25-Minute Detailed Presentation Script

Slide 1 – Title
Narration (1 min):
“Good morning/afternoon respected examiners and colleagues. I am Priti Das, and I will present
my PhD research on Computational Analysis of Post-Translational Modifications (PTMs) of
Proteins.

Proteins are central to all biological processes, and their functions are dynamically regulated by
PTMs. My research focuses on S-palmitoylation, a lipid-based PTM, and uses deep learning
methods to predict its occurrence. I aim to show not just prediction improvements, but also how
computational models can bridge molecular biology and clinical relevance.”

Slide 2 – Overview
Narration (2 min):
“Post-translational modifications increase protein diversity far beyond the genome’s coding
capacity.

 What are PTMs? Chemical alterations after protein synthesis.

 S-palmitoylation: covalent binding of palmitic acid to cysteine.
 Unique features: reversible, dynamic, and regulates protein localization, stability, and
interactions.

Why is it important? Abnormal palmitoylation has been linked to:

 Cancer (e.g., tumor cell signaling),

 Alzheimer’s and Parkinson’s (synaptic dysfunction),
 Cardiovascular diseases,
 Major depressive disorder.

Despite its importance, computational tools for palmitoylation are limited. My solution: a
Convolutional Neural Network (CNN) model tailored for site prediction.”

Slide 3 – Research Gap

Narration (2–3 min):
“There are several gaps in existing PTM research:

1. Data coverage – Positive palmitoylation datasets exist, but validated negative datasets
are very scarce. This leads to class imbalance.
2. Model limitations – Most tools are designed for a single PTM. A generic prediction
framework for multiple PTMs is missing.
3. Crosstalk is ignored – PTMs act combinatorially, influencing each other. Current
computational approaches rarely model this.
4. Network integration is weak – Protein–protein interaction networks are modeled as
static, binary events, while PTMs actually regulate interaction dynamics.
5. Disease mapping is incomplete – Rare PTMs like palmitoylation are underexplored in
clinical contexts, leaving gaps in understanding their role in disease.

Thus, my research addresses: better datasets, optimized deep learning, sex-specific analyses, and
future integration into network and disease frameworks.”

Slide 4 – Scope of Work

Narration (2 min):
“The scope of my work includes:

 Prediction of PTM sites using CNN models.

 Building a multi-PTM model that can generalize beyond palmitoylation.
 Curating reliable negative data to strengthen in-silico predictions.
 Integrating PTMs into interaction networks, moving beyond linear site prediction.
 Exploring disease associations, especially how aberrant PTMs contribute to pathological
states.

This research is not only computational but also designed to complement experimental biology
and translational medicine.”

Slide 5 – Motivation
Narration (2 min):
“My motivation comes from the fact that PTMs are biological regulators and potential
biomarkers.

1. Unified platforms: At present, each PTM has its own scattered tools. A multi-PTM,
multi-omics platform with standardized annotations is needed.
 2. Interpretability: Biologists want interpretable models — not just predictions but also
mechanistic insights.
3. Time-series PTM data: Diseases progress dynamically. Modeling temporal PTM
changes can reveal new biomarkers.
4. High-throughput palmitoyl-proteomics: Still underdeveloped compared to
phosphorylation profiling.
5. Clinical biomarkers: PTM-based biomarkers could support early diagnosis, prognosis,
and personalized treatment.

Thus, my CNN framework is one step toward bridging ML predictions and biomedical
applications.”

Slide 6 – Methodology
Narration (2 min):
“My methodology combines data curation, feature extraction, feature selection, and CNN
model training:

1. Dataset preparation: Using PANIMoni dataset (PXD025286).

2. Feature extraction: 566 features from AAIndex.
3. Windowing: 39 residues per sequence window, centered on cysteine.
4. Feature selection: Three strategies — K-Best, Genetic Algorithm (GA), and Hybrid
(UN).
5. CNN architecture: Trained on optimized features, evaluated using AUC, precision,
recall, and F1-score.

This systematic pipeline ensures robust and biologically meaningful predictions.”

Slide 7 – Dataset
Narration (2 min):
“The PANIMoni dataset provides high-quality mass spectrometry–validated palmitoylation sites.

 Female dataset: 1036 proteins, 1773 positive, 8934 negative.

 Male dataset: 1077 proteins, 1870 positive, 9279 negative.
 Combined dataset: 1180 proteins, 2083 positive, 10,403 negative.

The novelty is in using sex-specific data. Many computational studies ignore biological sex, but
evidence shows male and female proteomes can differ significantly in modification patterns.”
Slide 8 – Feature Extraction
Narration (2 min):
“Feature extraction is critical because raw protein sequences cannot be directly fed into CNNs.

 Using AAIndex, I derived 566 physicochemical features.

 Features include hydrophobicity, polarity, flexibility, charge distribution, etc.
 Each cysteine-centered window of 39 residues was converted into a feature vector.

Then, feature selection:

 K-Best ranks features statistically.

 GA searches for optimal combinations.
 Hybrid (UN) integrates both methods.

This reduced redundancy and improved model interpretability.”

Slide 9 – Results
Narration (3 min):
“My CNN model achieved strong results:

 Female dataset: K-Best features performed best, with ~79% AUC.

 Male dataset: Hybrid features gave ~76% AUC.
 Combined dataset: Hybrid features gave ~78% AUC.

Key findings:

 CNN outperformed RFCM-PALM and protein language models.

 Performance was balanced across precision, recall, and F1 — not biased toward one
class.
 The sex-specific difference suggests biological variation in palmitoylation regulation,
opening new research directions.

Thus, the CNN approach is both computationally strong and biologically relevant.”

Slide 10 – Research Gap (Expanded)

Narration (2 min):
“While my model addresses several challenges, broader gaps remain:
  Databases like dbPTM are growing but still fragmented.
 Phosphorylation has dozens of predictors, but palmitoylation predictors are still rare.
 Most AI tools focus on disease associations rather than mechanistic predictions.
 Patient-derived PTM datasets are missing — most are from cell lines.

These limitations highlight the need for continued integration of PTM biology with AI-driven
computational methods.”

Slide 11 – Research Plan

Narration (2–3 min):
“My future work includes:

1. Multi-PTM prediction: Extending the CNN pipeline to phosphorylation, acetylation,

ubiquitination, etc.
2. PTM-driven PPI networks: Incorporating modification sites into network rewiring,
moving beyond static PPI maps.
3. PTM–disease integration: Using omics data to build systematic PTM–disease
associations.
4. Causal AI models: Going from correlation to causation — predicting how PTMs interact
and influence signaling pathways.
5. Web-based tool: To make the model accessible to the broader community.

This roadmap takes the work from basic computational prediction to translational
applications.”

Slide 12 – Conclusion
Narration (2 min):
“To conclude:

 I developed a CNN-based prediction model for S-palmitoylation sites.

 The model outperforms existing state-of-the-art methods in both sex-specific and
combined datasets.
 This research contributes to filling critical gaps in PTM prediction, dataset development,
and interpretability.
 The approach is scalable to multiple PTMs and has potential applications in biomarker
discovery and therapeutic targeting.

Thus, this work represents an important step toward integrating computational predictions with
experimental and clinical research.”
Slide 13 – Acknowledgement
Narration (1 min):
“I would like to thank my supervisors — Dr. Subhadip Basu (Jadavpur University) and Dr.
Soumyendu Sekhar Bandyopadhyay (IEM Kolkata) — for their invaluable guidance. I also
thank my colleagues and collaborators for their constant support.”

10. What is a recurrent neural network (RNN)?

 Suitable for sequential/time series data. Variants include LSTM and GRU.

🤖 AI Algorithms and Concepts

11. What are heuristic search techniques in AI?

 A*: Greedy, Hill Climbing, Best-First Search.

12. What is knowledge representation in AI?

 Ontologies, Semantic Networks, Frames, Rules, and Logic.

13. What is the difference between strong AI and weak AI?

 Strong AI = General Intelligence; Weak AI = Task-specific intelligence.

14. What is reinforcement learning?

 Learning through interaction and rewards. Concepts: agent, environment, policy, Q-

learning.

📚 Ethics & Challenges

15. What are the ethical concerns in AI?

 Bias, Privacy, Job Displacement, Autonomous Weapons, Transparency.

16. What is explainable AI (XAI)?

 Making AI decisions interpretable and transparent to humans.

17. What is the role of AI in healthcare/education/finance/etc.?

 Domain-specific applications and impacts.

🧪 Practical & Research-Level Questions

18. How does backpropagation work?

 Gradient descent-based optimization in neural networks.

19. What is transfer learning?

 Using a pre-trained model on a new but related task.

20. What are common performance metrics in ML/AI?

 Accuracy, Precision, Recall, F1-Score, AUC-ROC.

1. What is Artificial Intelligence?

 Define AI and its types: Narrow AI, General AI, and Super AI.

2. What are the main areas of AI?

 Machine Learning, Natural Language Processing, Robotics, Computer Vision, Expert

Systems, Planning.

3. What is the difference between AI, ML, and Deep Learning?

 Definitions, overlaps, and hierarchical relationships among them.

4. What are the main types of machine learning?

 Supervised, Unsupervised, Semi-supervised, and Reinforcement Learning.

5. What is the Turing Test?

 A test to determine whether a machine can exhibit human-like intelligence.

📊 Machine Learning & Deep Learning

6. What is overfitting and underfitting?

 Definitions, causes, and techniques to avoid them (e.g., regularization, cross-validation).

7. What are activation functions?

 Examples: ReLU, Sigmoid, Tanh, Softmax.

8. What is a neural network? How does it work?

 Layers, weights, biases, forward propagation, and backpropagation.

9. What is a convolutional neural network (CNN)?

 Used for image processing and sequential data (like in your PTM project).

10. What is a recurrent neural network (RNN)?

 Suitable for sequential/time series data. Variants include LSTM and GRU.

🤖 AI Algorithms and Concepts

11. What are heuristic search techniques in AI?

 A*: Greedy, Hill Climbing, Best-First Search.

12. What is knowledge representation in AI?

 Ontologies, Semantic Networks, Frames, Rules, and Logic.

13. What is the difference between strong AI and weak AI?

 Strong AI = General Intelligence; Weak AI = Task-specific intelligence.

14. What is reinforcement learning?

 Learning through interaction and rewards. Concepts: agent, environment, policy, Q-

learning.

📚 Ethics & Challenges

15. What are the ethical concerns in AI?

 Bias, Privacy, Job Displacement, Autonomous Weapons, Transparency.

16. What is explainable AI (XAI)?

 Making AI decisions interpretable and transparent to humans.

17. What is the role of AI in healthcare/education/finance/etc.?

 Domain-specific applications and impacts.

🧪 Practical & Research-Level Questions

18. How does backpropagation work?

 Gradient descent-based optimization in neural networks.

19. What is transfer learning?

 Using a pre-trained model on a new but related task.

20. What are common performance metrics in ML/AI?

 Accuracy, Precision, Recall, F1-Score, AUC-ROC.

. What is Machine Learning?

  Define ML and explain its goal: systems that learn from data without being explicitly
programmed.

2. Types of Machine Learning?

 Supervised Learning (with labeled data)

 Unsupervised Learning (no labels)
 Semi-supervised Learning
 Reinforcement Learning

3. What is the difference between AI, ML, and Deep Learning?

4. Explain Overfitting and Underfitting.

 Definitions with causes, consequences, and prevention methods.

5. What are Bias and Variance?

 Trade-off explanation, and how it relates to overfitting/underfitting.

🔧 Algorithms and Models

6. What is Linear Regression?

 Assumes a linear relationship between input and output.

7. What is Logistic Regression?

 Used for binary classification problems; outputs probability values.

8. What is Decision Tree?

 A tree-like model of decisions with conditions at each node.

9. What is Random Forest?

 An ensemble of decision trees to improve prediction and reduce overfitting.

10. What is Support Vector Machine (SVM)?

 Finds the optimal hyperplane for separating classes.

11. What is K-Nearest Neighbors (KNN)?

 Instance-based learner that classifies based on the majority vote of neighbors.

12. What is Naïve Bayes?

 A probabilistic classifier based on Bayes' theorem with independence assumptions.

📈 Model Evaluation
13. What are common evaluation metrics in ML?

 Accuracy, Precision, Recall, F1 Score, AUC-ROC, Confusion Matrix.

14. What is Cross-Validation?

 A resampling technique to assess model performance (e.g., K-Fold CV).

15. What is the ROC Curve and AUC?

 Measures how well the model separates classes at various thresholds.

🧮 Optimization and Training

16. What is Gradient Descent?

 Optimization algorithm to minimize the loss function.

17. What is the difference between Batch, Stochastic, and Mini-Batch Gradient
Descent?

18. What is the Cost or Loss Function?

 Quantifies the error between predicted and actual values.

🧬 Data and Preprocessing

19. What is Feature Engineering?

 Creating new input features from raw data to improve model performance.

20. What is Feature Selection and Dimensionality Reduction?

 Techniques like PCA, LDA, and mutual information.

21. What is Data Normalization and Standardization?

 Techniques to scale features for better model performance.

🤖 Advanced / Practical Questions

22. What is Ensemble Learning?

 Combining multiple models (e.g., Bagging, Boosting).

23. What is the difference between Bagging and Boosting?

 Bagging: parallel models (e.g., Random Forest); Boosting: sequential (e.g., AdaBoost,
XGBoost).

24. What is Hyperparameter Tuning?

 Adjusting model settings (e.g., learning rate, tree depth) using GridSearchCV,
RandomSearch, or Bayesian optimization.

25. What is a Confusion Matrix?

 A table to visualize TP, FP, TN, FN for classification problems.

Slide 1 – Title
"Welcome to this presentation on S-Palmitoylation Site Prediction of Mouse Proteins using
Convolutional Neural Networks.
This work is jointly authored by Priti Das, Soumyendu Sekhar Bandyopadhyay, and Subhadip
Basu."

Slide 2 – Abstract
"S-palmitoylation is a reversible post-translational modification in which palmitic acid modifies
cysteine residues.
It plays a crucial role in many biological processes and is associated with several human
diseases.
In this study, we propose a convolutional neural network-based model to predict S-
palmitoylation sites in mouse proteins.
By selecting optimized physicochemical features, our method achieves approximately seventy-
eight percent AUC, and shows a fourteen percent improvement over state-of-the-art
approaches."

Slide 3 – Introduction
"Post-translational modifications enhance protein diversity and regulate biological functions.
S-palmitoylation involves the covalent binding of palmitic acid to cysteine residues.
This modification influences protein localization, stability, and activity.
It has been linked to several diseases, including major depressive disorder, cancer, Alzheimer’s,
Parkinson’s, and cardiovascular conditions.
Therefore, accurate prediction of S-palmitoylation sites is of great biological and clinical
importance."

Slide 4 – Dataset
"The dataset used in this study was generated using PANIMoni mass spectrometry and is
publicly available under the Proteome-Xchange Consortium, with ID PXD025286.
It includes three categories: female, male, and sex-independent datasets.
For example, the female dataset contains one thousand and thirty-six proteins with one thousand
seven hundred and seventy-three positive sites and over eight thousand nine hundred negative
sites.
Similar distributions are available for the male and combined datasets."

Slide 5 – Feature Extraction and Selection

"We extracted five hundred and sixty-six physicochemical features from the AAIndex database.
Each protein sequence was divided into a window of thirty-nine amino acids, with cysteine at the
center.
From these windows, position-specific amino acid propensities were calculated.
For feature optimization, we applied K-Best selection, genetic algorithms, and a combined
approach.
The optimized features were then used as inputs for CNN training."

Slide 6 – CNN Architecture

"Our proposed model is a one-dimensional convolutional neural network.
The architecture consists of convolutional layers, dropout layers, a flatten layer, fully connected
dense layers, and an output layer.
Activation functions such as ReLU and softmax are applied.
The network was trained and evaluated using five-fold cross-validation to ensure robustness."

Slide 7 – Results
"The results show that the K-Best feature set performs best for the female dataset, achieving
around seventy-nine percent AUC.
For the male and sex-independent datasets, the combined feature set performs better, with AUC
values of seventy-six and seventy-eight percent respectively.
Overall, the CNN model demonstrated balanced precision, recall, and F1 scores across
datasets."

Slide 8 – Comparison with State-of-the-Art

"We compared our proposed method against two state-of-the-art models: RFCM-PALM and a
protein language model proposed by Chakraborty and colleagues.
Our CNN significantly outperforms both methods on female and male datasets, with
improvements of fourteen to sixteen percent in AUC.
Although Chakraborty’s method performs slightly better for sex-independent data, our approach
achieves competitive accuracy overall."

Slide 9 – Conclusion and Future Work

"In conclusion, the proposed CNN-based approach effectively predicts S-palmitoylation sites in
mouse proteins, using optimized physicochemical features.
It outperforms existing methods on sex-dependent datasets.
In the future, this method can be extended to other post-translational modifications, such as
phosphorylation and acetylation, and applied to larger, publicly available datasets.
Thank you for your attention."
  Conclusion:
The proposed CNN-based framework provides an accurate and efficient method for
predicting S-palmitoylation sites. It is scalable to other PTMs and has potential
applications in guiding experimental biologists to prioritize candidates for laboratory
validation.
 Acknowledgement:
This work was conducted under the supervision of Dr. Subhadip Basu (Jadavpur
University) and Dr. Soumyendu Sekhar Bandyopadhyay (Institute of Engineering
and Management, Kolkata).

Why did you choose S-palmitoylation instead of other PTMs like phosphorylation or acetylation? Can
you explain in simple terms why S-palmitoylation is important for disease research? What biological
validation supports your computational findings? answer

ChatGPT said:

Great — those are exactly the kind of Q&A preparation questions you’ll face in your defense
or presentation.
Here are concise, confident answers you can give:

❓ Q1: Why did you choose S-palmitoylation instead of other PTMs like
phosphorylation or acetylation?

✅ Answer:
"We focused on S-palmitoylation because, unlike phosphorylation or acetylation, it is a
reversible lipid modification that directly regulates protein trafficking, localization, and
stability.
Although phosphorylation is well-studied with large datasets available, S-palmitoylation
remains relatively underexplored despite being equally important for neural signaling and
disease pathways.
By targeting S-palmitoylation, we aimed to address this knowledge gap and provide a
computational framework that complements experimental biology."

❓ Q2: Can you explain in simple terms why S-palmitoylation is important for
disease research?

✅ Answer:
"In simple terms, S-palmitoylation works like an on–off switch for proteins in the brain and
heart.
It controls whether a protein stays at the right place in the cell or gets removed.
When this modification is abnormal, it can disrupt signaling pathways and has been linked to
diseases such as Alzheimer’s, Parkinson’s, cancer, and even psychiatric disorders like
depression.
So, understanding and predicting S-palmitoylation sites may help us identify potential drug
targets and improve disease treatments."

❓ Q3: What biological validation supports your computational findings?

✅ Answer:
"Our computational predictions were trained and tested using experimentally validated datasets
generated by mass spectrometry, specifically PANIMoni-based datasets from wild-type and
knockout mice.
We validated our CNN model’s performance using cross-validation and compared it with state-
of-the-art tools.
While we have not yet conducted new wet-lab experiments, our results show consistent
improvement over benchmark methods, which indicates that the computational predictions are
biologically meaningful and can be used to guide future laboratory validation."

👉 Would you like me to also prepare short bullet-point answers (one-liners) for quick recall
during live Q&A, so you don’t have to memorize long responses?

You said:
Dataset-Related Questions How was the dataset curated and why did you choose PXD025286? How did
you handle class imbalance between positive and negative cysteine sites? Did you perform any
preprocessing, like removing redundant sequences or aligning proteins? Are the results specific to
mouse proteins, or can the model generalize to human data? answer

ChatGPT said:

Perfect 👍 These are the dataset-related Q&A you’ll likely face. Here are strong, concise
answers you can give:

❓ Q1: How was the dataset curated and why did you choose PXD025286?

✅ Answer:
"We used dataset PXD025286 from the ProteomeXchange Consortium because it was generated
with PANIMoni mass spectrometry, which provides high-confidence identification of S-
palmitoylation sites.
It contains wild-type and knockout mouse brain proteins with both positive (palmitoylated) and
negative (non-palmitoylated) cysteine sites.
We chose this dataset since it is publicly available, experimentally validated, and specifically
designed for studying palmitoylation in synaptic proteins."

❓ Q2: How did you handle class imbalance between positive and negative cysteine
sites?

✅ Answer:
"The dataset had many more negative than positive samples.
To address this, we used a bootstrapping approach: we divided the data into balanced subsets
where positive samples remained fixed, and negative samples were randomly resampled to
maintain a 1:1 ratio.
This ensured that the model did not get biased toward the majority negative class and improved
prediction reliability."

❓ Q3: Did you perform any preprocessing, like removing redundant sequences or
aligning proteins?

✅ Answer:
"Yes. We filtered out ambiguous cysteine sites that had conflicting modifications, such as
carbamidomethyl and N-ethylmaleimide.
We also removed redundant or duplicate entries to avoid bias.
For modeling, protein sequences were converted into fixed-length windows of 39 amino acids
with cysteine at the center, ensuring consistent input representation."

❓ Q4: Are the results specific to mouse proteins, or can the model generalize to
human data?

✅ Answer:
"The model was trained on mouse datasets, but since the physicochemical properties of amino
acids are universal, the method can generalize to other species, including humans.
Of course, accuracy depends on the availability of experimentally validated training data.
Future work will involve testing on human datasets to evaluate generalizability more
rigorously."
👉 Would you like me to also prepare Model Architecture & Performance Q&A (like why
CNN instead of LSTM/Transformer, overfitting concerns, why K-Best vs GA features) so you’re
ready for the technical panel?

You said:
Feature Engineering & Selection Why did you select a window size of 39 amino acids? Did you try other
lengths? Among the 566 AAIndex features, which types contributed most to prediction accuracy? Can
you explain the difference between K-Best and Genetic Algorithm feature selection? Why does the
combined (UN) feature set perform better for some datasets but not others?

ChatGPT said:

Excellent 👍 These are exactly the feature engineering & selection Q&A reviewers or
examiners love to ask. Here are strong, confident answers you can use:

❓ Q1: Why did you select a window size of 39 amino acids? Did you try other
lengths?

✅ Answer:
"We selected a window size of 39 because it provides 19 amino acids upstream and 19
downstream of the central cysteine, capturing both local and extended sequence context.
We experimented with smaller windows such as 21 and 31, but they missed longer-range
interactions.
Larger windows like 51 increased noise without significant improvement.
Empirically, 39 gave the best balance between capturing biologically relevant context and
maintaining model performance."

❓ Q2: Among the 566 AAIndex features, which types contributed most to prediction
accuracy?

✅ Answer:
"The most informative features came from hydrophobicity scales, flexibility indices, and
secondary structure propensities.
These are biologically meaningful because S-palmitoylation typically occurs in hydrophobic
membrane-associated regions and depends on local structural context.
In contrast, features unrelated to structure or membrane interaction contributed less to
predictive power."
❓ Q3: Can you explain the difference between K-Best and Genetic Algorithm
feature selection?

✅ Answer:
"K-Best is a filter-based method that ranks features individually based on statistical measures
such as their correlation with class labels.
It is fast and effective but ignores interactions between features.
The Genetic Algorithm, on the other hand, is an evolutionary search method that evaluates
subsets of features as a whole using AUC as the fitness score.
It can capture feature interactions but is computationally more expensive.
In our work, we combined both to balance efficiency and accuracy."

❓ Q4: Why does the combined (UN) feature set perform better for some datasets but
not others?

✅ Answer:
"The combined feature set, or union of K-Best and GA features, increases diversity and captures
complementary information.
This works well for male and sex-independent datasets, which are larger and more
heterogeneous, so additional features help capture subtle patterns.
However, for the female dataset, K-Best alone performed better because it reduced redundancy
and avoided overfitting in a smaller, more homogeneous dataset.
Thus, the optimal feature selection strategy is dataset-dependent."

👉 Do you want me to now also prepare the Model Architecture & Performance Q&A (e.g.,
Why CNN instead of LSTM or Transformers?, How did you prevent overfitting?, Why different
results across datasets?) so you’re fully covered for technical questions?

You said:
Model Architecture & Performance Why did you choose a 1D CNN instead of LSTM, Bi-LSTM, or
Transformers? How did you optimize hyperparameters like kernel size, dropout, and learning rate? Did
you consider overfitting, and how did you prevent it? How does the performance vary across folds in
cross-validation? Why does K-Best perform better for female data, while UN performs better for male
and combined?

ChatGPT said:

Perfect 👍 — these are the core model architecture & performance Q&A you’ll almost
certainly get. Here are clear, confident answers:
❓ Q1: Why did you choose a 1D CNN instead of LSTM, Bi-LSTM, or
Transformers?

✅ Answer:
"We chose a 1D CNN because our task is to detect local sequence motifs around cysteine
residues.
CNNs are very efficient at capturing spatial patterns in fixed-length windows, and they train
faster with fewer parameters compared to LSTM or Transformer models.
While LSTMs and Transformers excel at long-range dependencies, our analysis showed that
most useful signals for S-palmitoylation prediction lie in the local neighborhood of the cysteine.
Therefore, CNN provided the best trade-off between accuracy, interpretability, and
computational efficiency."

❓ Q2: How did you optimize hyperparameters like kernel size, dropout, and
learning rate?

✅ Answer:
"We used a grid search approach on validation folds.
Kernel sizes from 3 to 11 were tested, with 5 performing best for capturing short sequence
motifs.
Dropout rates from 0.2 to 0.5 were evaluated, and 0.3 gave the best balance between
generalization and information retention.
For learning rate, we tried values between 1e-2 and 1e-5; the optimal was 1e-3 with Adam
optimizer.
These hyperparameters were chosen based on maximizing AUC in cross-validation."

❓ Q3: Did you consider overfitting, and how did you prevent it?

✅ Answer:
"Yes, overfitting is a common risk in deep learning, especially with imbalanced datasets.
We addressed this by applying dropout layers, early stopping during training, and bootstrapped
data balancing.
We also used five-fold cross-validation, which confirmed that the model generalized well across
different subsets.
Performance consistency across folds suggested that overfitting was effectively controlled."

❓ Q4: How does the performance vary across folds in cross-validation?

✅ Answer:
"The performance variation was minimal, indicating stable generalization.
For example, in the female dataset with K-Best features, the AUC ranged from about 0.77 to
0.80 across folds.
For the male and combined datasets with UN features, the variation was within ±0.02.
This stability reinforces the robustness of our model."

❓ Q5: Why does K-Best perform better for female data, while UN performs better
for male and combined?

✅ Answer:
"The female dataset is relatively smaller and more homogeneous, so simpler feature sets like K-
Best are sufficient and reduce redundancy.
In contrast, the male and sex-independent datasets are larger and more diverse.
Here, combining K-Best with GA-selected features (UN) helps capture subtle and
complementary signals that improve prediction accuracy.
So, the best feature strategy depends on the dataset size and heterogeneity."

👉 Would you like me to also prepare Comparison & Future Scope Q&A (e.g., Why does your
model outperform RFCM-PALM but not Chakraborty et al. on sex-independent data?, How
significant is the improvement in real biology?, How can this work be extended?) so you’re
covered for discussion-type questions too?

You said:
Comparison with State-of-the-Art Why does your model outperform existing methods for male/female
datasets but not for sex-independent? How significant is the ~14–16% improvement in AUC in practical
biological terms? What are the limitations of RFCM-PALM and the Protein Language Model compared to
CNN?

ChatGPT said:

Great 👍 these are exactly the comparison & benchmarking Q&A you’ll get from reviewers.
Here are confident answers:

❓ Q1: Why does your model outperform existing methods for male/female datasets
but not for sex-independent?

✅ Answer:
"Our model outperforms existing methods on male and female datasets because CNNs are
effective at detecting local sequence motifs in relatively homogeneous data.
However, the sex-independent dataset is more heterogeneous since it combines male and female
data.
The protein language model used by Chakraborty et al. is pretrained on very large datasets and
is better at capturing long-range and diverse patterns.
Therefore, while CNN excels in focused, sex-specific contexts, the language model has an
advantage in handling more complex mixed data."

❓ Q2: How significant is the ~14–16% improvement in AUC in practical biological

terms?

✅ Answer:
"An improvement of 14 to 16 percent in AUC is highly significant in biological prediction tasks.
It means the model can distinguish positive palmitoylation sites from negatives much more
reliably.
In practical terms, this reduces false positives and false negatives, which is critical when
prioritizing candidates for experimental validation.
Since biological experiments are expensive and time-consuming, even a modest improvement
translates into a major impact by saving time, cost, and resources in the lab."

❓ Q3: What are the limitations of RFCM-PALM and the Protein Language Model
compared to CNN?

✅ Answer:
"RFCM-PALM, which is based on random forests, performs well but cannot fully capture
sequential patterns and spatial dependencies in amino acid windows.
It treats features more independently, which limits its ability to model local motifs important for
palmitoylation.
The protein language model is very powerful but computationally expensive and less
interpretable, requiring significant resources to train and deploy.
In contrast, our 1D CNN provides a balance: it captures local motifs more effectively than
RFCM-PALM, is more efficient than transformer-based language models, and is relatively
easier to interpret."

👉 Would you like me to also draft Conclusion & Future Scope Q&A (e.g., Can this model be
extended to other PTMs?, How will you validate predictions experimentally?, What are your
future research directions?) so you’ll be fully ready for the final discussion?

You said:
Future Scope & Applications Can this model be applied to predict other PTMs beyond S-palmitoylation?
How could experimental biologists use your tool in real lab work? Do you plan to develop a web server
or software for easy use by the research community? What are the computational limitations—can this
method scale to much larger datasets?

ChatGPT said:

Excellent 👌 these are exactly the closing Q&A that let you show vision and practical relevance.
Here are polished answers you can use:

❓ Q1: Can this model be applied to predict other PTMs beyond S-palmitoylation?

✅ Answer:
"Yes, the framework is generalizable. The CNN architecture and feature selection pipeline are
not specific to S-palmitoylation; they can be adapted to other PTMs such as phosphorylation,
acetylation, or ubiquitination.
The only requirement is a sufficient amount of experimentally validated training data.
So, with appropriate datasets, the same approach can be extended to a wide range of PTMs."

❓ Q2: How could experimental biologists use your tool in real lab work?

✅ Answer:
"Experimental biologists could use this tool to prioritize candidate cysteine sites for validation.
Instead of testing hundreds of potential modification sites, they can focus on the high-confidence
predictions generated by our model.
This reduces the number of wet-lab experiments needed, saving time, cost, and resources while
increasing the chances of successful validation."

❓ Q3: Do you plan to develop a web server or software for easy use by the research
community?

✅ Answer:
"Yes, one of our future goals is to deploy the model as a web-based prediction server or
lightweight software tool.
This would allow biologists to input protein sequences and receive predicted palmitoylation sites
instantly, without requiring coding or deep learning expertise.
Such accessibility would greatly increase the impact of our work within the broader scientific
community."
❓ Q4: What are the computational limitations—can this method scale to much
larger datasets?

✅ Answer:
"Currently, our CNN model is efficient enough to handle thousands of protein sequences, as
shown in this study.
For much larger datasets, such as full proteomes across multiple species, the main challenges
would be training time and memory requirements.
However, with GPU acceleration and distributed training, CNN models are well-suited for
scaling.
Thus, computational limitations are manageable, and the method can be adapted to big-data
biological applications.

Updated:

Slide 1: Title
“Good [morning/afternoon] everyone.
My name is Priti Das and today I will present my PhD work on Computational Analysis of Post-
Translational Modifications of Proteins.
This research is carried out under the supervision of Dr. Subhadip Basu and Dr. Soumyendu
Sekhar Bandyopadhyay.
The focus of my work is on predicting S-palmitoylation sites in proteins using a deep learning-
based approach.”

Slide 2: Contents
“Here is the outline of my presentation.
I will begin with an overview and motivation, followed by the research plan.
Then I will introduce the dataset, describe the feature extraction process, and explain the CNN
architecture.
After presenting the results, I will compare them with existing methods, discuss future work, and
finally conclude with acknowledgements and references.”
Slide 3: Overview
“Post-translational modifications, or PTMs, are chemical changes to proteins after translation.
They play a crucial role in regulating protein function, localization, and stability.
Among them, S-palmitoylation is a reversible modification, where palmitic acid covalently
attaches to cysteine residues.
This modification is linked to diseases such as cancer, Alzheimer’s, Parkinson’s, and
cardiovascular conditions.
Given its biological and clinical importance, I propose a convolutional neural network-based
model to predict S-palmitoylation sites in proteins.”

Slide 4: Research Gap

“Despite progress in PTM research, there are still several gaps.
Databases are fragmented across PTMs and species.
In phosphorylation, prediction tools give inconsistent results and longitudinal datasets are rare.
For palmitoylation, large-scale maps are lacking, and predictors are fewer compared to
phosphorylation.
Moreover, most AI tools focus on disease association, not mechanistic predictions.
Finally, patient-derived PTM datasets are limited, restricting clinical translation.
My work addresses some of these gaps by building a robust predictive model for
palmitoylation.”

Slide 5: Overall Research Plan

“The overall plan of my research involves three main steps.
First, developing a deep learning model for PTM site prediction.
Second, analyzing protein–protein interactions in the context of PTMs.
And third, linking PTMs with disease associations.
My CNN model with optimized features improves palmitoylation prediction and outperforms
state-of-the-art methods.
This approach can also be extended to other PTMs such as phosphorylation or acetylation.”

Slide 6: Motivation / Scope of Work

“The motivation of this work is to re-design computational systems for PTM research.
We need unified multi-omics platforms with standardized annotations for machine learning.
We need interpretable models that map kinase–substrate or PTM-specific interactions.
There is also a demand for time-series PTM datasets in diseases such as cancer or
neurodegeneration.
Finally, PTM-based biomarkers have great potential in early disease detection and treatment
response.
My work aims to contribute towards these goals by combining deep learning with optimized
feature selection.”

Slide 7: Methodology Proposal

“To summarize the methodology:
Post-translational modifications enhance protein diversity and function.
Specifically, I focus on S-palmitoylation, which regulates protein stability and localization.
Since it is linked to major diseases like depression, cancer, and Alzheimer’s, predicting
palmitoylation sites can provide important insights into disease biology.
This sets the foundation for my dataset preparation and feature extraction, which I will describe
next.”

Slide 8: Dataset
“The dataset used in my work is the PANIMoni mass spectrometry dataset, PXD025286.
It is divided into three categories: Female, Male, and Combined.
For example, the Female dataset contains 1036 proteins with 1773 positive and 8934 negative
sites.
The Male dataset has 1077 proteins with 1870 positive and 9279 negative sites.
The Combined dataset includes 1180 proteins, with 2083 positive and over 10,000 negative sites.
This dataset provides a strong foundation for training and testing my CNN model.”

Slide 9: Feature Extraction & Selection

“I extracted 566 physicochemical features from the AAIndex database.
Each sample uses a window size of 39 residues—19 before the cysteine, the cysteine itself, and
19 after.
Position-specific amino acid propensities were also calculated.
To avoid overfitting, I applied feature selection using three strategies: K-Best, Genetic
Algorithm, and a combined method.
These optimized features were then used for CNN training.”
Slide 10: Results
“The CNN model produced promising results.
For the Female dataset, the K-Best feature selection gave the best performance with around 79%
AUC.
For the Male dataset, the combined method performed best with 76% AUC.
For the sex-independent dataset, the combined method gave 78% AUC.
Precision, recall, and F1-scores were balanced across datasets.
Importantly, my CNN model outperformed the state-of-the-art methods like RFCM-PALM and
protein language models.”

Slide 11: Comparison with State-of-the-Art

“This slide shows the comparison.
While RFCM-PALM and protein language models performed well, my CNN with optimized
features consistently outperformed them.
This demonstrates the novelty and effectiveness of my approach, particularly in handling sex-
specific and combined datasets.”

Slide 12: Future Work

“In the future, I plan to extend this method to other PTMs such as phosphorylation and
acetylation.
I also want to apply it to patient-derived data, which will increase its translational value.
Another direction is building a unified multi-PTM platform and developing a webserver or
standalone tool for easy use by researchers.
Ultimately, this work aims to support the discovery of PTM-based biomarkers and therapeutic
targets.”

Slide 13: Conclusion

“To conclude, my CNN model with optimized features improves S-palmitoylation prediction
significantly.
It outperforms existing models and achieves balanced performance across male, female, and
combined datasets.
The approach is scalable and can be applied to other PTMs.
Overall, this research provides both computational and biological insights into PTM prediction.”
Slide 14: Acknowledgement
“I would like to sincerely thank my supervisors, Dr. Subhadip Basu and Dr. Soumyendu Sekhar
Bandyopadhyay, for their invaluable guidance and support throughout this work.
I also acknowledge my institute for providing the necessary facilities.”

Slide 15: References

“Here is a list of references that supported my research.
These include studies on PTMs, computational prediction models, and methodological
foundations.
For time constraints, I will not go through them in detail, but I am happy to discuss any reference
during questions.”

Slide 16: Thank You

“Thank you all for your attention.
I will be happy to take any questions.”
Slide 1 – Title

Narration:
“Good [morning/afternoon], everyone. I am Priti Das, and today I will present my doctoral
research titled Computational Analysis of Post-Translational Modifications of Proteins.

This research focuses on understanding the role of S-palmitoylation, a lipid modification of

proteins, and developing deep learning-based prediction models to identify modification sites.
My aim is not only to improve computational prediction but also to lay the groundwork for
integrating PTM information into biological and clinical research.”

Slide 2 – Overview

Narration:
“Post-translational modifications, or PTMs, are chemical modifications that proteins undergo
after translation. They expand protein diversity far beyond the coding potential of genes.

Among the many PTMs, S-palmitoylation is unique because it is reversible and highly dynamic.
It involves the covalent attachment of palmitic acid to cysteine residues, influencing where
proteins are located in the cell, how stable they are, and how they interact with other proteins.

Aberrant palmitoylation has been implicated in neurological disorders, cancers,

cardiovascular diseases, and even depression.

Therefore, computational prediction of palmitoylation sites can accelerate functional studies,

biomarker discovery, and drug development. To achieve this, I propose a convolutional
neural network (CNN)-based model trained on curated protein sequence datasets.”
Slide 3 – Research Gap

Narration:
“Despite the importance of palmitoylation, there are four major research gaps that motivated
this study:

1. Data imbalance and negative samples – While experimentally validated positive

palmitoylation sites exist, negative datasets are scarce. Without good negative data,
machine learning models suffer from bias.
2. Lack of a generalizable prediction framework – Current tools often focus on one PTM
at a time. There is no generic model that can handle multiple PTMs across different
species.
3. Underexplored PTM crosstalk – PTMs do not act in isolation. Crosstalk between
phosphorylation, ubiquitination, palmitoylation, and others regulates protein networks,
but most computational models treat them independently.
4. Incomplete disease mapping – Current PTM datasets are biased toward
phosphorylation, leaving other modifications like palmitoylation poorly studied in disease
contexts. Moreover, most protein–protein interaction (PPI) databases ignore PTM-driven
rewiring of interactions.

These gaps highlight the need for a robust, data-driven, and integrative computational
approach.”

Slide 4 – Scope of Work

Narration:
“The scope of my research can be divided into four objectives:

 PTM site prediction – Build models to predict modification sites in protein sequences.
 Multi-PTM modeling – Extend beyond palmitoylation to phosphorylation, acetylation,
and others.
 Dataset curation – Develop high-quality negative datasets to improve in-silico
predictions.
 Integration with biological networks – Incorporate PTM information into protein–
protein interaction networks, making them more biologically meaningful.

Ultimately, the goal is to bridge computational predictions with clinical applications,

especially in the context of PTM-driven diseases.”

Slide 5 – Motivation
Narration:
“The motivation for this research lies in the potential of PTMs as biological regulators and
clinical biomarkers.

 A unified PTM prediction platform would allow researchers to query different

modifications in a standardized manner.
 More interpretable ML models are needed for mapping kinase–substrate or enzyme–
substrate relationships.
 Time-series PTM datasets could reveal how modifications evolve during disease
progression.
 High-throughput profiling of the palmitoyl-proteome across tissues and diseases
remains largely unexplored.

In the long run, I envision AI models that can predict causal PTM networks, simulate PTM
crosstalk effects, and even propose disease biomarkers for precision medicine.”

Slide 6 – Methodology

Narration:
“The methodology involves three main stages:

1. Dataset preparation – I collected positive and negative S-palmitoylation sites from the
PANIMoni mass spectrometry dataset. This dataset provides male, female, and
combined samples, allowing us to compare biological variations.
2. Feature extraction – Protein sequences were encoded using 566 physicochemical
features derived from the AAIndex database. A window size of 39 residues centered on
cysteine was chosen, ensuring sufficient local context.
3. Model development – Feature selection methods such as K-Best, Genetic Algorithm,
and Hybrid approaches were applied to reduce dimensionality. These optimized
features were then fed into a 1D CNN architecture for training and evaluation.”

Slide 7 – Dataset

Narration:
“The dataset used is the PANIMoni proteomic dataset (PXD025286), which provides site-
specific palmitoylation information in mouse proteins.

 Female dataset: 1036 proteins, 1773 positive sites, and 8934 negative sites.
 Male dataset: 1077 proteins, 1870 positives, and 9279 negatives.
 Combined dataset: 1180 proteins, 2083 positives, and 10,403 negatives.
This dataset is valuable because it captures sex-specific variations in palmitoylation patterns, an
often-overlooked biological dimension.”

Slide 8 – Feature Extraction

Narration:
“Feature extraction was performed using the AAIndex database, which catalogs various amino
acid properties.

 A sliding window of 39 amino acids around cysteine residues was applied.

 566 physicochemical features such as hydrophobicity, charge, polarity, and flexibility
were extracted.
 To avoid overfitting, feature selection was carried out using:
o K-Best statistical ranking,
o Genetic Algorithm for evolutionary optimization,
o and a Hybrid UN approach combining both.

This ensured that only the most informative features were used to train the CNN.”

Slide 9 – Results

Narration:
“The CNN-based model achieved encouraging results:

 For the female dataset, K-Best features achieved the best AUC of about 79%,
outperforming existing methods.
 For the male dataset, the Hybrid approach performed best with an AUC of around 76%.
 For the sex-independent dataset, the Hybrid approach again excelled, achieving ~78%
AUC.

Across all datasets, the CNN model consistently outperformed RFCM-PALM and even protein
language models. Precision, recall, and F1-scores were well balanced, showing that the model
generalizes well without bias toward positive or negative classes.”

Slide 10 – Research Gap (Expanded)

Narration:
“Even though the model performed well, broader research gaps remain:
  Databases: Even with growth in repositories like dbPTM 2025, data remain fragmented
across species and PTM types.
 Palmitoylation: Compared to phosphorylation, large-scale palmitoylation maps are rare,
and predictors are still limited.
 AI approaches: Most AI models developed in 2025, such as PTMFusionNet, emphasize
disease association but not mechanistic predictions.
 Clinical translation: Many PTM studies are limited to cell lines. Patient-derived
datasets are still lacking, which restricts real-world applicability.”

Slide 11 – Research Plan

Narration:
“My future research plan includes:

1. Extending the CNN model to other PTMs like phosphorylation and acetylation, making
it a general-purpose predictor.
2. Modeling protein–protein interaction networks with PTM context, so we can
understand how PTMs influence binding dynamics.
3. Integrating PTM-disease associations using large omics datasets, with the goal of
identifying new biomarkers for early diagnosis and drug response.

This integrative approach aims to bridge computational predictions with translational

biology.”

Slide 12 – Conclusion

Narration:
“To conclude:

 I developed a CNN model for predicting S-palmitoylation sites in mouse proteins.

 The model achieved high accuracy and outperformed state-of-the-art approaches.
 This work highlights the importance of sex-specific datasets, optimized feature
selection, and deep learning in PTM prediction.
 The framework can be scaled to other PTMs and integrated into systems biology and
clinical research.

Ultimately, this research contributes to the long-term goal of using PTMs as biomarkers and
therapeutic targets.”

Slide 13 – Acknowledgement
Narration:
“I sincerely thank my supervisors, Dr. Subhadip Basu of Jadavpur University and Dr.
Soumyendu Sekhar Bandyopadhyay of IEM Kolkata, for their mentorship and guidance. I
also acknowledge my lab colleagues and collaborators for their support.”

✅ Now you have a fully detailed, defense-ready presentation script with deeper scientific
explanations.

Would you like me to also prepare a Q&A preparation section (possible examiner questions +
best answers) so you can defend this work confidently?

You said:
Q&A preparation section (possible examiner questions + best answers)
ChatGPT said:

Excellent choice ✅ Preparing for Q&A is critical for a PhD defense or seminar. Below, I’ll give
you a structured list of likely examiner questions (general + technical + future scope) with
strong suggested answers tailored to your research.

Q&A Preparation for PTM Prediction Research

🔹 General Research Understanding

Q1. Why did you choose S-palmitoylation instead of other PTMs like phosphorylation?
Answer:
S-palmitoylation is an understudied but biologically important PTM. Unlike phosphorylation,
which has many prediction tools and large datasets, palmitoylation predictors are rare and
datasets are limited. Moreover, its reversible and dynamic nature makes it particularly
interesting in regulating protein localization and signaling. This gap in research provided both an
opportunity and a challenge, making it a suitable target for computational modeling.

Q2. What is the novelty of your work compared to existing tools like RFCM-PALM or
protein language models?
Answer:
The novelty lies in three aspects:

1. Optimized feature selection – I systematically used K-Best, Genetic Algorithm, and

Hybrid approaches to maximize discriminative features.
 2. Sex-specific and combined dataset analysis – Most previous works ignored biological
variation between male and female proteomes.
3. Deep learning with CNNs – My model captures local sequence dependencies better than
random forest–based methods or generic protein language models, resulting in higher
accuracy and balanced performance.

Q3. Why did you choose CNN over other models like LSTM, SVM, or Random Forest?
Answer:
CNNs are particularly effective in detecting local sequence motifs and position-specific
patterns in amino acid windows. Unlike LSTMs, which are better for long-range dependencies,
CNNs are computationally efficient and suitable for residue-level site prediction. SVMs and
Random Forests rely heavily on feature engineering, while CNNs can automatically learn
hierarchical representations of sequence features.

🔹 Dataset & Methodology

Q4. How did you ensure the quality of negative datasets?
Answer:
Negative datasets are challenging because experimentally validated negatives are scarce. To
address this:

 I used non-palmitoylated cysteines from experimentally validated proteins as

negatives.
 I ensured that negatives were balanced across male, female, and combined datasets.
 Additionally, I avoided randomly sampling negatives from unrelated proteins, which
could introduce bias.

Q5. Why did you use a window size of 39 residues around cysteine?
Answer:
Window size is crucial for capturing the local sequence context. Previous studies suggest that
palmitoylation is influenced by residues within 15–20 positions upstream and downstream of
cysteine. A window of 39 ensures sufficient context without introducing excessive noise. I also
tested smaller and larger windows, and 39 gave the most stable performance.

Q6. Why did you extract 566 features from AAIndex? Isn’t that too many?
Answer:
Yes, 566 is a large feature set. That’s why I used feature selection methods to reduce
dimensionality and avoid overfitting. For example:

 K-Best ranked features by statistical relevance.

 Genetic Algorithm explored combinations for optimal subsets.
 Hybrid (UN) approach combined the strengths of both.
This ensured that only the most informative features were used for CNN training.

🔹 Results & Interpretation

Q7. Why did K-Best perform best for the female dataset, while Hybrid worked better for
male and combined datasets?
Answer:
This may be due to biological differences in palmitoylation patterns between male and female
proteomes.

 The female dataset may contain stronger discriminative signals, which K-Best captures
efficiently.
 The male and combined datasets may have more complex patterns, requiring the
exploratory power of Genetic Algorithms in the Hybrid approach.
This highlights the importance of considering sex-specific biological variation in
computational biology.

Q8. How does your model compare with RFCM-PALM and protein language models in
terms of interpretability?
Answer:
RFCM-PALM, being tree-based, offers feature importance but struggles with sequence context.
Protein language models capture context but are often “black boxes.” CNNs strike a balance —
filters can be mapped back to motif-like sequence patterns, making the results more
biologically interpretable.

Q9. What do your results mean biologically? How can experimentalists use them?
Answer:
My model provides high-confidence predictions of S-palmitoylation sites. Experimentalists
can:

 Focus validation efforts on predicted sites instead of screening entire proteomes.

  Use predictions to design mutagenesis studies, confirming the functional impact of
specific cysteine modifications.
 Integrate results with protein interaction studies to understand how palmitoylation
rewires cellular signaling.

🔹 Future Scope & Limitations

Q10. Can your model be extended to other PTMs?
Answer:
Yes. The pipeline is designed to be generalizable. With appropriate datasets, the same CNN
framework can be applied to other PTMs like phosphorylation, acetylation, ubiquitination, or
glycosylation. Feature encoding may need slight adaptation, but the methodology remains the
same.

Q11. What are the limitations of your study?

Answer:

 Dataset imbalance – Negative sites are still fewer compared to positives.

 Species-specificity – Current model is trained on mouse data; transferability to human
proteomes requires validation.
 Lack of clinical datasets – Most available data are from mass spectrometry of lab
samples, not patient-derived samples.

Q12. How do you plan to improve this research in the future?

Answer:

 Expanding the model to multi-PTM prediction frameworks.

 Integrating predictions with PPI networks to model signaling pathways dynamically.
 Collaborating with experimental labs to validate predictions in real systems.
 Developing a web-based tool so the wider community can access predictions easily.

🔹 Defense-Oriented Deep Questions

Q13. Why not use transformer-based models (like ProtBERT, ESM)?
Answer:
Transformer models are powerful but require massive data and computational resources. While
they capture long-range dependencies, they are less efficient for site-specific prediction with
small datasets. CNNs provided a better trade-off between accuracy, interpretability, and
computational cost. However, integrating transformer embeddings as features for CNN could
be a promising hybrid approach in the future.

Q14. How do you handle overfitting in your CNN model?

Answer:
I used several strategies:

 Dropout layers to prevent co-adaptation of neurons.

 Early stopping to avoid over-training.
 Cross-validation across male, female, and combined datasets to check generalization.
 Feature selection reduced redundancy, lowering the risk of overfitting.

Q15. How do you see this work contributing to clinical or pharmaceutical applications?
Answer:

 Predicted palmitoylation sites can be used as potential biomarkers for early disease
detection.
 Drug development can target enzymes involved in palmitoylation (palmitoyltransferases),
guided by site predictions.
 Personalized medicine could benefit from sex-specific palmitoylation patterns,
informing treatment strategies.

✅ This covers likely examiner questions and strong answers.

Would you like me to also prepare a “1-minute crisp answer version” of this Q&A set (in case
time is short during defense)?

You said:
yes

ChatGPT said:

Perfect 👍 Here’s a 1-minute crisp answer version of the Q&A set — short, confident, and
defense-ready. You can expand if examiners ask for more, but this will help you stay concise
under pressure.
Quick Q&A Answers (Defense Style)

General

 Q: Why S-palmitoylation?
A: It’s biologically important, reversible, and underexplored compared to
phosphorylation — a clear research gap.
 Q: Novelty of your work?
A: Optimized feature selection, sex-specific datasets, and CNN architecture that
outperformed existing tools.
 Q: Why CNN over LSTM/SVM/RF?
A: CNN captures local motifs efficiently, trains faster, and balances accuracy with
interpretability.

Dataset & Method

 Q: How did you build negative datasets?

A: From experimentally confirmed non-palmitoylated cysteines; balanced across
datasets.
 Q: Why window size = 39?
A: Captures ~20 residues on each side — enough context without noise.
 Q: Why 566 features, isn’t that too many?
A: Yes — that’s why I applied K-Best, GA, and Hybrid selection to reduce to only
informative ones.

Results

 Q: Why K-Best better for female, Hybrid for male/combined?

A: Likely due to sex-specific biological variation; Hybrid better handles complexity.
 Q: How does your model compare with RFCM-PALM?
A: Higher AUC, balanced metrics, and CNN motifs are more interpretable.
 Q: Biological meaning?
A: Provides high-confidence predictions for targeted experiments and disease studies.

Future & Limitations

  Q: Can it extend to other PTMs?
A: Yes, same pipeline works for phosphorylation, acetylation, etc., with dataset
adaptation.
 Q: Limitations?
A: Dataset imbalance, species-specific training, and lack of clinical data.
 Q: Future plan?
A: Multi-PTM prediction, PPI integration, experimental validation, and web tool
deployment.

Defense-Oriented

 Q: Why not transformers?

A: They need huge data/resources; CNN is more efficient for small site-specific datasets.
 Q: How did you avoid overfitting?
A: Dropout, early stopping, cross-validation, and feature reduction.
 Q: Clinical relevance?
A: Supports biomarker discovery, enzyme-targeted drug design, and sex-specific
precision medicine.