Slide 3: Overview

 Proteins are not functionally complete just after translation.

 They undergo Post-Translational Modifications (PTMs) – chemical changes that diversify protein
structure and regulate biological processes.
 Major PTMs include phosphorylation, ubiquitination, acetylation, methylation, glycosylation, and
palmitoylation.
 PTMs are directly linked to diseases such as cancer, Alzheimer’s, Parkinson’s, cardiovascular
disorders, and depression.
 Understanding and predicting PTM sites can thus reveal disease mechanisms and guide therapeutic
strategies.

Slide 4: Research Gap

 Despite their importance, PTM research still faces challenges:

1. Data, coverage & benchmarking: Lack of unified, large-scale, high-quality datasets.
2. Generic models missing: Most models are PTM-specific; no robust model exists for predicting
multiple PTMs.
3. Crosstalk ignored: PTMs often interact, but studies treat them as isolated events.
4. Poor integration with PPI networks: PTMs are rarely mapped in protein-protein interaction
networks.
5. Incomplete PTM-disease mapping: Rare PTMs are overlooked in disease associations.

Slide 5: Scope of the Work

 My PhD work aims to:

o Develop a generic PTM site prediction model.
o Curate high-quality negative data for reliable in-silico predictions.
o Integrate PTMs into protein interaction networks.
o Perform computational analysis of PTM-disease associations.
 As a case study, I focus on S-palmitoylation prediction using deep learning.

Slide 6: S-Palmitoylation (Focus Area)

 S-palmitoylation is a reversible PTM where palmitic acid attaches to cysteine residues.

 This modification regulates protein stability, localization, and signaling.
 Accurate prediction is crucial for biological insights and clinical relevance.
 My proposed method: a Convolutional Neural Network (CNN) for predicting S-palmitoylation sites in
mouse proteins.

Section 1: General Research Understanding

Q1. Why did you choose to focus on Post-Translational Modifications (PTMs)?
A: PTMs are crucial regulators of protein function, influencing localization, stability, and signaling. They are
linked to several major diseases such as cancer, Alzheimer’s, and cardiovascular conditions. Despite their
importance, PTMs remain underexplored computationally, especially in terms of crosstalk, disease mapping,
and integration with protein-protein networks. This gap motivated me to contribute a computational
approach for PTM prediction.

Q2. Why did you select S-palmitoylation as your case study?

A: S-palmitoylation is a reversible lipid modification that plays a key role in neuronal signaling and disease
processes, yet its site-specific prediction is less explored compared to phosphorylation or ubiquitination. It
provided a biologically meaningful and computationally challenging problem, making it a good test case for
developing predictive models.

Q3. What is novel about your research compared to existing works?

A:

1. I curated high-quality negative data, which is often neglected in PTM prediction.

2. I applied feature selection strategies (N-Best, Genetic Algorithm, UN method) to improve predictive
performance.
3. I developed a CNN-based model that outperformed state-of-the-art methods like RFCM-PALM and
Protein Language Models.
4. The framework can be generalized to other PTMs, moving towards a generic PTM prediction model.

Section 2: Technical & Methodological

Q4. Why did you choose CNN instead of other models like LSTM or Random Forest?
A: CNNs are effective at capturing local sequential patterns in amino acid windows. Since PTM prediction
relies on short sequence motifs around the modification site, CNNs are a natural fit. In contrast, LSTMs are
designed for long-term dependencies, which are less critical here. Random Forests and traditional ML methods
require hand-crafted features, whereas CNNs learn feature representations automatically, improving accuracy.

Q5. How did you ensure your model was not overfitting?
A:

 Used cross-validation and a hold-out test set for unbiased evaluation.

 Performed regularization techniques such as dropout.
 Applied feature selection to avoid redundant or noisy features.
 Compared across male, female, and combined datasets to ensure generalizability.

Q6. Why did you use 39 residues as the window size?

A: Prior biological studies suggest that local sequence context around the cysteine residue (approximately
±19 residues) is sufficient to capture the determinants of S-palmitoylation. Smaller windows risk losing
important context, while larger windows add noise. Thus, 39 residues provided the best trade-off between
information and noise.

Q7. How does your feature selection improve performance?

A: Using all 566 AAIndex features may introduce redundancy. Feature selection methods (N-Best, GA, UN)
helped in identifying the most informative subsets, reducing dimensionality, improving training efficiency,
and enhancing prediction performance. For example, N-Best worked best for the female dataset, while UN
combined selection worked best for male and sex-independent datasets.

Q9. Why do you think the female dataset achieved the highest performance?
A: Biological sex differences may influence palmitoylation dynamics. Female-specific proteins might have
more distinct sequence motifs around modification sites, which CNNs could capture more effectively. This
suggests potential sex-specific biology in PTMs, which is an interesting avenue for future research.

Q10. How did you validate biological relevance beyond computational metrics?
A:

 Compared predicted motifs with known consensus sequences from experimental studies.
 Performed cross-dataset evaluation (Male vs Female vs Combined).
 Plan to collaborate with experimental biologists to validate predictions in lab-based assays.

Section 4: Broader Impact

Q11. Can your model be generalized to other PTMs?
A: Yes. The pipeline—feature extraction, feature selection, and CNN-based classification—is generic. By
replacing cysteine-centered windows with other residue types and retraining on relevant datasets, the same
method can predict other PTMs such as phosphorylation, ubiquitination, or acetylation.

Q13. What are the limitations of your work?

A:

1. Limited by the availability of curated datasets.

2. Focused on a single PTM (S-palmitoylation) for case study.
3. Predictions rely on sequence context only, while 3D structural data is not yet integrated.

Q14. What is your future research plan?

A:

 Extend the framework to multiple PTMs.

 Integrate structural and network-based features.
 Develop a web server or software tool for community use.
  Collaborate with experimental groups for lab validation of predicted PTM sites.

🔹 Scope of the Work – Detailed Explanation

1. Prediction of PTM Sites in Protein Sequences (Current Work)

o This is the core contribution of my thesis.
o I focused on S-palmitoylation prediction using CNN models trained on curated datasets.
o Extracted features from AAIndex, performed feature selection, and validated the model across
male, female, and combined datasets.
o The results show significant improvement over existing state-of-the-art methods.
o This part is complete and forms the experimental basis of my doctoral work.

2. Building a Generic Model for Multiple PTMs (Future Work)

o While my current model is specific to S-palmitoylation, the methodology (feature engineering +
CNN) can be extended to other PTMs such as phosphorylation, acetylation, and ubiquitination.
o Future work will involve training on larger and more diverse datasets to build a generalized
PTM predictor instead of a PTM-specific one.

3. Curation of High-Quality Negative Data (Future Work)

o PTM datasets often suffer from false negatives (sites assumed unmodified may simply be
untested).
o A major bottleneck is creating balanced and biologically meaningful negative datasets.
o My future plan is to develop robust negative data generation pipelines to improve in-silico
PTM prediction reliability.

4. Integration of PTMs in Protein Interaction Networks (Future Work)

o PTMs regulate protein–protein interactions, but most PPI databases ignore PTM effects.
o The next step is to integrate PTM predictions into network biology frameworks, showing how
PTMs affect interaction dynamics, signaling pathways, and disease progression.

5. Computational Analysis of PTM-Disease Associations (Future Work)

o Many diseases (e.g., neurodegeneration, cancer) are linked to PTM dysregulation.
o Using computational approaches, I aim to map predicted PTM sites to known disease
pathways.
o This will help in biomarker discovery and potential therapeutic targeting.

🔑 How to Present This Slide

 “The first point is my present PhD contribution”

  “The remaining points represent my planned future research direction”

🔹 Slide-by-Slide in Current Work vs Future Work

Pattern

Slide: Overview
✅ Current Work

 Explained the biological importance of PTMs (e.g., phosphorylation, ubiquitination, glycosylation,

palmitoylation).
 Established their role in protein diversity, regulation, and disease mechanisms.
 Identified PTMs as a relevant problem area for computational prediction.

🚀 Future Work

 Expand beyond descriptive overview → build a comprehensive PTM knowledgebase integrating

sequence, structural, and disease information.
 Contribute to PTM ontologies and annotations for broader research communities.

Slide: Scope of the Work

(already done earlier, but summarizing)
✅ Current Work – Prediction of S-palmitoylation sites using CNN.
🚀 Future Work – Generic PTM predictor, negative dataset curation, PPI integration, disease association.

Slide: Feature Extraction & Selection

✅ Current Work

 Extracted 566 physicochemical features from AAIndex.

 Used 39-residue sliding window around cysteine.
 Applied N-Best, Genetic Algorithm, and Combined methods for feature selection.

🚀 Future Work
  Use deep representation learning (protein language models, embeddings like ESM or ProtBERT) as
alternative feature sets.
 Explore 3D structure-based features (solvent accessibility, secondary structure, dynamics).
 Develop hybrid feature engineering pipelines.

Slide: Comparison with State-of-the-Art

✅ Current Work

 Compared CNN with RFCM-PALM and protein language models.

 CNN consistently showed higher AUC and balanced performance across datasets.

🚀 Future Work

 Benchmark against emerging large-scale PTM predictors (transformers, graph neural networks).
 Create a standard benchmarking pipeline for future PTM prediction research.

🎤 20-Minute Presentation Script

Slide 1: Title (1 min)

“Computational Analysis of Post-Translational Modifications of Proteins”

 Good [morning/afternoon] to everyone present.

 My name is Priti Das, and today I will present my PhD research on the computational analysis of
post-translational modifications (PTMs) of proteins, with a focus on predicting S-palmitoylation
sites using deep learning.
 I will also highlight the research gaps, my contributions, and the future directions this work can take.

Slide 3: Overview (1.5 min)

 Proteins are the fundamental building blocks of life. But their functions are not fully defined by
genetic sequences alone.
 After synthesis, proteins undergo post-translational modifications (PTMs)—chemical changes that
add functional diversity.
 Common PTMs include phosphorylation, ubiquitination, acetylation, methylation, glycosylation,
and palmitoylation.
 PTMs regulate protein activity, stability, localization, and interactions, and are linked to diseases
such as cancer, Alzheimer’s, Parkinson’s, and cardiovascular disorders.
 Therefore, understanding and predicting PTM sites is essential both for basic biology and for
therapeutic applications.
Slide 4: Research Gap (2 min)
 Despite their importance, several gaps remain:
1. Data & Benchmarking: Available datasets are scattered, noisy, and often lack proper
benchmarks.
2. No Generic Model: Most prediction tools are PTM-specific, not generalizable.
3. Crosstalk Ignored: PTMs interact with each other, but most models treat them as independent.
4. Poor Integration with Protein-Protein Interaction (PPI) Networks: PTM effects on
interactions are rarely considered.
5. Incomplete Disease Mapping: Only common PTMs are studied; rare PTMs are neglected.
 These gaps highlight the need for robust computational approaches that are both accurate and
generalizable.

Slide 5: Scope of the Work (2 min)

 My PhD work primarily focused on:
o Prediction of S-palmitoylation sites in protein sequences using CNN.
 For future directions, this framework can be extended to:
o Build a generic PTM predictor across multiple modification types.
o Curate high-quality negative data, which is a major bottleneck.
o Integrate PTM predictions into PPI networks to study signaling pathways.
o Analyze PTM-disease associations computationally for biomarker discovery.
 So, the core contribution is S-palmitoylation prediction, while the other points open up long-term
research opportunities.

Slide 6: S-Palmitoylation (Focus Area) (1.5 min)

 Among many PTMs, I selected S-palmitoylation as my case study.
 It is a reversible modification where palmitic acid binds to cysteine residues.
 This modification regulates protein trafficking, localization, and stability, especially in neurons.
 Dysregulation of S-palmitoylation is linked to cancer and neurodegenerative diseases.
 However, site-specific prediction of S-palmitoylation is underexplored, making it an ideal case for
testing deep learning models.

Slide 8–9: Feature Extraction & Selection (2 min)

 To train the CNN, I extracted 566 physicochemical features from the AAIndex database.
 Used a 39-residue sliding window around cysteine (19 before + cysteine + 19 after).
 Also calculated position-specific amino acid propensities.
 Then applied feature selection methods:
 o N-Best (selects top-ranked features).
o Genetic Algorithm (GA) (evolutionary selection).
o Combined Method (UN) (hybrid approach).
 These reduced dimensionality, removed noise, and improved predictive performance.

Slide 11: Comparison with State-of-the-Art (2 min)

 Compared my CNN with RFCM-PALM and protein language models.
 CNN outperformed both in AUC and overall balance.
 Especially for female dataset, CNN showed a clear margin of improvement.
 This confirms that deep learning with feature selection is more effective than both classical ML and
large protein language models.

🎤 Research Gap – Detailed Presentation Script (3

min)
When presenting, you can structure it like this:

Introduction (10–15 sec)

 "Now let me discuss the research gaps that motivated this work. Despite the increasing interest in post-
translational modifications, there are still major limitations in existing computational studies."

Gap 1: Data, Coverage, and Benchmarking (30–40 sec)

 Current PTM datasets are fragmented, noisy, and incomplete.

 Positive samples exist, but negative data is often assumed, not experimentally validated.
 Benchmark datasets are lacking → makes it difficult to fairly compare methods.
 Without curated, large-scale, high-quality datasets, prediction models remain less generalizable.

Gap 2: Lack of a Generic Prediction Model (30–40 sec)

 Most existing tools are PTM-specific, e.g., one model for phosphorylation, another for ubiquitination.
 No unified or generic framework exists that can handle multiple PTMs simultaneously.
 This makes computational PTM research fragmented and reduces translational applications.
Gap 3: Limited Understanding of Crosstalk and Proteoforms (30–40 sec)

 PTMs do not work in isolation; they interact and influence each other (crosstalk).
 For example, phosphorylation at one site can block or enable ubiquitination at another.
 Current models treat PTMs as independent events, ignoring such biological interactions.
 Also, proteoforms (different modified versions of the same protein) are rarely modeled.

Gap 4: Incomplete Integration with Protein-Protein Interaction (PPI) Networks (30–40 sec)

 PTMs regulate how proteins interact, but most PPI databases ignore PTM effects.
 A protein’s interaction partners can change depending on its modification state.
 Without PTM-aware network modeling, we miss key insights into signaling pathways.

Gap 5: Incomplete PTM-Disease Mapping (30–40 sec)

 While some common PTMs (like phosphorylation) are linked to diseases, rare PTMs are
underexplored.
 Current studies provide incomplete disease associations.
 This makes it harder to use PTMs as biomarkers or therapeutic targets.

Closing Transition (10–15 sec)

 "These gaps clearly highlight the need for better computational approaches. My research specifically
addresses these challenges by building a CNN-based framework for S-palmitoylation prediction, while
also laying the groundwork for a more general PTM prediction system."

🎤 Footer Speech for Results Slide

"To summarize the results, my CNN-based framework achieved strong predictive performance across male,
female, and sex-independent datasets, with balanced precision, recall, and F1-scores. The best AUC was
observed for the female dataset at nearly 79%, while male and combined datasets reached 76% and 78%
respectively. These results demonstrate not only the robustness of the model but also its potential biological
relevance in capturing sex-specific patterns in palmitoylation. Most importantly, the CNN consistently
outperformed existing state-of-the-art methods, which confirms the novelty and strength of my approach. This
sets the stage for the next slide, where I present a direct comparison with current state-of-the-art models."
🎤 Feature Extraction & Selection – Detailed
Presentation Script
Introduction (10–15 sec)

"Once I prepared the dataset, the next step was to convert raw protein sequences into numerical
representations suitable for machine learning. This process is called feature extraction, followed by feature
selection to identify the most informative features for prediction."

Feature Extraction (1.5 min)

1. AAIndex Database
o I extracted 566 physicochemical features from the AAIndex database, which contains
numerical indices representing amino acid properties.
o These include hydrophobicity, polarity, flexibility, charge, and steric parameters.
o The goal is to capture the biochemical environment around cysteine residues where
palmitoylation occurs.
2. Sliding Window Approach
o For each candidate cysteine site, I used a 39-residue window: 19 residues before + cysteine
itself + 19 after.
o This ensures local sequence context is included, as PTMs are often determined by short motifs.
3. Position-Specific Amino Acid Propensity (PSAAP)
o In addition to physicochemical descriptors, I calculated propensity scores for each amino acid
position.
o This captures motif-like patterns — for example, whether certain residues tend to occur near
palmitoylated cysteines.
4. Final Feature Vector
o After combining physicochemical indices and PSAAP values, each window is represented as a
high-dimensional vector for classification.

Feature Selection (1 min)

1. Why Feature Selection?

o With 566+ features per sample, the dataset is high-dimensional.
o Not all features are informative → many may introduce noise or redundancy, leading to
overfitting.
o Feature selection helps to reduce dimensionality, speed up training, and improve
generalization.
2. Methods Used
o N-Best: Ranks features by statistical relevance and selects the top-performing subset.
o Genetic Algorithm (GA): An evolutionary method that searches for optimal feature subsets by
simulating natural selection.
 oCombined (UN method): A hybrid approach combining statistical filtering with heuristic search
for robust feature selection.
3. Observations
o N-Best performed best for the female dataset (~79% AUC).
o Combined method (UN) worked best for male (~76%) and sex-independent (~78%) datasets.
o This indicates that different biological groups may require different feature subsets for
optimal prediction.

Closing Transition (20–30 sec)

"In summary, careful feature extraction allowed us to capture meaningful biochemical patterns, and feature
selection ensured that only the most informative signals were used for CNN training. This combination directly
contributed to the improved performance of the model, which I will now present in the results section."

🔹 Detailed Explanation of Feature Selection

Methods
1. N-Best Method (Filter-based Selection)

 Process:
o Each feature is scored individually based on its ability to discriminate between positive and
negative samples.
o Statistical measures like F-score, chi-square, or information gain are often used.
o Features are then ranked from most informative to least informative.
o The top N features are selected as the optimal subset.
 Strengths: Fast, interpretable, and avoids overfitting by removing irrelevant features.
 Limitation: Considers features independently, so it may miss combinations of weaker features that are
powerful together.

2. Genetic Algorithm (GA) (Wrapper-based Selection)

 Process:
o GA is an evolutionary optimization technique inspired by natural selection.
o It starts with a population of random feature subsets.
o Each subset is evaluated using the model’s performance (e.g., classification accuracy, AUC).
o The best-performing subsets are “selected” to create the next generation through crossover
(combining features from parents) and mutation (randomly altering some features).
o This process repeats until convergence, leading to an optimized subset of features.
 Strengths:
o Considers interactions between features.
o Can explore a large feature space effectively.
 Limitation: Computationally expensive compared to N-Best.
3. Combined (UN Method)

 Process:
o This method integrates both filter-based ranking (like N-Best) and wrapper-based search
(like GA).
o First, features are pre-filtered statistically to remove irrelevant ones.
o Then, GA is applied on the reduced feature set to identify the most discriminative
combination.
 Strengths:
o Balances speed and accuracy.
o Reduces risk of overfitting by combining global statistical relevance with local subset
optimization.
 Limitation: More complex pipeline, but gives robust and stable results.

🔑 In the Context of Your Work

 Female dataset: N-Best was sufficient → suggests strong discriminative individual features.
 Male & Combined datasets: UN method worked best → suggests that feature interactions were more
important, captured better by GA after pre-filtering.

🎤 Dataset + Feature Extraction & Selection

(Detailed Script)

Categories

1. Male dataset
2. Female dataset
3. Combined dataset (sex-independent)

Experimental Conditions

 Each category includes:

o Knockout (KO) samples
o Wild-type (WT) samples

Why This Dataset?

 Most previous PTM studies do not consider sex-specific differences, but emerging research shows that
modifications like palmitoylation may vary between male and female samples.
 By analyzing separately and jointly, I was able to investigate:
 o Whether sex influences palmitoylation patterns.
o How predictive performance changes when datasets are combined vs separated.

Transition:
"Once I had the dataset prepared, the next step was to represent protein sequences numerically, which brings
me to feature extraction and selection."

Feature Extraction (1.5 min)

1. AAIndex Database
o Extracted 566 physicochemical properties of amino acids.
o These include hydrophobicity, polarity, charge, flexibility, steric properties, etc.
o Such descriptors capture the biochemical environment around potential palmitoylation sites.
2. Sliding Window (39 residues)
o For each cysteine candidate site: 19 amino acids before + cysteine + 19 after.
o This window ensures local sequence context, which is crucial since PTMs depend on nearby
motifs.
3. Position-Specific Amino Acid Propensity (PSAAP)
o Calculated how likely each amino acid appears at each position relative to cysteine.
o Helps capture motif signatures that characterize palmitoylation.
4. Final Feature Vector
o Each site is represented as a high-dimensional vector combining AAIndex descriptors and
PSAAP scores.

Feature Selection (1 min)

 With 566+ features, dimensionality reduction was necessary to avoid noise and overfitting.
 I applied three strategies:

N-Best Method

 Ranked features by statistical relevance (e.g., chi-square, F-score).

 Selected the top-ranked N features.
 Fast and effective when strong individual features exist.

Genetic Algorithm (GA)

 Inspired by evolutionary selection.

 Starts with random feature subsets → evaluates performance → applies crossover and mutation →
iterates until optimal subset is found.
 Captures feature interactions that simple ranking misses.

Combined (UN Method)

 First, a statistical filter removes irrelevant features.

 Then GA is applied on this reduced set.
  Provides a balance of speed and robustness.

Closing Transition (20–30 sec)

"In summary, the dataset provided a rich basis for studying palmitoylation across male, female, and combined
groups. By extracting physicochemical features from AAIndex and carefully selecting them using three
complementary methods, I ensured that the CNN model was trained on the most informative patterns. These
steps directly contributed to the strong results I will now present."

👉 This combined dataset + feature pipeline explanation will take 3–4 minutes, keeping your total presentation
within 20 minutes.

Would you like me to also prepare a table-style summary (Dataset → Features Extracted → Selection
Method → CNN Input) so you can show your entire pipeline in one visual slide for clarity?