Science Reviews - Biology, 2024, 3(1), 9-15 Martina Elena Tarozzi

Artificial intelligence for Next generation sequencing

data analysis

Martina Elena Tarozzi, indipendent researcher. Florence, Tuscany, Italy

https://doi.org/10.57098/SciRevs.Biology.3.1.2

Received March 22, 2024. Revised April 09, 2024. Accepted April 10, 2024.

Abstract: In the rapidly evolving field of genomics, our capacity to decipher genetic data encoded in DNA has

been transformed by Next Generation Sequencing (NGS) technologies. These advanced technologies produce

an enormous volume of data, posing substantial challenges in extracting meaningful biological insights.

Artificial intelligence (AI) algorithms offer distinctive possibilities to unravel the biological information

embedded in such extensive and intricate datasets. This review offers a synopsis of AI classifications and

algorithms, elucidating how these techniques can be employed on sequencing data. Subsequently, a selection of

the most typical, promising, or illustrative applications of AI on NGS data to tackle unresolved technical or

biological issues are showcased.

Introduction

Artificial intelligence (AI) algorithms and se-

quencing technologies represent two groundbreak-

ing innovations that witnessed outstanding im-

provements in the last few decades. In both AI and

sequencing technologies, the first ilestones date

back to the early ‘50s, and the rapid and simultane-

ous advancements in the two fields resulted in the

new hybrid research branch of computational biol-

ogy[1]. The large and complex datasets produced

by sequencing experiments contain the information

needed to understand many unanswered biological

and medical questions, but that information is often

difficult to extract. As the biomedical sector is be-

coming more data-intensive and AI algorithms

more able to handle biological complexity, the inter-

connection between these two research fields is

bound to strengthen.

Overview of artificial intelligence, machine

learning and deep learning

Artificial Intelligence (AI) is a broad term that

covers a plethora of computational approaches and

algorithms able to mimic cognitive abilities (Figure

1). The term Machine Learning (ML) refers to sev-

eral algorithms able to perform different tasks, such

as pattern recognition, classification and prediction

tasks based on models derived from existing data.

The key feature of these methods is the absence of

coded algorithms given by the developer to de-

scribe the steps towards which input data are trans-

formed in output results. The ML method therefore

learns from the data to create a hierarchy of con-

cepts, each one deﬁned by its relation to simpler

concepts, that it uses to perform a task. By building

knowledge from previous data and training, this

approach avoids the need for the developer to ex-

plicitly define all the knowledge that the machine

needs. The goal of many ML tasks is to optimize the

model performance so that they can be generalized

on independent datasets (generalization perfor-

mance).

Martina Elena Tarozzi Science Reviews - Biology, 2024, 3(1), 9-15

Figure 1: Schematic representation of differences and shared features between Artificial Intelligence (AI), Machine

Learning (ML) and Deep Learning (DL).

ML methods are usually defined either as su-

pervised or unsupervised. In unsupervised learn-

ing, no predefined labels are provided for the ob-

jects under study. Here, the goal is to explore the

data and discover similarities between objects (e.g.,

samples) based solely on the input data. Clustering

and most dimensionality reduction techniques rep-

resent examples of unsupervised algorithms. Unsu-

pervised methods are commonly used in explora-

tory data analysis and for quality control tasks to

identify potential issues such as outliers and batch

effects, as well as to discover recurrent patterns and

unknown sources of variation in highly dimen-

sional datasets [2,3]. Examples of unsupervised di-

mensionality reduction techniques are principal

component analysis (PCA), t-distributed stochastic

neighbor embedding (t-SNE)[4] and Uniform Man-

ifold Approximation and Projection (UMAP) [5].

PCA is a fast, linear transformation, which maps the

data over a space whose coordinates are linear com-

binations of input features that capture most of the

variance of the data. This algorithm is usually pre-

ferred when the aim is to separate the data points as

far as possible. On the other hand, methods such as

UMAP or t-SNE rely on more complex

mathematical assumptions and steps, to “guess” the

manifold on which data are located. Such methods

preserve only local similarities and hence produce a

higher clusterization of the data in the embedding

space, with data subpopulations more separated

among themselves compared to PCA embeddings

[6,7]. Clustering methods are used to identify in an

unsupervised way groups of similar data points

based on the measure of similarity of choice. Exam-

ples of commonly used clustering algorithms in se-

quencing data analysis are hierarchical clustering

with dendrograms, K-means and Density-Based

Spatial Clustering of Applications with Noise

(DBSCAN) [8].

On the other hand, supervised learning in-

volves using labeled data, where each input has an

associated output. This allows the algorithm to

learn a set of rules to predict the correct output for

new input data based on its features, attributes, or

labels. If the output is qualitative, then the process

is called classification [9,10]. Otherwise, in the case

of quantitative values, it is called regression. Well-

trained ML models can learn rules about the under-

lying patterns and relationships in the data that can

then be applied to make accurate predictions on

Science Reviews - Biology, 2024, 3(1), 9-15 Martina Elena Tarozzi

new, unseen data. These rules can display new in-

sights into the relevant features used to correctly

identify the studied classes[11]. Examples of classi-

fiers are decision trees, random forests, support vec-

tor machines and neural networks. Examples of re-

gression methods are linear or generalized linear

models, with the possible addition of random ef-

fects (mixed-effect models), penalization terms

(regularized models), non-linearity (e.g., kernel re-

gression), or basis functions (e.g., splines). Never-

theless, it is worth underlying that supervised and

unsupervised learning are not formally deﬁned

terms, and the boundaries between them are often

blurred since many ML methods can be used to per-

form both types of tasks[12].

Neural networks are ML algorithms consist-

ing of interconnected artificial neurons, which rep-

resent the building blocks of neural networks and

deep learning algorithms[13]. Artificial neurons

have as input a vector of values and compute

weighted sums of these values followed by a non-

linear transformation. The activation function de-

fines whether the input values reach the threshold

needed to activate the neurons and consequently

determines the output of the node given the set of

inputs received. The weights are parameters that

are adjusted during the training step, as learning

proceeds[11]. The term Deep Learning (DL) refers

to multi-layered artificial neural networks with a

high number of hidden layers used to extract pro-

gressively higher-level features from data. The first

layer of neurons, also referred to as the input layer,

is the one that receives the experimental input data,

followed by layer two, made of neurons that receive

as input the outputs of layer one, and so forth for

deeper hidden layers[13]. The goal of these

networks is to model a function f. In a classifier, for

example, the function y = f (x) maps an input x to a

category y. A feed-forward network defines a map-

ping y = f (x; θ) and learns from the training set the

value of the parameters θ to model the function.

Training a neural network involves optimizing its

parameters (weights) to minimize a certain error

metric, which is typically defined by a loss function.

Such loss function measures the difference between

the predicted output of the neural network and the

actual output. Optimal weight updates are enabled

by backpropagation, which is a well-established al-

gorithm in neural network training, as it enables the

computation of the loss function's gradient with re-

spect to the network's weights.

Neural networks can be categorized into three

main types based on the type of architecture in

which neurons are organized: feed-forward, recur-

rent, and convolutional (Figure 2). In feed-forward

networks, the connections between nodes do not

form a cycle, and the information proceeds exclu-

sively forward from the input layer to the hidden

layers and lastly to the output layer. On the contrary,

recurrent neural networks have connections that

form cycles, allowing the output of a node to affect

subsequent input to the same nodes. Convolutional

neural networks are made of convolution kernels

processing input data, followed by pooling layers

simplifying the information to its most meaningful

concepts, and ultimately followed by hidden fully

connected layers for further data processing, like for

example a prediction task. The ultimate output of

the neural network represents its prediction or clas-

sification of the input data, which is built based on

its experience of recurrent patterns learned from the

data.

Figure 2: Schematic representation of the three main types of neural networks: feed-forward neural network (a), recurrent neural

networks (b) and convolutional neural networks (c). In all three representations, orange circles stand for the input neurons and

green circles for the output neurons. In panel a, blue circles represent the hidden layers. In panel b, light blue circles indicate

recurrent hidden layers, while in panel c, yellow circles indicate the kernels and the gray 3D squares represent convolutional layers.

Martina Elena Tarozzi Science Reviews - Biology, 2024, 3(1), 9-15

Applications of AI in genomics and transcriptomics

The complexity of data generated by high-

throughput sequencing technologies can make tra-

ditional analysis methods insufficient for identify-

ing patterns and extracting insights.ML and DL

methods have been applied to sequencing data with

a vast number of scopes. Here we provide a selec-

tion of some of the most relevant fields of applica-

tion. This section aims at providing common, prom-

ising or exemplifying applications of AI methods on

NGS data in biology and bioinformatics, while it

should not be considered a complete overview of all

its possible applications in biology.

Liquid biopsies and personalized medicine

Liquid biopsies are minimally invasive diagnostic

methods that analyze bodily fluids, such as blood,

urine, or cerebrospinal fluid, to detect and monitor

diseases, and are especially relevant in early diag-

nosis of cancer and neurodegenerative diseases

[14,15]. These samples contain cell-free Nucleic Ac-

ids (cfNA), circulating tumor DNA (ctDNA), circu-

lating tumor cells (CTCs), exosomes, and other bi-

omarkers that allow the extraction of genomic, tran-

scriptomic and epigenomic information, which can

be used for early detection, monitoring of progres-

sion and support personalized therapeutic deci-

sions to target the disease [15]. These types of data

are extremely complex, subject to many confound-

ers and for most features characterized by a high

signal-to-noise ratio. AI algorithms have signifi-

cantly advanced data analysis and interpretation of

this data and consequently the whole field in sev-

eral aspects, such as in risk assessment and early di-

agnosis [16], disease subtype classification [17],

treatment response prediction [18] and in monitor-

ing minimal residual disease [19]. For example,

SVM were effectively used to predict the probabil-

ity of reoccurrence based on gene expression data

or specific gene signature in different types of can-

cers[20,21], improving the monitoring of the molec-

ular profile of the patient’s tumor and the predic-

tion of personalized treatments at different times.

Furthermore, ctDNA methylation patterns have

been extensively studied with several ML classifica-

tion or regression methods as well as with neural

networks to achieve effective early detection both in

cancer research[22] and in the context of neuro-

degenerative diseases[23]. In this context, AI

reaches some of the most notable results in terms of

tangible impacts in molecular biology and medicine,

and it is expected that its role in personalized med-

icine will increase in the near future.

Regulatory genomics

Regulatory genomics is the field of genomics that

studies gene expression regulation trying to iden-

tify regulatory regions (such as enhancers, promot-

ers, transcription start sites (TSS), and genome ac-

cessibility) and the regulatory hierarchy between

these regions and other genes. In this context, deep

learning and more specifically Convolutional Neu-

ral Networks have been applied with the best re-

sults. One of the commonly used architectures in-

volves treating the input DNA sequence as categor-

ical variables. Each position in the sequence is one-

hot encoded, resulting in a vector where only one

channel corresponds to the A-C-G-T nucleotides

(with a value of 1) provided to the input layer.

These kernels are followed by convolutional layers,

which simplify the information to extract the most

relevant concepts. Convolutional filters are initially

trained on specific regions of interest with known

regulatory properties. The knowledge gained by

the convolutional neural network (CNN) during

training can then be applied to new regions for ac-

curate predictions. This architecture has been suc-

cessfully applied to various types of sequencing

data, particularly in the context of epigenomic stud-

ies. This overall architecture has been used on dif-

ferent types of sequencing data and has provided

particularly interesting results in terms of epige-

nomic studies. For example, this type of architec-

ture has been applied to DNAase-seq data to pre-

dict cell-type specific regions of accessible chroma-

tin [24], to identify promoters and distal regulatory

regions along mammalian genomes [25], to predict

cell-type specific gene expression from DNA se-

quencing data and alterations of it associated to var-

iant alleles [26], and to identify genomic regions re-

sponsible for the three-dimensional chromatin fold-

ing in the nucleus [27] from genomic and Hi-C data.

Considering that both the experimental and compu-

tational technologies used in these studies are rela-

tively young, this is arguably one of the most prom-

ising research fields for the next decades, with the

potential to answer many of the open questions in

functional genomics.

Improvement of genome editing specificity

Science Reviews - Biology, 2024, 3(1), 9-15 Martina Elena Tarozzi

During the past decade, technological innovations

in molecular biology have made genome editing

easier, allowing the modification of the DNA se-

quence at a single nucleotide resolution[28]. The

most successful technique to perform genome edit-

ing is CRISPR/cas9, where the identification of the

target genomic region is mediated by a guide RNA

(gRNA). The gRNA is a chimeric RNA consisting of

a ~20nt guide sequence that identifies through base

complementarity the target site in the genome and

precisely directs the Cas9 protein to it. Some mis-

matches in the guide sequence can be tolerated and

do not affect the ability to align and cut DNA, re-

sulting in off-target cleavages [29,30]. Accurate

gRNA design maximizes on-target efficacy (sensi-

tivity) and minimizes off-target effects (specificity).

ML and DL models have been used in this context

to predict gRNA sequencing with high sensitivity

and specificity, and several specific tools have been

released in the last few years, for example, Deep-

SpCas9 [31], DeepCRISPR [32] , DeepCpf1 [33],

CRISPRscan [34], among many other. These tools

differ in terms of models and network architectures,

nevertheless, the fundamental overall architecture

described in the previous section still applies also in

this context. Of course, other more complex net-

works exist, like the one used by DeepCRISPR[32],

where a hybrid deep neural network is used com-

bining unsupervised and supervised representation

learning to model single gRNAs using a set of ge-

nome-wide sgRNAs. Despite the outstanding re-

sults already achieved, this type of application of

ML and DL on sequencing data still has some rele-

vant limitations [35]. In this context, the large

amount of data needed to train prediction models is

not always available, and quite often the available

data show some challenges caused by the heteroge-

neity of sequencing platforms and cell types.

Future Perspectives

Both computational and molecular biology

are continuing to grow at an impressive pace, as

they did in the last decades. As shown in this review,

integrating AI, ML and DL techniques with NGS

data holds immense promise, but it also presents

challenges. For example, the quality and quantity of

NGS data pose hurdles— e.g. noise, biases, and ar-

tifacts- that can impact analyses and computational

resources and strain existing infrastructure due to

DL model demands. In addition, these models re-

quire a huge amount of data which is not always

available, therefore one of the currently most rele-

vant challenges in the field is addressing small sam-

ple sizes and class imbalance affects model robust-

ness. Another limitation is that interpretability re-

mains an issue: black-box models lack transparency

and the understanding of the molecular processes

underneath a given prediction or classification

based on nucleotides pattern often remains under

understood. Bridging the gap between AI expertise

and biological domain knowledge is critical and

could improve transferability and generalization

across diverse biological contexts as well. Last but

not least, ethical concerns arise from handling sen-

sitive genomic data. Finding a balance between the

extraordinary potential of AI in the medical field

and the need for a careful safeguard of individual

privacy are really hot topic involving both law and

tech experts which currently remains an open issue.

In summary, while AI, ML, and DL offer exciting

prospects for genomics research, overcoming these

challenges is crucial for realizing their full potential.

Taken together these innovations are a new frontier

of research and have the potential to strongly affect

our possibilities to understand and interact with ge-

netic information, which will also require ethical

considerations.

References

1. Muir P, Li S, Lou S, et al. The real cost of sequencing: Scaling computation to keep pace with data

generation. Genome Biol. 2016; 17:1–9

2. Jia W, Sun M, Lian J, et al. Feature dimensionality reduction: a review. Complex Intell. Syst. 2022

83 2022; 8:2663–2693

3. Meng C, Zeleznik OA, Thallinger GG, et al. Dimension reduction techniques for the integrative

analysis of multi-omics data. Brief. Bioinform. 2016; 17:628–641

4. Hinton G; L van der M. Visualizing Data using t-SNE. Ann. Oper. Res. 2014; 219:187–202

Martina Elena Tarozzi Science Reviews - Biology, 2024, 3(1), 9-15

5. McInnes L, Healy J, Melville J. UMAP: Uniform Manifold Approximation and Projection for

Dimension Reduction. 2018;

6. Yang Y, Sun H, Zhang Y, et al. Dimensionality reduction by UMAP reinforces sample

heterogeneity analysis in bulk transcriptomic data. Cell Rep. 2021; 36:

7. Li W, Cerise JE, Yang Y, et al. Application of t-SNE to human genetic data. J. Bioinform. Comput.

Biol. 2017; 15:1–14

8. Jin J, Wang H, Shu Z, et al. A Density-Based Algorithm for Discovering Clusters in Large Spatial

Databases with Noise. Cailiao Yanjiu Xuebao/Chinese J. Mater. Res. 2017; 31:219–225

9. Xu C, Jackson SA. Machine learning and complex biological data. Genome Biol. 2019; 20:76

10. Bzdok D, Krzywinski M, Altman N. Points of significance: Machine learning: Supervised

methods. Nat. Methods 2018; 15:5–6

11. Murphy KP. Machine Learning A Probabilistic Perspective. MIT Press Cambridge, Massachusetts

London, Engl. 2018; 16:

12. Omta WA, Heesbeen RG van, Shen I, et al. Combining Supervised and Unsupervised Machine

Learning Methods for Phenotypic Functional Genomics Screening:

https://doi.org/10.1177/2472555220919345 2020; 25:655–664

13. Zou J, Huss M, Abid A, et al. A primer on deep learning in genomics. Nat. Genet. 2019; 51:12–18

14. Gong X, Zhang H, Liu X, et al. Is liquid biopsy mature enough for the diagnosis of Alzheimer’s

disease? Front. Aging Neurosci. 2022; 14:891

15. Lone SN, Nisar S, Masoodi T, et al. Liquid biopsy: a step closer to transform diagnosis, prognosis

and future of cancer treatments. Mol. Cancer 2022 211 2022; 21:1–22

16. Liu L, Chen X, Petinrin OO, et al. Machine learning protocols in early cancer detection based on

liquid biopsy: A survey. Life 2021; 11:1–39

17. Peneder P, Stütz AM, Surdez D, et al. Multimodal analysis of cell-free DNA whole-genome

sequencing for pediatric cancers with low mutational burden. Nat. Commun. 2021; 12:

18. Gu F, Wang X. Analysis of allele specific expression-a survey. Tsinghua Sci. Technol. 2015;

20:513–529

19. Im YR, Tsui DWY, Diaz LA, et al. Next-Generation Liquid Biopsies: Embracing Data Science in

Oncology. Trends in Cancer 2021; 7:283–292

20. Zhou J, Li L, Wang L, et al. Establishment of a SVM classifier to predict recurrence of ovarian

cancer. Mol. Med. Rep. 2018; 18:3589–3598

21. Xu G, Zhang M, Zhu H, et al. A 15-gene signature for prediction of colon cancer recurrence and

prognosis based on SVM. Gene 2017; 604:33–40

22. Constantin N, Sina AAI, Korbie D, et al. Opportunities for Early Cancer Detection: The Rise of

ctDNA Methylation-Based Pan-Cancer Screening Technologies. Epigenomes 2022; 6:1–27

23. Bahado-Singh RO, Radhakrishna U, Gordevičius J, et al. Artificial Intelligence and Circulating

Cell-Free DNA Methylation Profiling: Mechanism and Detection of Alzheimer’s Disease. Cells 2022;

11:1–19

24. Kelley DR, Snoek J, Rinn JL. Basset: Learning the regulatory code of the accessible genome with

deep convolutional neural networks. Genome Res. 2016; 26:990–999

25. Onimaru K, Nishimura O, Kuraku S. Predicting gene regulatory regions with a convolutional

neural network for processing double-strand genome sequence information. PLoS One 2020;

15:e0235748

Science Reviews - Biology, 2024, 3(1), 9-15 Martina Elena Tarozzi

26. Kelley DR, Reshef YA, Bileschi M, et al. Sequential regulatory activity prediction across

chromosomes with convolutional neural networks. 2018;

27. Fudenberg G, Kelley DR, Pollard KS. Predicting 3D genome folding from DNA sequence with

Akita.

28. Jinek M, Chylinski K, Fonfara I, et al. A programmable dual-RNA-guided DNA endonuclease in

adaptive bacterial immunity. Science (80-. ). 2012; 337:816–821

29. Doench JG, Hartenian E, Graham DB, et al. Rational design of highly active sgRNAs for CRISPR-

Cas9–mediated gene inactivation. Nat. Biotechnol. 2014 3212 2014; 32:1262–1267

30. Doench JG, Fusi N, Sullender M, et al. Optimized sgRNA design to maximize activity and

minimize off-target effects of CRISPR-Cas9. Nat. Biotechnol. 2015 342 2016; 34:184–191

31. Kim HK, Kim Y, Lee S, et al. SpCas9 activity prediction by DeepSpCas9, a deep learning–based

model with high generalization performance. Sci. Adv. 2019; 5:

32. Chuai G, Ma H, Yan J, et al. DeepCRISPR: Optimized CRISPR guide RNA design by deep

learning. Genome Biol. 2018; 19:1–18

33. Kim HK, Min S, Song M, et al. Deep learning improves prediction of CRISPR–Cpf1 guide RNA

activity. Nat. Biotechnol. 2018 363 2018; 36:239–241

34. Moreno-Mateos MA, Vejnar CE, Beaudoin JD, et al. CRISPRscan: designing highly efficient

sgRNAs for CRISPR-Cas9 targeting in vivo. Nat. Methods 2015 1210 2015; 12:982–988

35. Wang J, Zhang X, Cheng L, et al. An overview and metanalysis of machine and deep learning-

based CRISPR gRNA design tools. RNA Biol. 2020; 17:13