In-Silico Approach to Searching and Qualifying Histone Deacetylase 6 Novel Inhibitors

Nguyen Huy Trung Kien

Abstract:

Histone deacetylase 6 (HDAC 6) is appreciated for its role in regulating and influencing the epigenome landscape of the chromatin structure, and notably it has been shown in numerous research findings that the protein relates to pathogenesis, thus posing a significant pharmacological value, in various cancers, such as breast cancer and multiple myeloma. Utilizing a ANN-based QSAR model, this paper proposes a list of top 10 novel inhibitors with known IC50 value (the concentration needed to inhibit a biological entity and reduce the respective biological processes by half) from a dataset of over 50 thousands molecules. And lastly, to obtain the visual illustrations of multiple binding configurations of the ligands in HDAC6, as well as to predict the binding energy values and find the root mean square deviation values (rmsd) of the protein-ligand interactions, docking had been practiced on Chimera, using the Autodock Vina tool, to collect these statistics.

Amongst the top 10 molecules being found, it is determined that the lowest docking score is -8.3 kcal·mol^-1 and the highest one is-7.0 kcal·mol^-1. Entropy change and solvation effects render certain intermolecular interactions hard to be accounted for, and therefore, the molecular docking has a drawback of not being entirely confident in the accuracy of the binding energy calculated. This research could be further improved by performing more docking trials using a variety of platforms, such as PyRosetta, GOLD, and MOE-Dock, to approach closer to the true docking score values.

Keywords: Histone deacetylase 6, ANN-based QSAR model, IC50, protein-ligand docking, 1H-pyrazolo[3,4-b]pyridine derivatives.

Introduction:

Human’s epigenome is known for being dynamic, or rather, being able to “turn on” and “turn off” by multiple different biological entities. A large part of this plasticity is devoted to the epigenetic mechanisms that underlie the regulation of gene expression. In addition to DNA methyltransferase (DNMT) that works by transferring the methyl group (-CH₃) from the methyl donor S-adenosyl methionine (SAM) to the 5-position of cytosine residues in DNA, histone deacetylase – a histone modifying protein – also plays a significant role in epigenetic regulations by removing the acetyl group from the lysine residue in the tail regions of histone protein (Barry and Townsend). In that, histone deacetylase 6 (HDAC 6) stands out as a distinct regulator that does not only participate in the processes of histone deacetylation and acetylation but also target a wide-range of other functional cytoplasmic proteins, including the cylindromatosis tumor suppressor gene, α-tubulin, cortactin, heat shock protein 90, β-catenin, and Survivin, which regulates a variety of molecular processes such as cellular proliferation, apoptosis, cellular adhesion and cellular migratory (Aldana-Masangkay and Sakamoto). Henceforth, the protein is closely related to tumor’s growth, invasion and metastasis.(Li et al; Anh Tran et al; Kim et al.)

In a study that investigates the relationship between breast cancer and HDAC6 protein, it has been found that there tends to be an overexpression of HDAC6 in cancer patients. In another study, higher HDAC6 mRNA expression in 139 female patients with invasive breast cancer indicates a better prognosis (Saji et al.). Being an estrogen-regulated gene, HDAC6 decreases in its level when being treated with tamoxifen, which competes with 17β-estradiol (E2) at the estrogen receptor (Aldana-Masangkay and Sakamoto; Yu and Bender). Moreover, such an overexpression of the protein also presents in low-grade and high-grade ovarian carcinomas and acute myeloid leukemia. (Bazzaro et al; Bradbury et al.)

The overexpression of HDAC6 protein has a positive relationship with the process of tumorigenesis by hypoacetylating α-tubulin and granting the cells certain phenotypes such as cell migration and chemotaxis (Yoo et al.). Thereby, more HDAC6-specific inhibitors, i.e. ricolinostat, should be discovered and searched for.

Recently, numerous publications in the field have found and strongly corroborated that the HDAC6 protein, being treated with selective inhibitors, has immense potential in cancer immunotherapy by controlling the expression of tumor-associated antigens, programmed death receptor-1 (PD-1), and programmed death receptor ligand-1 (PD-L1), thereby decreasing the anti-inflammatory phenotype of pro-tumorigenic macrophage and the immunosuppressive pathways caused by tumor’s proteins. (Li et al; Knox et al.)

Further investigations in potential novel HDAC6 inhibitors will aid in developing new drugs that could be used to treat certain types of cancer such as multiple myeloma, melanoma, gastrointestinal cancer, urothelial cancer, and triple-negative breast cancer. (Kuroki et al; Alothaim et al.)

Method:

2.1 The Identification and preparation of the target protein

HDA6 is identified through literature as one of the proteins that plays a prominent role in epigenetics and tumorigenesis. After choosing HDAC6 as a target for the study, the virtual screen experiment starts by preparing the protein. ChemBL is where the id for Histone deacetylase 6 has been taken from. Please visit the link for more information. In addition to providing the information for the target protein, ChemBL id of HDAC6 also encodes in itself data of molecules with known IC50 values.

2.2 Extract compound set and generate molecular fingerprints

To generate a dataset of fingerprints for the inhibitors, coding was carried out on Google Colab. The process mainly involved installing ChemBL web resource and RDkits. Afterwards, command lines that filter the large database and selectively choose inhibitors of HDAC6 with known IC50 values. In the export file, information regarding molecules’ ChemBL ids, canonical SMILES and pChEMBL values ( -log₁₀ (molar IC50) ). The resulting file will therefore be stored in Google Drive as “.txt” format.

2.3 ANN-QSAR model

Typically, the QSAR model works by associating chemical structure’s descriptors of compounds and their physicochemical properties (H-bond, molecular weight, electronegativity,etc. .) and biological activities (pharmacokinetic properties such as absorption, distribution, metabolism, toxicity) (Mandlik et al; Afshar et al.). By utilizing multiple linear regression and partial least square regression, QSAR model can predict linear relationships, based on observed trend and correlations (Mandlik et al; Davis). Nonetheless, in face of more complex situations, such as when there’s an unknown compound to analyze, ANNs aid in the process by granting the ability to decipher for non-linear relationships. (Mandlik et al.)

After the input csv file is loaded, the dataset is processed further by having the training features (molecular descriptors) separated from the label (IC₅₀ values). Then, given that this is a binary classification task, the molecules will be sorted into 2 cohorts (active and inactive), with active compounds possessing IC₅₀ value greater than 6 μM and smaller than 15 μM:

train_labels = np.where(np.logical_and(np.greater_equal(train_labels,6.0),np.less_equal(train_labels,15)), 1, 0)

After that, the dataset is split into a training set–a model that approximates the distribution of the population–and an independent test set–the model that will be used to test the trained model. When generating such an unbiased split, a standard 80:20 split represents a prudent choice.

2.4 Defining a neural network architecture

The model is set to go from input features to the first hidden layer of 32 neurons, which use ReLU as the activation function, then to a 1-neuron output layer at the end, which uses Sigmoid. Next, the fit function initiates the training process.

2.5 Making predictions

With the data available, predicted probabilities can be obtained using the predict () function; if the probability is greater than 0.5 we assign 1 (active), and 0 (inactive) otherwise.

2.6 Model evaluation

A confusion matrix can scrutinize the classification task done by our machine learning model.

from sklearn.metrics import confusion_matrix

print(confusion_matrix(y_test, prediction_classes))

2.7 Rescue overtraining with Dropout

Training a large neural network on such a small dataset can oftentimes lead to overtraining, causing an inability to generalize and produce an erroneous output due to statistical noise in the training data when an unseen data is used (Brownlee). One of the ways to remedy this is via fitting different neural networks, generated by dropping out neurons that act as feature detectors, on the same dataset and to averager the predictions from each model (Brownlee).

y1_model_dropout = tf.keras.Sequential([

layers.Dense(32, activation = activations.relu),

layers.Dropout(0.05),

layers.Dense(1, activation = activations.sigmoid)

])

y1_model_dropout.compile(loss = tf.keras.losses.BinaryCrossentropy(), optimizer = tf.optimizers.Adam(),

metrics=[

tf.keras.metrics.AUC(name=’auc’),

tf.keras.metrics.BinaryAccuracy(name=’accuracy’),

tf.keras.metrics.Precision(name=’precision’),

tf.keras.metrics.Recall(name=’recall’)

]

)

2.8 Perform virtual screening using the trained QSAR model

Drug virtual screening on 50 thousands external molecules is obtained from the Mcule database. Similar to the training process, the process starts with computing the molecular fingerprints for 50K external compounds and then using the trained QSAR-ANN model to make predictions regarding their inhibitory effects on HDAC6.

mcule_predictions = hdac6_model_dropout.predict(hdac6_antagonist_mcule_fp)

Afterwards, the list of molecules can be sorted based on their descending predicted inhibitory effect on HDAC6.

sorted_mcule_dataset = mcule_dataset.sort_values(by=’prediction’,ascending=False)

2.9 Protein-ligand docking experiment

After obtaining the pdb file of HDAC6 from Protein Data Bank and the 3D SDF of each of the top 10 molecules identified from PubChem, the study continues by using Chimera to prepare the protein and the molecule for docking. For the protein HDAC6 part, its structure is first removed from the binding ubiquitin. After that, the structure editing function in Chimera is utilized, where appropriate tasks such as solvent removal, hydrogen additions and charges additions are performed. Similarly, when loading in the ligand molecule, its structure is also edited by adding hydrogen and charges for the docking process, which is then executed by using Autodock Vina. A full protocol for the global docking process could be found in the following link, which details every single step needed to complete the experiment.

Results:

Active compounds extracted

From ChemBL, it is indicated that there are 4913 compounds with known IC_5o values, with the lowest pChEMBL value ( -log₁₀ [molar IC₅₀] ) being 4.85 μM and the highest one being 7.59 μM.

Among those molecules, there are 1687 inactive and 3218 active.

Training QSAR-ANN model

After the fit function initiates the training process, metrics regarding loss, accuracy, precision and recall during training will be visualized using Matplotlib. In the visualization, it is clear that as the number of epoch increases, meaning to say as we train the model, AUC (Area Under The Curve), accuracy, precision and recall increase, while loss decreases. Specifically, the value of AUC, accuracy, precision and recall are 0.87, 0.88, 0.91, and 0.92, respectively. Upon evaluating the testing data with the confusion matrix, it is found that there are 61 false positives and 52 false negatives.

After performing dropout to rescue the model from overtraining, there are slight improvements in the performance metrics, compared to the ANN without the dropout, with the values of AUC, accuracy, precision and recall being 0.88, 0.89, 0.92, and 0.92, respectively.

To further improve, additional layers are added to the network, and the number of neurons are increased to 1000. It is apparent that AUC, accuracy, precision and recall’s values all rise, and the numbers of false positives and false negatives also go down.

Perform virtual screen of a test dataset

Overall, there are a total of 1687 inactive molecules and 3218 active molecules. In this dataset, the vast majority of the molecules are active, and some are inactive, as indicated in the visualization of the distribution of the IC50 values of HDAC6 inhibitors. There are many compounds with IC₅₀ values of 100. Those compounds with the IC₅₀ value of were actually inactive because measurable IC₅₀ values are always less than 15.

Since this is a binary classification task, we will divide the compounds in the training set into 2 groups : active and inactive. We will use np.where to convert the IC₅₀ values into binary values (1 for 6 < IC50 < 50, 0 otherwise).

After performing evaluation on test data with the confusion matrix, the model demonstrates that there are about 50 false positives and 54 false negatives. In addition to that, the values of AUC, accuracy, precision and recall are 0.88, 0.89, 0.92, and 0.92, respectively. And lastly, the distribution of the predictions on the test set indicate most of the predictions are active and then some are inactive, which is expected as we have more active compounds in the test set. In fact, the distribution of the predicted labels for the test set are very close to the distribution of their real labels.

Drug virtual screening on the 50K external molecules dataset

After plotting the predicted HDAC6 inhibiting activity, it is crystal-clear that there are compounds with relatively high predictions (> 0.8).

After sorting the compounds based on their descending predicted HDAC6 inhibiting activity, the model then shows the following table, summarizing the molecules’ inhibiting activities.

Top molecules with highest prediction scores are then selected by the model, resulting in the following cohort of compounds (shown in the table below). The 10 candidates are:

1. 3-methyl-4-[4-(methylsulfanyl)phenyl]-1-([1,2,4]triazolo[4,3-b]pyridazin-6-yl)-4,5-dihydro-1H-pyrazolo[3,4-b]pyridin-6-ol
2. 4-[3-methyl-1-(3-methyl-[1,2,4]triazolo[4,3-b]pyridazin-6-yl)-6-oxo-5,7-dihydro-4H-pyrazolo[3,4-b]pyridin-4-yl]benzonitrile
3. 1-propan-2-yl-4-(1,3,5-trimethylpyrazol-4-yl)-2,4,5,7-tetrahydropyrazolo[3,4-b]pyridine-3,6-dione
4. 4-(2-bromophenyl)-1-(6-methoxypyridazin-3-yl)-3-methyl-5,7-dihydro-4H-pyrazolo[3,4-b]pyridin-6-one
5. 4-(3-methoxy-2-propoxyphenyl)-1-(6-methoxypyridazin-3-yl)-3-methyl-5,7-dihydro-4H-pyrazolo[3,4-b]pyridin-6-one
6. 4-(3-chlorophenyl)-1-(6-methoxypyridazin-3-yl)-3-methyl-5,7-dihydro-4H-pyrazolo[3,4-b]pyridin-6-one
7. 4-(2-methoxyphenyl)-3-methyl-1-(3-methyl-[1,2,4]triazolo[4,3-b]pyridazin-6-yl)-5,7-dihydro-4H-pyrazolo[3,4-b]pyridin-6-one
8. 4-(4-ethoxyphenyl)-3-methyl-1-(6-morpholin-4-ylpyridazin-3-yl)-5,7-dihydro-4H-pyrazolo[3,4-b]pyridin-6-one
9. 4-(1,5-dimethylpyrazol-4-yl)-1-propan-2-yl-2,4,5,7-tetrahydropyrazolo[3,4-b]pyridine-3,6-dione
10. 4-(3-chlorophenyl)-3-methyl-1-(3-methyl-[1,2,4]triazolo[4,3-b]pyridazin-6-yl)-5,7-dihydro-4H-pyrazolo[3,4-b]pyridin-6-one

All of these molecules are later shown to be 1H-pyrazolo[3,4-b]pyridine derivatives, which pose immense pharmacological significance.

Protein-ligand docking result

After selecting the top 10 screened molecules based on their IC₅₀ values, the study proceeds to dock the ligands to HDAC6 using Chimera and AutoDock Vina, thereby recording several parameters such as docking score (binding affinity), root-mean-square deviation and the presence of hydrogen bonds in the interaction sites. Lower docking scores correspond to higher binding affinity and stable interaction(s), thus those molecules representing options as better drug candidates. The figure below shows the results of Chimera AutoDock.

Compound	Binding Affinity (kcal/mol)	Mode	RMSD lower Bond	RMSD upper Bond
	-7.9	1	0.0	0.0
3-methyl-4-[4-(methylsulfanyl)phenyl]-1-([1,2,4]triazolo[4,3-b]pyridazin-6-yl)-4,5-dihydro-1H-pyrazolo[3,4-b]pyridin-6-ol	-7.7	2	25.002	28.052
	-7.6	3	25.937	28.118
	-7.6	4	27.192	29.056
	-7.5	5	24.576	26.976
	-7.3	6	23.079	25.006
	-7.1	7	24.09	26.732
	-7.0	8	22.897	24.874
	-7.0	9	25.969	28.448
	-7.0	10	23.611	26.082
4-[3-methyl-1-(3-methyl-[1,2,4]triazolo[4,3-b]pyridazin-6-yl)-6-oxo-5,7-dihydro-4H-pyrazolo[3,4-b]pyridin-4-yl]benzonitrile	-8.2	1	0.0	0.0
	-8.2	2	6.457	8.833
	-7.9	3	23.275	25.211
	-7.8	4	0.918	2.073
	-7.6	5	22.11	23.646
	-7.5	6	26.183	29.496
	-7.5	7	1.662	2.211
	-7.4	8	32.529	35.344
	-7.3	9	18.885	22.246
	-7.3	10	7.132	11.056
1-propan-2-yl-4-(1,3,5-trimethylpyrazol-4-yl)-2,4,5,7-tetrahydropyrazolo[3,4-b]pyridine-3,6-dione	-7.0	1	0.00	0.0
	-6.8	2	24.168	27.184
	-6.7	3	23.624	26.836
	-6.7	4	23.49	28.232
	-6.6	5	25.511	29.323
	-6.5	6	27.223	29.175
	-6.5	7	27.003	28.623
	-6.5	8	26.163	29.059
	-6.4	9	22.262	24.433
	-6.3	10	15.35	17.8
4-(2-bromophenyl)-1-(6-methoxypyridazin-3-yl)-3-methyl-5,7-dihydro-4H-pyrazolo[3,4-b]pyridin-6-one	-8.0	1	0.0	0.0
	-7.4	2	22.406	23.324
	-7.3	3	23.146	25.706
	-7.2	4	23.495	25.731
	-7.2	5	22.936	25.28
	-7.2	6	34.388	36.555
	-7.2	7	32.238	34.734
	-7.1	8	20.842	21.935
	-7.0	9	22.212	25.219
	-6.9	10	21.656	23.462
4-(3-methoxy-2-propoxyphenyl)-1-(6-methoxypyridazin-3-yl)-3-methyl-5,7-dihydro-4H-pyrazolo[3,4-b]pyridin-6-one	-7.8	1	0.0	0.0
	-7.5	2	24.56	27.824
	-7.5	3	1.741	3.482
	-7.2	4	2.875	8.39
	-7.1	5	33.978	36.394
	-7.0	6	32.158	34.599
	-7.0	7	21.142	23.793
	-6.8	8	34.137	37.048
	-6.7	9	32.142	34.982
	-6.7	10	19.149	20.916
4-(3-chlorophenyl)-1-(6-methoxypyridazin-3-yl)-3-methyl-5,7-dihydro-4H-pyrazolo[3,4-b]pyridin-6-one	-8.3	1	0.0	0.0
	-7.8	2	4.679	7.978
	-7.8	3	23.375	24.979
	-7.8	4	24.874	26.8
	-7.7	5	24.155	26.878
	-.76	6	23.234	25.669
	-7.5	7	23.659	25.552
	-7.4	8	5.146	7.35
	-7.3	9	20.452	24.416
	-7.2	10	23.602	24.924
4-(2-methoxyphenyl)-3-methyl-1-(3-methyl-[1,2,4]triazolo[4,3-b]pyridazin-6-yl)-5,7-dihydro-4H-pyrazolo[3,4-b]pyridin-6-one	-7.9	1	0.0	0.0
	-7.8	2	4.238	8.446
	-7.6	3	26.707	29.552
	-7.4	4	28.33	30.595
	-7.4	5	32.805	35.469
	-7.2	6	44.19	46.368
	-7.2	7	14.997	17.359
	-7.0	8	28.077	30.715
	-7.0	9	35.097	37.432
	-6.9	10	34.891	37.21
4-(4-ethoxyphenyl)-3-methyl-1-(6-morpholin-4-ylpyridazin-3-yl)-5,7-dihydro-4H-pyrazolo[3,4-b]pyridin-6-one	-8.3	1	0.0	0.0
	-7.9	2	22.723	25.992
	-7.6	3	11.17	16.411
	-7.5	4	18.133	20.61
	-7.5	5	4.323	6.691
	-7.5	6	19.205	22.276
	-7.3	7	33.702	36.282
	-7.3	8	3.203	9.479
	-7.1	9	3.174	4.587
	-7.0	10	41.748	44.75
4-(1,5-dimethylpyrazol-4-yl)-1-propan-2-yl-2,4,5,7-tetrahydropyrazolo[3,4-b]pyridine-3,6-dione	-7.4	1	0.0	0.0
	-7.2	2	23.964	26.585
	-6.9	3	22.484	24.965
	-6.6	4	34.369	35.947
	-6.3	5	34.038	36.4
	-6.3	6	2.136	3.688
	-6.2	7	1.85	5.773
	-6.2	8	23.879	25.203
	-6.2	9	22.799	24.172
	-6.2	10	22.292	22.292
4-(3-chlorophenyl)-3-methyl-1-(3-methyl-[1,2,4]triazolo[4,3-b]pyridazin-6-yl)-5,7-dihydro-4H-pyrazolo[3,4-b]pyridin-6-one	-8.1	1	0.0	0.0
	-8.1	2	4.893	8.279
	-8.0	3	26.093	28.21
	-7.8	4	23.663	26.816
	-7.7	5	26.661	28.452
	-7.6	6	1.918	2.169
	-7.3	7	36.23	39.921
	-7.3	8	22.356	24.117
	-7.2	9	4.073	7.089
	-7.2	10	17.401	19.987

The result indicates that the selected molecules have relatively strong binding affinity with the drug target, with 4-(3-chlorophenyl)-1-(6-methoxypyridazin-3-yl)-3-methyl-5,7-dihydro-4H-pyrazolo[3,4-b]pyridin-6-one and 4-(4-ethoxyphenyl)-3-methyl-1-(6-morpholin-4-ylpyridazin-3-yl)-5,7-dihydro-4H-pyrazolo[3,4-b]pyridin-6-one having the highest binding affinity of -8.30 kcal/mol. The interaction of the ligands and the target protein were visualized respectively through Chimera as shown below.

Interaction of HDAC6 with 4-(3-chlorophenyl)-1-(6-methoxypyridazin-3-yl)-3-methyl-5,7-dihydro-4H-pyrazolo[3,4-b]pyridin-6-one

Interaction of HDAC6 with 4-(4-ethoxyphenyl)-3-methyl-1-(6-morpholin-4-ylpyridazin-3-yl)-5,7-dihydro-4H-pyrazolo[3,4-b]pyridin-6-one

Conclusion and Discussion:

As investigated by ChemDraw, the chosen compounds pose a similar parent chain which render them to be the derivatives of 1H-pyrazolo[3,4-b]pyridine.

1H-Pyrazolo[3,4-b]pyridines are bicyclic heterocyclic compounds that garner great interests from medicinal chemists for its resemblance with adenine and guanine, which are formed from the fusion of pyrazole and pyridine and have 2 possible tautomeric isomers: 1H and 2H. Studies have shown that 1H- tautomers have a greater stability with an energy difference by 37.03 kJ/mol (Donaire-Arias et al.). Literature suggests that different possible substituents enable it to have different biological activities. In addition, a wide range of substituents also make 1H-Pyrazolo[3,4-b]pyridine widely used as a scaffold for different small molecules with therapeutic purposes. Specifically, 15% of the 1H-Pyrazolo[3,4-b]pyridine derivatives are utilized for medicinal purposes and in that, 22675 molecules are anti-tumor agents (Donaire-Arias et al.).

Given lines of finding, it is clear that these molecules represent a brand-new class of small molecules with the potential anticarcinogenic activities by interacting with HDAC6. In the market, there are several drugs that have been approved as HDACs inhibitors for cancer treatment, such as Vorinostat, Belinostat, Panobinostat, Chidamide and Romidepsin. Among these, Vorinostat, Benlinostat and Panobinostat are hydroxamic acid derivatives with anticancer properties. Chidamine is a ortho-aminoaniline (benzamide)derivative, and Romidepsin is a thio-ω(lactam-carboxamide) derivative, both of which possess antitumor properties.

Protein-ligand docking has always been a breakthrough milestone in computational chemistry and biology that aids in the process of structure-based drug design. Upon performing molecular docking, this in silico study hypothesizes that ¹ 4-(3-chlorophenyl)-1-(6-methoxypyridazin-3-yl) -3-methyl-5,7-dihydro-4H-pyrazolo [3,4-b]pyridin-6-one and ² 4-(4-ethoxyphenyl)-3-methyl-1-(6-morpholin-4-ylpyridazin-3-yl)-5,7-dihydro-4H-pyrazolo [3,4-b]pyridin-6-one pose high binding affinity with their target and thus appear to be potential ligands for binding and inhibiting the activity of HDAC6, representing potential drug molecules. Nonetheless, further docking processes on various platforms and in silico studies should be conducted to determine if they substantiate the aforementioned results and to generate a more effective and potent drug through structure-based drug designing approach.

References:

Afshar, M., et al. “Multiobjective/Multicriteria Optimization and Decision Support in Drug Discovery.” Comprehensive Medicinal Chemistry II, Elsevier, 11 Apr. 2007, https://www.sciencedirect.com/science/article/pii/B008045044X002753.

Aldana-Masangkay, Grace I., and Kathleen M. Sakamoto. “The Role of HDAC6 in Cancer.” BioMed Research International, Hindawi, 7 Nov. 2010, https://www.hindawi.com/journals/bmri/2011/875824/.

Alothaim, Tahiyat, et al. “HDAC6 Inhibitors Sensitize Non-Mesenchymal Triple-Negative Breast Cancer Cells to Cysteine Deprivation.” Nature News, Nature Publishing Group, 26 May 2021, https://www.nature.com/articles/s41598-021-90527-6.

Anh Tran, Andy Dong, et al. HDAC6 Deacetylation of Tubulin Modulates Dynamics of Cellular Adhesions. https://www.researchgate.net/profile/Ralph-Mazitschek/publication/51374655_HDAC6_deacetylation_of_tubulin_modulates_dynamics_of_cellular_adhesions/links/0fcfd5101566fb599e000000/HDAC6-deacetylation-of-tubulin-modulates-dynamics-of-cellular-adhesions.pdf.

Barry, Seán P., and Paul A. Townsend. “What Causes a Broken Heart-Molecular Insights into Heart Failure.” International Review of Cell and Molecular Biology, Academic Press, 25 Sept. 2010, https://www.sciencedirect.com/science/article/abs/pii/S1937644810840031.

Bazzaro, Martina, et al. “Ubiquitin Proteasome System Stress Underlies Synergistic Killing of Ovarian Cancer Cells by Bortezomib and a Novel HDAC6 Inhibitor.” American Association for Cancer Research, American Association for Cancer Research, 14 Nov. 2008, https://aacrjournals.org/clincancerres/article/14/22/7340/73341/Ubiquitin-Proteasome-System-Stress-Underlies.

Bradbury, C A, et al. “Histone Deacetylases in Acute Myeloid Leukaemia Show a Distinctive Pattern of Expression That Changes Selectively in Response to Deacetylase Inhibitors.” Nature News, Nature Publishing Group, 25 Aug. 2005, https://www.nature.com/articles/2403910.

Brownlee, Jason. “A Gentle Introduction to Dropout for Regularizing Deep Neural Networks.” Machine Learning Mastery, 6 Aug. 2019, https://machinelearningmastery.com/dropout-for-regularizing-deep-neural-networks/.

Davis, A.M. “Quantitative Structure–Activity Relationships.” Comprehensive Medicinal Chemistry III, Elsevier, 13 June 2017, https://www.sciencedirect.com/science/article/pii/B9780124095472123480.

Donaire-Arias, Ana, et al. “1h-Pyrazolo[3,4-b]Pyridines: Synthesis and Biomedical Applications.” Molecules (Basel, Switzerland), MDPI, 30 Mar. 2022, https://www.ncbi.nlm.nih.gov/pmc/articles/PMC9000541/.

Kim, Eunah, et al. “Histone and Non-Histone Targets of Dietary Deacetylase Inhibitors.” Current Topics in Medicinal Chemistry, U.S. National Library of Medicine, 2016, https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5087604/.

Knox, Tessa, et al. “Selective HDAC6 Inhibitors Improve Anti-PD-1 Immune Checkpoint Blockade Therapy by Decreasing the Anti-Inflammatory Phenotype of Macrophages and down-Regulation of Immunosuppressive Proteins in Tumor Cells.” Nature News, Nature Publishing Group, 16 Apr. 2019, https://www.nature.com/articles/s41598-019-42237-3.

Kuroki, Hiroo, et al. “Histone Deacetylase 6 Inhibition in Urothelial Cancer as a Potential New Strategy for Cancer Treatment.” Oncology Letters, Spandidos Publications, 1 Jan. 2021, https://www.spandidos-publications.com/10.3892/ol.2020.12315.

Kwon, Sunyoung, et al. “Comprehensive Ensemble in QSAR Prediction for Drug Discovery – BMC Bioinformatics.” BioMed Central, BioMed Central, 26 Oct. 2019, https://bmcbioinformatics.biomedcentral.com/articles/10.1186/s12859-019-3135-4.

Li, Ting, et al. “Histone Deacetylase 6 in Cancer – Journal of Hematology & Oncology.” BioMed Central, BioMed Central, 3 Sept. 2018, https://jhoonline.biomedcentral.com/articles/10.1186/s13045-018-0654-9.

Mandlik, Vineetha, et al. “Application of Artificial Neural Networks in Modern Drug Discovery.” Artificial Neural Network for Drug Design, Delivery and Disposition, Academic Press, 12 Feb. 2016, https://www.sciencedirect.com/science/article/pii/B9780128015599000065.

Saji, Shigehira, et al. “Significance of HDAC6 Regulation via Estrogen Signaling for Cell Motility and Prognosis in Estrogen Receptor-Positive Breast Cancer.” Nature News, Nature Publishing Group, 4 Apr. 2005, https://www.nature.com/articles/1208646.

Yoo, Jung, et al. “HDAC6-Selective Inhibitors Enhance Anticancer Effects of Paclitaxel in Ovarian Cancer Cells.” Oncology Letters, Spandidos Publications, 1 Mar. 2021, https://www.spandidos-publications.com/10.3892/ol.2021.12462.

Yu, F, and W Bender. “The Mechanism of Tamoxifen in Breast Cancer Prevention.” Breast Cancer Research : BCR, BioMed Central, 2001, https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3300587/#:~:text=Tamoxifen%20(TAM)%20is%20known%20to,activation%20and%20to%20initiate%20carcinogenesis.