1

1. Understanding genetic mutations and their impact.

2. Our DNA is made up of genes, and genes contain instructions to make proteins,
which are essential for life. However, only about 1-2% of our DNA directly codes
for proteins. The rest (about 98%) is called non-coding DNA.
3. A genetic mutation is a change in a sequence of your DNA. Genetic mutations occur
during cell division when your cells divide and replicate. Symptoms of genetic
conditions depend on which gene has a mutation. There are many different diseases
and conditions caused by mutations.

Variants(muatation)can happen in both coding and non-coding regions.

Coding Variants → Changes in protein-coding regions, which can alter protein

structure or function.
Non-Coding Variants → Changes in non-coding regions, which do not directly
affect proteins but can influence gene activity.

Where Do Non-Coding Variants(muatation) Occur?

Non-coding variants(mutation) can appear in regions like:

1. Promoters – Control when a gene is turned on/off.

2. Enhancers – Boost gene activity.
3. Introns – Non-coding sections within genes.
4. Regulatory RNAs – Like microRNAs, which control gene expression.
5. Intergenic regions – Areas between genes that may have unknown functions.

Examples of Non-Coding Variants(mutations) in Diseases

 Cancer – Some non-coding variants increase the risk of cancer by disrupting gene
regulation.
 Heart Disease – Non-coding mutations in enhancers can affect cholesterol-related
genes.
Diabetes – Some non-coding variants alter insulin-related genes.

AI in Genomic Analysis - The role of AI in detecting and analyzing genetic

variations.

How AI Helps with Non-Coding Variants(mutations)The Role of AI in

Genomic Analysis

AI algorithms, particularly machine learning and deep learning models, enhance the accuracy
and efficiency of genomic data interpretation.

Traditional methods of identifying genetic mutations often involve labor-intensive processes

that require extensive manual examination of sequencing data.

AI streamlines this process by analyzing vast amounts of genetic information at an

unprecedented speed, reducing both time and costs associated with genomic studies
 2

Because non-coding variants are hard to interpret, AI models analyze large

datasets to predict their impact. AI helps in:

✅ Identifying which non-coding variants are important.

✅ Understanding how they influence gene expression.
✅ Improving personalized medicine and disease prediction.

Researchers continue to develop methodologies to improve AI's reliability and reduce biases
in genomic interpretations.

1. AI can analyze multi-omics data(Multi-omics refers to integrating multiple types of

biological data (genomics, epigenomics, transcriptomics, proteomics, etc.) to gain a
comprehensive understanding of biological systems, diseases, and genetic regulation. to
predict disease risk and develop targeted treatments.
2. You can use deep learning (CNNs, Transformers) or graph neural networks to integrate
multi-omics data and improve the functional annotation of regulatory variants.

AI leverages deep learning, machine learning, and statistical models to analyze vast
genomic datasets and predict how a non-coding variant affects gene expression, chromatin
structure, and regulatory elements like enhancers and promoters.

3. Literature Gap: Current AI models struggle to interpret non-coding mutations,

which play a crucial role in diseases but are harder to analyze than coding mutations.

While AI models like DeepSEA and ExPecto have made progress, they are still limited in
non-coding variant interpretation:

 DeepSEA (Zhou et al., 2015)

✅ Uses CNNs to predict chromatin accessibility, TF binding, and histone
modifications.
❌ Cannot fully predict the downstream effects of non-coding mutations on gene
expression.
 ExPecto (Zhou et al., 2018)
✅ Uses deep learning to predict gene expression from DNA sequences.
❌ Struggles with long-range regulatory interactions and context-specific effects.

To improve interpretation, AI needs to integrate multiple layers of biological data,

including:

 Epigenetic Data (e.g., DNA methylation, histone modifications)

 Chromatin Accessibility (e.g., ATAC-seq, DNase-seq)
 RNA Expression (e.g., single-cell RNA-seq, GTEx data)
 3D Genome Organization (e.g., Hi-C data for enhancer-promoter interactions)

Here is a literature survey summarizing key deep learning models for predicting gene
expression and their limitations:

Here is a literature survey table including DeepSEA (Zhou et al., 2015) and related models,
along with their accuracy percentages:
 3

Accuracy /
Model Study Approach Strengths Limitations
Performance
Accurately
CNN-based predicts TF Cannot fully
Zhou model for binding, predict downstream
DeepSEA et al., chromatin chromatin effects of non- AUC-ROC: 91-94%
2015 feature accessibility, and coding mutations
prediction histone on gene expression
modifications
Struggles with
Deep learning long-range
Zhou Predicts gene Spearman correlation:
model for gene regulatory
ExPecto et al., expression from ~0.82 (median), up to
expression interactions and
2018 DNA sequences 0.9 for some tissues
prediction context-specific
effects
CNN-based
Kelley Captures long- Limited
model for long- Pearson correlation:
Basenji et al., range interactions generalizability
range regulatory 0.85
2018 up to 200kb across cell types
prediction
Transformer-
More accurate
Avsec based Requires high
long-range Pearson correlation:
Enformer et al., architecture for computational
dependency 0.90+
2021 gene expression resources
modeling
prediction
Deep learning
Min et Predicts enhancer
for enhancer- Needs large labeled
DeepMEL al., activity and gene AUC-ROC: ~85-88%
promoter datasets for training
2021 expression
interactions

AUC-ROC (Area Under the Receiver Operating Characteristic Curve) is used for binary
classification tasks like TF binding prediction.
 Spearman and Pearson correlations measure the model’s ability to predict gene
expression levels compared to experimental data.

 Higher values indicate better performance. Enformer shows state-of-the-art

performance due to its Transformer-based architecture.

Methodology
In this section, we explore the techniques used in AI-driven genomic analysis. The process
generally involves:

1. Data Collection – Gathering genomic sequences from databases like NCBI and
Ensembl.
2. Preprocessing – Cleaning and normalizing the data for AI models.
3. Feature Selection – Identifying key genetic markers using AI algorithms.
4. Model Training – Using Deep Learning models like CNNs or Transformers to
analyze mutations.
 4

5. Prediction and Validation – Evaluating how mutations affect gene expression and
disease risk.
Through these steps, AI enhances the accuracy and efficiency of genomic studies

The proposed work focuses on:

[Link] can use existing deep learning models (e.g., Enformer, Basenji2, DeepSEA, ExPecto)
instead of building a new one.
[Link] work will focus on comparing and optimizing models instead of creating a complex
AI architecture.
[Link] fine-tuning models or combining multi-omics data to improve accuracy.
Objectives of the Research

1️⃣ To Evaluate the Performance of Pre-Trained AI Models

 Compare the accuracy and effectiveness of AI models like DeepSEA, Enformer,

ExPecto, and Basenji2 in detecting non-coding genetic mutations.

2️⃣ To Integrate Multi-Omics Data for Improved Mutation Prediction

 Use ClinVar, gnomAD, ENCODE, and ATAC-seq datasets to analyze how AI

models perform when integrating chromatin accessibility, histone marks, and gene
expression data.

3️⃣ To Identify the Most Effective AI Model for Regulatory Mutation Detection

 Benchmark AI predictions against experimentally validated mutations to determine

which model provides the most reliable predictions.

Future advancements

 Predicting the impact of mutations on gene expression.

 Prioritizing regulatory variants that influence gene activity.
 Enhancing model accuracy by incorporating more diverse datasets.
 Developing AI-driven models that can analyze complex genetic interactions.

Future advancements will allow AI to detect mutations more precisely, aiding in early
diagnosis and treatment planning

conclude,

AI is revolutionizing genetic research by analyze genetic data to:

 Predict disease risk and treatment response.

 Create accurate models for disease forecasting.
 Integrate genetic and environmental factors for a holistic view

Q& A
 5

Definition of DNA, Gene, and RNA

1. DNA (Deoxyribonucleic Acid)
Definition: DNA is the genetic material that stores instructions for building and maintaining
an organism. It is composed of nucleotides (A, T, C, G) and has a double-helix structure.
Functions:
✅ Stores genetic information.
✅ Passes genetic traits from parents to offspring.
✅ Provides instructions for making proteins through gene expression.

2. Gene
Definition: A gene is a specific segment of DNA that contains the instructions to make a
protein or functional RNA. Genes are the basic units of heredity.
Functions:
✅ Determines traits (eye color, height, disease risk).
✅ Codes for proteins that perform essential biological functions.
✅ Can be turned ON or OFF based on environmental and regulatory factors.

3. RNA (Ribonucleic Acid)

Definition: RNA is a single-stranded molecule that helps convert DNA instructions into
proteins.
Types of RNA & Functions:
 mRNA (Messenger RNA) → Carries genetic instructions from DNA to ribosomes for protein
synthesis.
 tRNA (Transfer RNA) → Helps assemble amino acids into proteins.
 rRNA (Ribosomal RNA) → Forms part of ribosomes, which make proteins.
 miRNA & siRNA → Regulate gene expression (turn genes ON or OFF).

Relationship Between DNA, Gene, and RNA

1. DNA stores genes, which contain instructions.
2. Genes are transcribed into RNA (mRNA).
3. mRNA is translated into proteins, which carry out biological functions.
Would you like a diagram or further explanation on any of these? 😊
Genetic analysis :
Genetic Analysis: Understanding DNA and Its Variations
Genetic analysis is the study of DNA, genes, and genetic variations to understand how they
influence traits, diseases, and biological functions. It is widely used in medicine, agriculture,
evolutionary biology, and forensic science.

1. Steps in Genetic Analysis

1.1. DNA Extraction and Sequencing
 DNA Extraction: Isolating DNA from cells (blood, saliva, or tissue samples).
 DNA Sequencing: Determining the exact order of nucleotides (A, T, C, G) in a DNA molecule.
o Sanger Sequencing (older, accurate for small sequences).
o Next-Generation Sequencing (NGS) (modern, high-throughput for whole genomes).
1.2. Variant Detection (Mutation Analysis)
 Identifying Single Nucleotide Variants (SNVs), Insertions/Deletions (InDels), and Structural
Variants (SVs) in DNA.
 Tools used: GATK, VarScan, DeepVariant (AI-based).
1.3. Functional Annotation of Variants
 6

 Predicting the impact of mutations on genes and proteins.

 Common databases:
o ClinVar (disease-related variants).
o CADD, SIFT, PolyPhen-2 (predict mutation pathogenicity).
1.4. Gene Expression Analysis
 Measuring which genes are active and at what levels.
 Methods:
o RNA Sequencing (RNA-seq) – Quantifies gene expression levels.
o Microarrays – Older method, detects known gene expression patterns.
1.5. Epigenetic Analysis
 Studying DNA methylation, histone modifications, and chromatin accessibility, which
regulate gene expression without changing DNA sequence.
 Techniques:
o ATAC-seq – Measures chromatin accessibility.
o ChIP-seq – Detects protein-DNA interactions.

2. Applications of Genetic Analysis

2.1. Medical Genetics & Disease Research
✅ Identifies genetic mutations associated with diseases (e.g., cancer, inherited disorders).
✅ Helps in personalized medicine (e.g., targeted cancer therapies based on tumor genetics).
2.2. Evolutionary and Population Genetics
✅ Studies genetic variation in populations to understand human evolution and ancestry.
✅ Used in conservation biology to protect endangered species.
2.3. Agricultural and Animal Genetics
✅ Improves crop yield, disease resistance, and livestock breeding through genetic
modifications (e.g., GMO crops).
2.4. Forensic and Ancestry Analysis
✅ DNA fingerprinting for crime investigations and personal ancestry tracking (e.g., 23andMe,
AncestryDNA).

3. Role of AI in Genetic Analysis

 Deep Learning for Mutation Prediction: Models like DeepSEA, Enformer predict regulatory
effects of mutations.
 Machine Learning for Disease Diagnosis: AI-powered tools analyze genomic data to predict
disease risks.
 Multi-Omics Integration: AI combines genomics, transcriptomics, and epigenomics for more
accurate insights.
Would you like details on a specific aspect of genetic analysis? 😊
Gene Regulation: The Control of Gene Expression
Gene regulation is the process by which cells control when, where, and how much a gene
is expressed. It ensures that genes are activated or repressed at the right time, in the right cell
type, and under the right conditions.
Why is Gene Regulation Important?
 Allows cells to respond to environmental changes (e.g., stress, nutrients).
 Helps in cell differentiation (e.g., a neuron and a muscle cell have the same DNA but
express different genes).
 Prevents diseases—dysregulated gene expression can lead to cancer, genetic
disorders, etc.
 7

AI models struggle to interpret non-coding mutations more than coding

mutations due to several key reasons:
1. Lack of Clear Functional Annotation
 Coding Mutations: Their impact is relatively easier to assess since they directly alter
amino acids, affecting protein function (e.g., missense or nonsense mutations).
 Non-Coding Mutations: These mutations occur in regulatory regions (e.g.,
enhancers, promoters, introns, UTRs), where their effects are indirect and context-
dependent.
2. Complex and Context-Specific Regulation
 Gene expression is controlled by a network of transcription factors (TFs),
enhancers, chromatin modifications, and 3D genome organization.
 Non-coding mutations may disrupt these long-range interactions, leading to subtle
or cell-type-specific effects that AI struggles to capture.
3. Insufficient High-Quality Training Data
 Coding mutations: Well-documented in databases like ClinVar and COSMIC.
 Non-coding mutations: Fewer experimentally validated datasets exist, making AI
training more challenging.
4. Difficulty in Predicting Long-Range Effects
 Non-coding regions do not act locally; enhancers can regulate genes located
hundreds of kilobases away.
 AI models like DeepSEA and ExPecto struggle with long-range interactions,
whereas newer architectures like Enformer try to address this issue.
5. Epigenetic and Chromatin Accessibility Influence
 The same non-coding mutation may have different effects depending on the
chromatin state (open vs. closed chromatin).
 AI models need multi-omics data (ATAC-seq, ChIP-seq, Hi-C) to infer these effects
accurately, but integrating such data remains difficult.
6. Limited Interpretability of Deep Learning Models
 AI models like CNNs and Transformers predict outcomes well, but they do not
provide clear biological explanations for how non-coding mutations alter gene
regulation.
 Unlike coding mutations, where effects can be directly linked to protein structure,
non-coding effects require indirect inference.
Emerging Solutions
 Transformer-based models (e.g., Enformer): Capture long-range interactions
better.
 Multi-modal AI approaches: Combine genomics, epigenomics, and 3D chromatin
data.
 Large-scale experimental validation (MPRA, CRISPR screens): Generate better
training datasets for AI models.
Common Deep Learning & Machine Learning Architectures in Gene
Mutation and Regulation
Machine Learning (ML) and Deep Learning (DL) have significantly advanced our
understanding of gene mutations and gene regulation by predicting their effects on gene
expression, chromatin accessibility, and disease association. Below are some of the most
commonly used architectures in this field:
 8

1. Convolutional Neural Networks (CNNs)

Use Case:
✅ Detecting DNA sequence patterns, predicting transcription factor (TF) binding, chromatin
accessibility, and histone modifications.
How it Works:
 CNNs apply filters/kernels to scan DNA sequences for regulatory motifs.
 They extract spatial patterns that are crucial for gene regulation.
Examples:
Model Study Application
Zhou et al., Predicts TF binding, chromatin accessibility, and histone modifications
DeepSEA
2015 from DNA.
Kelley et al.,
Basenji Predicts gene expression over long genomic regions.
2018
Kelley et al.,
Basset Classifies functional regulatory elements in DNA sequences.
2016
Strengths:
✅ Identifies local sequence motifs efficiently.
✅ Performs well on genomic sequence-based tasks.
Limitations:
❌ Struggles with long-range dependencies (e.g., enhancer-promoter interactions).
❌ Requires large amounts of labeled data.

2. Recurrent Neural Networks (RNNs) & Long Short-Term Memory

(LSTM)
Use Case:
✅ Modeling sequential dependencies in gene expression and alternative splicing.
How it Works:
 RNNs process sequential genomic data, learning patterns across time/position.
 LSTMs solve vanishing gradient issues, allowing long-term dependencies to be captured.
Examples:
Model Study Application
Hybrid CNN + BiLSTM for predicting chromatin accessibility and TF
DanQ Quang & Xie, 2016
binding.
Strengths:
✅ Captures sequential dependencies in gene expression regulation.
✅ Effective for RNA-related predictions (e.g., splicing).
Limitations:
❌ Computationally expensive.
❌ Less effective than CNNs for spatial feature extraction.

3. Transformer-Based Models
Use Case:
✅ Predicting long-range regulatory interactions, gene expression, and chromatin accessibility.
How it Works:
 Transformers use self-attention to model long-range dependencies between genomic
elements.
 Unlike RNNs, they process sequences in parallel, improving efficiency.
 9

Examples:
Model Study Application
Avsec et al., Predicts gene expression and regulatory interactions over 1Mbp+
Enformer
2021 sequences.
Singh et al.,
Nucleome Predicts 3D genome organization and enhancer-promoter interactions.
2022
Strengths:
✅ Captures long-range regulatory interactions (e.g., enhancers affecting distant genes).
✅ More interpretable than CNNs/RNNs for sequence-based predictions.
Limitations:
❌ Requires high computational power.
❌ Needs large-scale datasets for training.

4. Graph Neural Networks (GNNs)

Use Case:
✅ Modeling gene regulatory networks (GRNs) and 3D chromatin interactions (e.g.,
enhancer-promoter loops).
How it Works:
 GNNs represent biological systems as graphs, where nodes (genes) and edges (interactions)
are learned dynamically.
 Used for predicting regulatory interactions and identifying key mutations affecting gene
expression.
Examples:
Model Study Application
DeepGraphGO Kulmanov et al., 2020 Predicts gene function using biological networks.
Geneformer Lotfollahi et al., 2023 Uses transformer + GNN for gene regulatory modeling.
Strengths:
✅ Captures complex interactions in regulatory networks.
✅ Effective for predicting mutational impact on pathways.
Limitations:
❌ Requires detailed network annotations.
❌ Computationally intensive for large biological datasets.

5. Hybrid Models (CNN + RNN / Transformer)

Use Case:
✅ Combining CNNs’ ability to detect local motifs with RNNs/Transformers' ability to
capture long-range dependencies.
Examples:
Model Study Application
DanQ Quang & Xie, 2016 CNN + BiLSTM for TF binding site prediction.
BPNet Avsec et al., 2021 CNN + Transformer for regulatory sequence modeling.
Strengths:
✅ Captures both short-range and long-range dependencies.
✅ More accurate for complex regulatory sequence predictions.
Limitations:
 10

❌ More computationally expensive than pure CNNs or RNNs.

❌ Requires careful hyperparameter tuning.

6. Traditional Machine Learning Approaches

Common ML Algorithms:
✅ Random Forests (RF) – Used for predicting pathogenic mutations (e.g., PolyPhen-2).
✅ Support Vector Machines (SVMs) – Used for classifying functional vs. non-functional
mutations.
✅ Gradient Boosting (e.g., XGBoost, LightGBM) – Applied in SNP effect prediction.
Examples:
Model Study Application
PolyPhen-2 Adzhubei et al., 2010 Predicts if a coding mutation is damaging.
CADD Kircher et al., 2014 Scores functional effects of genetic variants.
Strengths:
✅ Less data-intensive than deep learning.
✅ Works well on small datasets.
Limitations:
❌ Struggles with high-dimensional genomic data.
❌ Requires manual feature engineering.

Conclusion: Choosing the Right Model

Task Best ML/DL Approach
TF Binding, Chromatin Accessibility CNNs (DeepSEA, Basset), CNN+LSTM (DanQ)
Gene Expression Prediction Transformers (Enformer), CNNs (Basenji)
Regulatory Network Prediction Graph Neural Networks (Geneformer)
3D Genome Structure Transformers + GNNs (Nucleome)
Mutation Effect Prediction Gradient Boosting (CADD), CNNs (DeepSEA)
Future Trends
🚀 Self-Supervised Learning (SSL): Models like Geneformer use pre-trained
representations from large-scale genomics data.
🚀 Multi-Modal AI: Integrating genomics, epigenomics, and transcriptomics for better
mutation effect prediction.
🚀 Explainable AI (XAI): Making deep learning more interpretable in genomic applications.
Would you like me to focus on a specific architecture in more detail? 😊