = 4). Statistical analysis was performed using Student’s t
test. * p < 0.05; **** p < 0.001 vs. vehicle-treated cells. (4c) Zileuton model
of action. In response to reactive oxygen species (ROS) stress, AA is released
from membrane phopholipidsby phospholipases. Free AA can be converted
to bioactive eicosanoids through the cyclooxygenase (COX), lipoxygenase
(LOX), or P-450 epoxygenasepathways. LOX enzymes (5-LO, 12-LO, and 15-
LO) catalyze the formation of LTs, 12(S)hydroperoxyeicosatetraenoicacids
and lipoxins(LXs), respectively. COX isozymes (constitutive COX-1 and
inducible COX-2) catalyze the formation prostaglandin. The P-450
epoxygenasepathway catalyzes the formation of hydroxyeicosatetraenoic
acids (HETEs) and epoxides. Zileuton was shown to inhibit 5-LO as well as
prostaglandin production through suppressing prostaglandin biosynthesis by
inhibition of arachidonic acid release in macrophages. Zileuton can also
activate NRF2.
Kubick, N., Pajares, M., Enache, I., Manda, G., & Mickael, M. E. (2020). Repurposing Zileuton as a
Depression Drug Using an AI and In Vitro Approach. Molecules, 25(9), 2155.
https://doi.org/10.3390/molecules25092155

17. Summary
•We have developed a pipeline capable of identifying markers
associated with cancers that metastasize to the brain, as well as
repurposing drugs targeted at these specific cancers.

18. Announcements
•Our lab integrates both computational and experimental research
approaches.
•Available Positions:
•PhD Candidates
•MSc Students