UC Riverside

UC Riverside Electronic Theses and Dissertations

Title
Applications of Genetic Algorithms in Bioinformatics

Permalink
https://escholarship.org/uc/item/9087560g

Author
Piserchia, Zachary

Publication Date
2018

Peer reviewed|Thesis/dissertation

eScholarship.org Powered by the California Digital Library

Applications of Genetic Algorithms in Bioinformatics

A Thesis submitted in partial satisfaction

Master of Science

in

Genetics, Genomics and Bioinformatics

by

Zachary T. Piserchia

December 2018

Thesis Committee:
Dr. Daniel Koenig, Chairperson
Dr. Connie Nugent
Dr. Jaimie Van Norman
 Copyright by
Zachary T. Piserchia
2018
The Thesis of Zachary T. Piserchia is approved:

Committee Chairperson

University of California, Riverside

Applications of Genetic Algorithms in Bioinformatics

Zachary T. Piserchia
Master of Science, Graduate Program in Genetics, Genomics and Bioinformatics
University of California, Riverside, December 2018
Dr. Daniel Koenig, Chairperson

The field of Bioinformatics has advanced rapidly in recent years. Breakthroughs

in sequencing technology have driven an explosion in the production of genetic data. As

a result, the development of methods to interpret these data accurately and in a timely

manner is a major bottleneck in bioinformatics research. This task is further complicated

by imperfect knowledge of genetic features, as performance-altering assumptions can

easily be introduced during algorithm development, and methodology may quickly

become obsolete. Genetic algorithms are an evolution-inspired class of machine learning

algorithms that show great promise to resolve these problems. These algorithms

gradually refine solutions through natural selection, evolving a solution to a problem in

bioinformatics rather than manually designing a search strategy. Due to this learning

process determining how genomic features are identified, these genetic algorithms do not

rely on human knowledge of the problem. Consequently, biases are largely limited to the

data used to determine how fitness is evaluated. Genetic algorithms also make better use

of computational resources by reducing search space and utilizing parallel computation.

In order to examine the potential for genetic algorithms to facilitate research in the field,

iv
this work explores many different implementations of these algorithms in bioinformatics.

These genetic algorithm-based tools outperform existing methodology for many different

problems, and their predictions can be used to guide experimental research.

v
 TABLE OF CONTENTS

Abstract iv

List of Figures viii

Introduction 1

Genetic Algorithms Background 4

Genomics 11

Gene finding 11

Modelling Disease Risk 14

Multiple Alignment 18

Promoter Prediction 19

Transcription Factor Binding Sites 20

Determining Groups for Expression Analysis 23

Transcriptomics 25

Prediction of RNA Secondary Structure 25

Prediction of RNA-RNA Interactions 28

Discovery of Gene Regulatory Networks 29

RNA Editing Site Prediction 29

Proteomics 30

Protein Structural Prediction 31

Drug Discovery 32

vi
Discussion 34

Advantages of Genetic Algorithms 34

Weaknesses of Genetic Algorithms 36

Conclusion 38

Bibliography 39

vii
 LIST OF FIGURES

Figure 1 - Crossover in Genetic Algorithms 9

Figure 2 - Mutation in Genetic Algorithms 10

viii
Introduction

Growth in the field of DNA sequencing has generated a tremendous amount of

genetic data, but much of it has only been superficially examined (Galperin and Koonin

2010). As a result, developing methods to decipher this information is one of the largest

undertakings in the field of bioinformatics. Applications in bioinformatics largely focus

on the discovery and annotation of genomic features from sequence data, as well as the

characterization of genetic variants within populations (Manzoni et al. 2018) (Horner et

al. 2010). While there is often a focus on genes, investigating regulatory elements such

as promoter sequences is an equally important task. Problems in transcriptomics include

the prediction of RNA structure from sequence data, the discovery of RNA interactions,

and determining interactions between expressed genes (Manzoni et al. 2018).

Bioinformatics can be used in proteomics research to predict protein structures from

sequence data but is also closely associated with pharmacology. Drug discovery and

development can be greatly facilitated by computational predictions (Mandal, Johnson,

Wu, and Bornemeier 2007). It is crucial for these areas of scientific research to

accurately and efficiently analyze sequence data in order to provide meaningful

predictions.

Despite the great value of these applications, interpreting sequence data is

incredibly difficult. The extreme complexity and size of this data necessitates the use of

computational approaches, but even these methods face challenges. Accuracy and

runtime are principal concerns in the design of analysis algorithms, and improving one

1
often comes at the cost of the other (Ruffalo, LaFramboise, and Koyutürk 2011). In

addition, traditional programming based on human knowledge of genetics problems

would inherit the assumptions and biases of our currently imperfect understanding of

genomic data.

Machine learning approaches present a strong opportunity to mitigate these

problems. While this class of algorithms is not completely free of bias, they can be used

to infer the rules required to identify genomic features from previous data. Consequently,

the strength of machine learning to predict the presence and function of genomic

structures is limited by data. Providing larger data sets reduces bias and error rate,

improving the predictions of machine learning approaches (Brain and Webb 2000). It

follows that as more genomic features become empirically verified, computational

learning predictions will also improve. While many approaches to machine learning

exist, three major approaches are Neural Networks (NNs), Support Vector Machines

(SVMs), and Genetic Algorithms (GAs). Of these approaches, Genetic Algorithms are

particularly suitable for bioinformatics applications, as they require little a priori

knowledge of a problem they are applied to (Spears and De Jong 1990).

Neural networks are a machine learning method that is modeled after neurons in

the brain. By using a series of interconnected nodes and weights representing action

potentials, NNs can make decisions based on input data (Gómez-Ramos and Venegas-

Martínez 2013). As this closely resembles the biological pathways for sensation and

perception, NNs are extremely adept at image recognition and identifying patterns

2
(Schmidhuber 2015). Generally, NNs use supervised learning called backpropagation

where a training set with a known solution is used to learn the weights of individual

neurons. This method proceeds through several layers of neurons, and as such is a very

slow process (Dharwal and Kaur 2016). However, once these weights have been

learned, NNs can perform tasks extremely quickly. PilotNet, a 9-layer NN has been able

to successfully steer and drive a vehicle without human input (Bojarski et al. 2017). In

this example, the neural network uses pixels from video feed to determines the patterns

constituting the road and traffic to steer the vehicle accordingly (Bojarski et al. 2017).

Neural networks have been applied to problems in bioinformatics, but the runtime cost of

training NNs is a significant constraint on throughput. NNs can also be combined with

genetic algorithms to improve performance.

Support vector machines (SVMs) are a learning approach applied to classification

problems. SVMs work through the use of training sets to calculate functions that

separate data into two or more groups (Devi Arockia Vanitha, Devaraj, and Venkatesulu

2015). SVMs are very effective classifiers, and can be used to solve problems where data

is not linearly separable through transformation (Winters-Hilt and Merat 2007). In

bioinformatics, SVMs have been used to identify genes from unknown sequence data.

GISMO is one such approach that uses a SVM in order to identify potential genes

(Krause et al. 2007). GISMO learns oligonucleotide positions and frequencies (motifs)

associated with genes in training data, and then uses this information to identify genes in

unknown sequence data (Krause et al. 2007). This method can also be improved through

the use of Genetic Algorithms to optimize initial tuning parameters.

3
Genetic Algorithms Background

Genetic algorithms (GAs) are an approach to machine learning with roots in

evolution and genetics. Rather than relying on human knowledge to develop a program,

GAs “evolve” a solution to problems using the principles of natural selection. Similar to

evolution in biology, a population of algorithms is evaluated to determine individual

“fitness.” Individuals in the population represent different parameters for a program, and

their fitness is ranked based on how well the resulting program performed (Holland

1992). A simulation of reproduction is then used to produce the next generation, with

high-fitness individuals contributing more to the gene pool (Beasley, Bull, and Martin

1993). Over many of these generations, the average fitness of the population gradually

improves through evolution and the algorithm arrives at a solution (Beasley, Bull, and

Martin 1993). While this general framework forms the basis for GA approaches, many

different components of the algorithm must be considered in detail to successfully apply

them to a problem.

The first step in GA development is to determine how the genotype of individuals

in the population will be defined. Rather than using DNA, GAs use binary strings called

“chromosomes” to represent an individual’s genes. These 0’s and 1’s are then translated

into parameters and other components (i.e. genes) of a program, which are then used to

solve the problem to which the GA is being applied (Holland 1992). For example, a

chromosome in a pathfinding GA in robotics might include parameters such as the speed

at which the robot will move, and at what distance from an obstacle it should be before

4
turning. In this case, the “genes” could be translated from binary to represent a decimal

number for speed and/or distance. In order to do this, the binary values would need to be

encoded and decoded (Haupt R. and Haupt S. 2004). Quantization is one method of

encoding that takes non-overlapping values in a range and assigns them a unique binary

value. This allows them to represent a larger range of values than pure binary, and also

makes decimal numbers possible (Haupt R. and Haupt S. 2004). Choosing how to

represent the problem and how many bits will be required for the chromosome is

particularly important, as it will affect the performance of the algorithm where the

maximum error will be equal to the intervals chosen during quantization (Haupt R. and

Haupt S. 2004). Longer chromosomes may (or may not) be helpful in yielding a robust

solution, but will greatly increase runtime costs. Conversely, shorter chromosomes may

result in the GA running faster, but may result in a solution that is too simple and cannot

be widely applied outside test data (Kim and Weck, 2005).

Populations in GAs consist of a static number of individual chromosomes, which

are repeatedly replaced by those of the next generation. Population size must be carefully

considered in this step, as small populations are extremely vulnerable to genetic drift

(Rogers and Prugel-Bennett 1999). Because the forces of drift can undermine expected

evolutionary changes and decrease fitness, a small population size or selecting too few

parents can lead to drift and cripple learning in GAs (Rogers and Prugel-Bennett 1999).

In contrast, using a large population size increases runtime costs. As GAs’ runtime is

primarily derived from population size, finding a balance between learning rate and

runtime is needed for a successful GA (Oliveto and Witt 2014).

5
 While chromosomes define genotype in GAs, and mutation and crossover provide

genetic variation, learning cannot occur unless we can evaluate phenotype in some way to

derive a fitness value. This is accomplished through implementing a fitness function,

which takes in an individual’s chromosome and assigns it a fitness value (Manning,

Sleator, and Walsh 2013). This involves translating the chromosome into parameters for

a program (as described in chromosome design) and then actually running the program to

arrive at a solution. The solution is then judged on its quality and accuracy to arrive at a

fitness value. These determinations can be made in a variety of ways, but two of the

most common are to rank members of the population in relation to the individual with the

highest fitness, or to rank according to the average fitness of the population (Blickle and

Thiele 1996). In a robotics example, a robot solving a maze might be assigned a fitness

value based on how close it got to the exit, and the time it took for the robot to do so.

This step is absolutely crucial for the success of a GA as it is the main learning step. It is

also the most runtime intensive part of GAs as many individuals must be evaluated every

generation.

Generations and repopulation are another core part of GAs, as they control the

length of the learning period and how the gene pool of the next generation is formed. As

a rule, the initial population of a GA starts with randomized chromosomes; this starts the

population off with great diversity (Diaz-Gomez and Hougen 2007). To illustrate the

importance of this rule, consider a graph where all possible solutions have their fitness

ranked (a fitness landscape). This graph would have many hills and valleys, representing

different solutions to the problem with varying degrees of effectiveness. If a GA were to

6
start with a monotypic population, there is potential for the algorithm to get stuck

between two small local hills and miss a substantially larger global hill (Leung and Wang

2001). The algorithm would not explore solutions beyond the local peaks due to the

fitness decline after leaving them. Consequently, the algorithm could accept a solution

substantially less effective than the optimal (Leung and Wang 2001). Starting with a

random population greatly ameliorates this problem by seeding many different “start

points” for learning to branch from. This does not guarantee an optimal solution will be

reached, but greatly increases the probability of reaching a near-optimal solution.

After fitness evaluations, reproduction occurs to create the population for the next

generation. Although several variations of reproduction exist, the most common is to

take the two individuals with the highest fitness to act as the parents for the next

generation. Through crossover between the two parents, offspring chromosomes are

generated until the population size is reached (Holland 1992). Single-point, two-point,

and multi-point crossover can occur and similar to biological crossover, the parental

chromosomes exchange material in regions adjacent to the crossover site(s) (see Figure 1)

(Spears 1995). In order to maintain diversity and also allow for the introduction of novel

genotypes, mutation can also occur. Mutation in GAs usually occurs at a fixed rate, and

changes a single bit: a 0 becomes 1 or vice versa (see Figure 2) (Holland 1992).

Choosing mutation rate is also an important part of GA implementation, as it affects the

learning rate of the algorithm. This occurs as new solutions are explored through the

introduction of novel genotypes produced by these mutations (Haupt, 2000). If the

mutation rate is too low, it will take an extremely large number of generations to

7
thoroughly explore the solution space. Conversely, a mutation rate that is too high may

result in fast initial learning rates, but fail in the long term by inhibiting convergence

(Haupt, 2000). A final consideration is determining when to stop the algorithm and use

the solution it has reached. Ensuring an accurate solution is found and mitigating runtime

costs are both important factors in this decision, and this choice may vary based upon the

problem. As such, several methods exist to terminate GAs, but the most common are

setting limits to the number of generations and convergence (Bhandari, Murthy, and Pal

2012). Running for a fixed number of generations is a common approach, but also relies

on the assumption that the algorithm has run for enough time to achieve a meaningful

solution. The convergence method stops the algorithm once the population has become

the same genotype and no new major changes occur, as this means all individuals have

the same fitness and are no longer improving. Convergence is taken as evidence that the

best solution has been found and the algorithm stops (Bhandari et al. 2012).

8
Figure 1: Single crossover occurs between two parents
at a point selected at random to produce offspring

9
Figure 2: Mutation occurs after crossover,
resulting in a completely novel genotype

10
Genomics

The increased the pace of genomics research is largely due to massively parallel

sequencing techniques, such as Illumina, which have resulted in a wealth of genetic data.

Annotating and deciphering these new data is one of the greatest undertakings in the

field, and genetic algorithms may greatly aid these efforts. GAs greatly benefit from the

increased availability of data, drawing upon more potential knowledge to make their

predictions. The automated nature of machine learning also requires minimal

experimental research, greatly streamlining data analysis. This makes GAs an ideal

approach for solving problems in genomics.

Gene-finding

One of the first steps in interpreting genome sequences is the identification of

genes. Several models for gene prediction exist, including Position Weighted Matrices

(PWMs), Spliced Alignment, and Hidden Markov Models (HMMs). PWMs align

unknown sequences with those in a database to make gene predictions. By using

conserved sequence sites at specific positions, the PWM can be used to predict gene

features based on the alignment with functionally related sequences (Mathé, Sagot,

Schiex, and Rouzé 2002). Glimmer3 (a gene-finding software used in prokaryotes) uses

a PWM in order to identify ribosomal binding sites to help identify the start of a gene

(Delcher, Bratke, Powers, and Salzberg 2007). Spliced alignment splits an unknown

sequence into many pieces and aligns it to known sequences in a database. By

11
discovering similarities with known coding sequences, these alignments can be used to

predict the presence of genes. However, this method suffers from inaccuracy in how the

predictions are grouped. Often, a single gene may be split into several predictions, or

multiple genes can be lumped together and predicted as one (Mathé et al. 2002). HMM

approaches use training data to learn the probability of a particular base (A, T, C, or G)

appearing at the current position in a sequence based on the identity of previous

nucleotides (Mathé et al. 2002). As individual nucleotides often do not provide enough

information, groups of nucleotides or “k-mers” are often used. If a k-mer at the current

position is extremely improbable, it may represent a state transition (such as the sequence

transitioning from exon to intron). Since exons and introns have statistically different

nucleotide composition, these “hidden states” can be learned by the HMM. Through the

use of training data sets, the algorithm can mark boundaries between coding and

noncoding sequences to find genes (Mathé et al. 2002). Examples of HMM-based gene

prediction software include GENSCAN (Burge and Carlin 1997), AUGUSTUS (Stanke

and Morgenstern 2005), and HMMGene (Krogh 1997).

While all of these methods have successfully made gene predictions, most have

several flaws. Current gene prediction software only finds transcribed regions of genes,

and skips over or mislabels valuable regulatory features such as promoter sequences

(Mathé et al. 2002). Comparative methods often struggle with the definition of

“similarity;” as there is no general consensus on criteria for similarity, arbitrary

guidelines are often used. This can result in comparative approaches predicting too many

or too few genes, and identifying which genes are false positives is not readily apparent

12
(Mathé et al. 2002). Furthermore, it is difficult to rely upon evolutionary comparisons for

the discovery of novel genes, as they may be substantially different from documented

genes in a database. Long genes, long introns, and overlapping genes all pose difficulties

in the developing gene prediction software, as they require identification of “special

cases” (Mathé et al. 2002). Consequently, there is room for further improvement in gene-

prediction through the use of GAs.

A GA approach to finding genes and annotating genomes was created by

Chowdhury et al. (2016). Because over- or under- predicting genes often complicates the

process of annotation, the problem was broken down into a sub-problem of finding exons

(Chowdhury, Garai A., and Garai G. 2016). The authors decided to use homology with

known exons in a database as a means to predict new exon boundaries from unknown

genomic sequence data. To do this, chromosomes in the GA delineated possible exon

boundaries, and fitness was evaluated using alignment scores to database exons

(Chowdhury et al. 2016). The resulting approach was not only successful in a test case

on human chromosome 21, but was also able to predict exons more accurately than

GENSCAN, a well-known and widely used annotation tool (Burge and Carlin 1997).

The success with this method demonstrates that the application of GAs may greatly

advance efforts in the field.

A genetic algorithm approach by Hwang et al. (2013) can be used to predict

which genes are essential. By definition, essential genes are required for an organism to

live, meaning identifying them through lab experiments is difficult. While genome-wide

13
knockout experiments have been successful in identifying many essential genes, this

process is extremely costly and time intensive (Hwang, Ha, Ju, and Kim 2013). Due to

these issues, in silico prediction of essential genes is a desirable alternative. In order to

accomplish this, the set of all an organism’s genes must be classified into essential and

non-essential groups. To do this, the authors chose to combine GAs with the area under

a ROC (receiver operating characteristic) curve. ROC curves compare a classifier’s true

positive rate (Y axis) against false positive rate (X axis) for many different settings. As a

high true positive rate and low false positive rate are required for accurate predictions, the

area under the curve or AUC demonstrates how effective a given classifier is. For this

reason, the authors chose to use AUC as a measurement of fitness in the GA. By using a

training set of known essential and non-essential genes, the GA was successfully able to

optimize a linear discriminant classifier (Hwang et al. 2013). The resulting GA_AUC

classifier was highly successful, and outperformed other popular classification methods

including logistic regression, polynomial kernels, RBF (radial basis function) support

vector machines, and multilayer perceptrons (Hwang et al. 2013).

Modelling Disease Risk

While important, the discovery of genes and regulatory elements alone still

presents incomplete genetic information. In order to learn more practical information in

the form of phenotypic predictions, interactions between genomic features must also be

considered. Due to the exponentially large number of possible interactions, as well as the

multidimensional nature of the data involved, statistical models such as linear regression

14
have very little power to predict these associations (Moore, Hahn, Ritchie, Thornton, and

White 2004). For the same reasons, these predictions are also extremely computationally

intensive: solving this problem efficiently would require a reduction in the size of search

space involved. GAs can be used to work around this problem as their populations

explore many possible solutions to the problem, but are not a comprehensive search. For

this reason, Moore et al. (2004) used GAs in order to develop a method for predicting

models of disease risk.

In order to model interacting genotypes, penetrance functions can be used to

predict the probability that a particular genotype will have a disease outcome. By

manipulating these functions, a model first developed by Frankel and Schork (1996) can

be used for multiple genes and alleles (Moore et al. 2004). Using this model, it is

possible to predict disease risk for a given genotype if the set of penetrance probabilities

is known. To calculate these values, as well as discover new and interesting models of

disease risk, Moore et al. (2004) used a genetic algorithm with a fitness function that

maximized the effect of gene interactions while minimizing the effect of individual

genotypes. After running the resulting algorithm was successful at finding 1,000

different models for disease risk, including 3, 4, and 5 interacting gene models (Moore et

al. 2004). The potential for GAs to make these complex predictions on phenotype shows

great promise for the future of personalized medicine.

Another approach by Yang et al. (2004) combined GAs with a local search

algorithm to study the influence of single nucleotide polymorphisms (SNPs) on disease

15
and predict genes associated with disease. This is a difficult problem, as multiple

interacting SNPs may alter disease outcome, and SNPs with negligible effects on

phenotype individually may collectively cause disease (Yang, Moi, Lin, and Chuang

2004). In addition, considering increased numbers of SNPs exponentially increases the

computational complexity of the problem, to the point where traditional statistical

methods become infeasible (Yang et al. 2004). Different types of genetic interactions

must also be considered for accurate predictions; however, this increases complexity even

further. For this reason, the authors investigated two main genetic models: a ZZ model,

where two high risk alleles cause disease, and an XOR model, where heterozygosity at

specific loci causes disease.

In order to discover interesting and novel patterns of disease, the GA’s fitness

function selected for SNPs that influenced disease more collectively, than they did

individually. To further improve the method, the authors also included a local search

algorithm to direct evolution within the GA. At the end of each generation, the local

search was used to examine possible crossover exchanges; if an exchange would improve

fitness, then it was carried out when creating the next generation (Yang et al. 2004). This

combined approach had two advantages: the GA helped to mitigate the complexity of the

problem and arrive at a solution faster, while the inclusion of the local search helped the

GA escape local optima. Compared to a GA-only implementation, the local search GA

was also found to improve the solution, finding more significant (based on chi-squared

values) genetic models for disease (Yang et al. 2004). While this did improve the

solution, the authors also noted that it did come with a slightly increased runtime cost, but

16
remained within the same order of approximation (Yang et al. 2004). It is likely that the

discrepancy between the local search GA and regular GA increases with the complexity

of model considered, making it a more suitable approach for complex interactions.

A final GA-based method for determining genes associated with disease was

developed by Tahmasebipour et al. (2015) These researchers noted that the commonly

used Genome Wide Association studies (GWAs) and comparative analysis approaches to

disease gene prediction are limited by their focus on predicting single disease genes

(Tahmasebipour K. and Houghten S. 2015). Because disease can be the result of many

interacting alleles, SNPs that have a low individual contribution to disease risk may have

a large effect in the presence of other specific alleles. In order to account for this

possibility, the authors to reframed the problem, representing disease association as a

network of nodes representing genes, mRNA, proteins, and other agents (Tahmasebipour

and Houghten 2015).

To do this, the authors implemented a GA that used a chromosome representing a

set of disease genes. Fitness was evaluated using a criterion of “collaboration”: sets

were ranked based upon interaction with other genes in the group, as well as interaction

with known disease genes based on known protein-protein interaction (PPI) networks

(Tahmasebipour and Houghten 2015). In order to ensure the algorithm did not discover

the same networks, a limit was placed on how many known PPI genes could exist

together in a single set. In this way, candidate disease genes are evaluated based on their

association with known disease genes, using the “guilt by association” principle

17
(Tahmasebipour and Houghten 2015). To test the algorithm, the authors used breast

cancer as an example, and the algorithm was able to successfully discover candidate

disease genes. These predictions were validated by Genotator (Wall et al. 2010), a tool

that ranks gene-disease associations based on clinical data. Several breast cancer gene

candidates found by the GA overlapped with the top 1% of Genotator’s candidates. In

addition, the GA was able to discover many candidates not identified by Genotator

(Tahmasebipour and Houghten 2015). The algorithm also had higher sensitivity than

CIPHER (Guzman and D'Orso 2017), a leading method for the disease-gene association

problem (Tahmasebipour and Houghten 2015).

Multiple Alignment

A separate, but equally important application of genomic data is to compare

evolutionary relationships between taxa. To make these comparisons and draw

meaningful conclusions, it is first necessary to align them. While programs like

ClustalW (Thompson, Higgins, and Gibson 1994) and Multalin (Corpet 1988) can

complete this task, there is still opportunity for the development and improvement of

algorithms for multiple alignment. A method developed by Nizam et al. (2011) uses a

cyclic GA to perform a multiple alignment (Nizam, Ravi, and Subbaraya 2011). By

using a fitness function that scores aligned residues, the algorithm is able to solve for the

gap positions. The resulting program was able to outperform ClustalW and Multalin by

generating multiple alignments with better scores.

18
 Multiple sequence alignment has a wide variety of applications, ranging from the

construction of phylogenetic trees to comparing protein structure (Nizam et al. 2011). As

these predictions are reliant upon the accuracy of the alignment, the improved Cyclic

Genetic Algorithm Multiple Sequence Alignment (CGA-MSA) approach will facilitate

efforts in these areas.

Promoter Prediction

The identification of genes is only a first step, as many other regulatory elements

affect gene expression in vivo. Promoter sequences are particularly useful information in

determining how a gene is expressed as they are directly associated with a downstream

gene (or genes in operon structures). Prediction of these elements is a notoriously

difficult problem due to high variability even within the same species. Furthermore,

extrapolating information from one species to another is impeded by the presence of

different regulatory architecture (Azad, Shahid, Noman, and Lee 2011). Search

approaches using motifs such as the TATA box and initiators exist, but these approaches

are prone to frequent false positives (Azad et al. 2011).

Previous approaches to promoter prediction have mainly focused upon the

identification of select features or signals. The presence of TATA or CAAT boxes, CpG

islands, known transcription factor binding sites, and pentamer motifs have all been used

to predict the presence of a promoter in unknown sequence data (Xie, Wu, Lam, and Yan

2006). Various forms of pattern recognizing algorithms have been applied to the

classification of promoters, including neural networks, discriminant analysis, and

19
individual component analysis. While they have had some success, these methods have

struggled due to the fact that individual sequence patterns indicating promoters cannot be

widely applied (Xie et al. 2006).

To combat these problems, Azad et al. (2011) developed a method combining

genetic algorithms with a support vector machine (SVM) to classify sequences into

promoter and non-promoter categories. In order to do this, a positive training set

containing verified promoters, and a negative training set containing non-promoter

sequences was provided to the SVM. By learning different weights for patterns (motifs)

in the data, the SVM learns to predict whether an unknown sequence is a promoter or not.

In order to determine which motifs should be used, the researchers used a GA. By using

a fitness function that weighted the probability of a random triplet being found in the

promoter dataset versus the non-promoter dataset, the GA determined which triplets

served as the best indicator of a promoter sequence. As triplets alone have limited

statistical power, the authors decided to use hexamers comprised of triplet pairs found by

the GA (Azad et al. 2011). The resulting method outperformed existing methods, having

a higher average sensitivity and specificity (Azad et al. 2011). The resulting

PROMOBOT tool remains a cutting edge tool for plant promoter identification with an

average sensitivity of 85% (Umarov and Solovyev 2017).

Transcription Factor Binding sites

In addition to promoter sequences, the identification of new Transcription Factor

Binding Sites (TFBSs) is also an important source of regulatory information. The

20
discovery of these sites is particularly difficult as few have been experimentally verified,

and the length of these sequences is particularly short (Jayaram, Usvyat, and Martin

2016). Due to this lack of data and the statistical limits that accompany short sequences,

it is difficult to use comparative methods such as homology searches on their own. This

possibility is further complicated by an unclear scope of sequence conservation: regions

flanking binding sites may also suggest common ancestry (Levitsky et al. 2007). Most

current TFBS prediction software use position weighted matrices (PWMs), which have

been suggested to be a state-of-the-art method for this task (Jayaram et al. 2016). By

scoring the probability of each base in particular positions of a sequence, PWMs can be

used to identify “matches” to motifs that could indicate the presence of a binding site.

While currently one of the best methods for TFBS prediction, PWMs presume sites to be

independent even though neighboring nucleotides often matter. Furthermore, many of

these methods are only able to identify clusters of TFBSs and not individual sites

(Jayaram et al. 2016). Owing to these problems, improved methodology is necessary to

move forward in TFBS prediction.

Due to the great difficult in predicting TFBSs, Levitsky et al. (2007) developed a

new GA-based method to predict binding sites. The authors noted that many PWMs

assume that single or select sites are essential to TFBS functionality and their reliance on

a few select bases results in a high false positive rate (Levitsky et al. 2007). In order to

mitigate these problems, the authors decided to approach the problem in a different way,

by identifying local dinucleotide pairings (LDPs) that could be potential sites before

analysis. To do this, the GA’s population consisted of a set of LDP frequencies, and the

21
fitness function used a Markov model to select for local dinucleotide pairings that were

statistically different from random sequence data (Levitsky et al. 2007). Once found,

LPDs are tested using discriminant analysis to predict if an unknown sequence is a TFBS.

A flexible window size also allows for flanking regions to be used; this can improve the

strength of predictions through the inclusion of regulatory elements highly associated

with TFBSs (Levitsky et al. 2007). This bottom-up approach to prediction was found to

outperform widely used PWM methods, and had a lower false positive rate (Levitsky et

al. 2007).

Since TFBSs are not the only class of short regulatory sequences, it is worth

noting that these solutions have the potential to be applied more widely. Motif discovery

is one of the largest classes of problems in bioinformatics, as the search for genomic

features is largely guided by the presence of sequence motifs indicating their presence

(Zara-Mirakabad et al. 2009).

Individual “monad” motifs are often the focus of research in this area, but more

complicated dyad motifs (where one motif closely follows another) are often overlooked.

This is likely a byproduct of the fact that many monad-focused approaches overlook

dyads due to low individual contribution to the signal, despite their combined importance

(Zara-Mirakabad et al. 2009). Dyads possess the advantage of more nucleotides,

statistically strengthening predictions; this also provides the opportunity for spacers

between the two motifs, as is often found in binding sites. For these reasons, Zara-

Mirakabad et al. implemented and tested a genetic algorithm approach to finding dyad

22
motifs. Using chromosomes that outlined the positions of possible dyads, researchers

evaluated fitness using a Position Frequency Matrix (similar to PWMs) to score the

signaling strength of potential motifs (Zara-Mirakabad et al. 2009). The resulting

algorithm was not only successfully able to predict dyads, but also outperformed existing

dyad-prediction software such as AlignAce (Roth, Hughes, Estep, and Church 1998) and

MITRA (Eskin and Pevzner 2002) (Zara-Mirakabad et al. 2009).

GA-based methods to TFBS prediction have outperformed the widely used

position weighted matrices, resulting in an extremely useful research tool. This improved

accuracy is particularly useful in identifying candidate TFBSs, which can be used to

direct experimental research. In addition, GA-based predictions do not require the use of

comparative methods, allowing them to be applied where data based on evolutionarily

conservation is unavailable. This is particularly helpful in discovering TFBSs, as there is

currently little experimentally verified binding data to make comparisons across multiple

species. The advantages of GAs circumvent many previous obstacles in discovering

TFBSs, providing new opportunities for regulatory research.

Determining Groups for Expression Analysis

In order to translate sequence data into meaningful phenotypic predictions, it is

important that specific motifs be linked to in vivo mRNA expression. The advent and

increasing accessibility of techniques such as RNA-seq as well as other methods such as

microarrays have made this possible, but is deciphering these data can be difficult (Ooi,

and Tan 2003). Different genes may contribute positively or negatively towards a

23
particular phenotype, and some may have a stronger influence than others. It is also

possible for a single gene to influence several different phenotypes. Analysis of mRNA

expression data therefore requires these different “classes” of genes to be organized by

sign and strength with respect to a particular phenotype.

This is a particularly difficult problem, as the number of possible combinations of

groupings grows exponentially as more genes are analyzed. Traditional mathematical

methods have struggled to address this problem; there is substantial drop-off in accuracy

for these methods when more than two or three groups are predicted for expression data

(Ooi et al. 2003). Since differences in expression can affect many different biological

pathways, the two or three classes provided by these methods are insufficient for

meaningful predictions to be made. To accurately represent biological data, methods for

expression analysis must be able to organize expressed genes into many different classes.

In order to do this, Ooi et al. (2003) developed a method combining GAs and maximum

likelihood to predict classes of genes from expression data, in order to determine tumor

phenotype. The authors used a chromosome representing different sets of genes, and

evaluated fitness with a maximum likelihood classifier. The classifier in turn used a

training set containing multiple tumor samples, allowing the gene classes to be scored

based on the probability that they would result in tumor growth (Ooi et al. 2003).

Implementing the program with a GA also resulted in several unique advantages: the

optimal group size for classifying the expression data was solved alongside the

membership of the group; and the use of a population allowed for a parallel search for

multiple possible gene classes at the same time (Ooi et al. 2003). The resulting

24
GA/MLHD program was more accurate than existing expression class prediction

methods, resulting in an improved means of analyzing expression data (Ooi et al. 2003).

This algorithm has been applied to microarray data from multiple cancer cell lines,

suggesting GAs can provide practical diagnostic applications in the field of medicine

(Ghaheri, Shoar, Naderan, and Hoseini 2015).

Transcriptomics

Once protein structures have been predicted and even verified, it is still crucially

important to consider the effects of regulation to make accurate in vivo predictions. Due

to the effects of regulatory non-coding RNAs such as miRNA, coding sequences alone

provide an incomplete view of gene expression (Reuter and Mathews 2010). Despite

their importance, approaches to determine RNA structures require tremendous effort. In

the case of ribosomal RNA secondary structure, predictions were slowly improved over

20 years of research before being validated by crystal structure (Dowell and Eddy 2006).

The use of machine learning approaches to predict transcripts and RNA structures can

greatly facilitate this process: isolation and verification are much easier once the basic

structure is known.

Prediction of RNA Secondary Structure

Previous attempts at RNA secondary structure predictions have mainly used

comparative analysis and thermodynamics based methods (Gardner and Giegerich 2004).

These approaches are similar to those used in protein structural prediction; comparative

25
analysis uses sequence homology to predict structure, while thermodynamics methods

seek to find the lowest free energy conformation of RNA secondary structure (Gardner

and Giegerich 2004). Many thermodynamics methods also struggle due to their reliance

on the assumption that the RNA alignment is correct; if this is not the case, their

predictions will be inaccurate (Dowell and Eddy 2006). This is likely the reason why the

accuracy of thermodynamics methods is lacking, with 73% accuracy at best for even

short sequences of RNA (Reuter and Mathews 2010). This number drops even further

when applied to longer sequences. Comparative analyses have also been applied, and are

known as one of the most accurate current methods (Dowell and Eddy 2006). These

analyses use a multiple alignment between homologous RNAs, using sequence

conservation to predict structural similarity. Despite this, they require an extremely large

number of sequences in order to find sufficient conservation to accurately identify folding

sites (Reuter and Mathews 2010). Sankoff algorithm approaches have also been applied

to the problem of RNA structural alignment (Havgaard and Gorodkin 2014). These

methods simultaneously compute RNA alignment and folding to predict a consensus

structure. However, this approach has mostly been limited to pairwise alignments due to

prohibitive computational costs (Dowell and Eddy 2006).

While alignments may aid in the prediction unknown RNA secondary structures,

this method is insufficient on its own. It does not consider that even small changes in

RNA sequences can affect structure and stability. It follows that it is possible for even

high-scoring alignments to have very different structures in vivo. At the same time,

alignments provide a means to predict RNA secondary structure without experimental

26
validation (Notredame, O’Brien, and Higgins 1997). As ab initio methods are currently

lacking due limited understanding of RNA folding in vivo, there is opportunity to

improve the method.

A hybrid method combining alignment and thermodynamics might improve RNA

structure predictions. Notredame et al. (1997) implemented this with PRAGA, a genetic

algorithm-based approach that combines structural alignments with dynamic

programming and free energy to predict RNA secondary structure. The algorithm begins

with some dynamic programming in order to create a main “backbone” for the alignment.

This creates an accurate main alignment to which finer tuning can be applied by the GA.

The use of the GA reduces the search space of the problem as not all alignments are

scored, minimizing expensive fine-scale calculations (Notredame et al. 1997). Combined

with the dynamic programming backbone, this allows the algorithm to still make accurate

predictions with relatively low runtime. In addition, the reduction in complexity allows

the algorithm to run on longer alignments than other methods (Notredame et al. 1997).

To compensate for the reliance on previously verified RNA structures in the alignment

portion of the algorithm, researchers also included thermodynamics to score the chemical

stability of the resulting structure. Again, in order to reduce the runtime cost of these

calculations, this was combined with alignment score in the fitness function for the GA.

Due to these improvements, PRAGA was not only able to successfully predict RNA

secondary structures in silico, but also succeeded in discovering pseudoknots, which

many previous methods have failed to find (Notredame et al. 1997). PRAGA established

GAs as a means of predicting RNA structure, and served as the backbone for the newer

27
Consensus Folding GA (Cofolga2) algorithm (Taneda 2008). Cofolga is able to run

faster than contemporary methods for RNA secondary structure prediction while

maintaining high sensitivity (Taneda 2008).

Prediction of RNA-RNA interactions

In order to accurately predict expression, it is necessary not only to understand

RNA structures but also to determine which structures interact. Many forms of RNA are

regulatory in nature, and modify expression levels through hybridization with other

transcripts. For example, miRNAs bind to specific mRNA transcripts and mark them for

degradation (Bhaskaran and Mohan 2014). These interactions are also of interest due to

the potential for novel disease treatments via gene therapy. Previous algorithms that have

attempted to solve this problem have struggled with incredibly high runtime requirements

due to the exponential complexity of RNA-RNA conformations (Montaseri, Zare-

Mirakabad, and Moghadam-Charkari 2014). In order to improve the efficiency of this

process, Montaseri et al. (2014) developed a GA to predict interactions between two

RNA molecules. By determining fitness using a free energy function, the author’s GA

was able to predict RNA-RNA interactions with accuracy comparable to currently

existing methods (Montaseri et al. 2014). In addition, the program was able to arrive at

these solutions with less runtime. Due to the runtime bottleneck in this field, this means

the application of GAs can expedite the slow process of identifying RNA-RNA

interactions.

28
Discovery of Gene Regulatory Networks

While expression data is extremely useful in predicting phenotype, it can also be

used to discover patterns of gene regulation. For this reason, Sîrbu et al. (2010) reviewed

several approaches to finding Gene Regulatory Networks (GRNs) from DNA microarray

data. They implemented and tested seven different evolutionary algorithms (a broader

class of algorithms related to GAs) for this task. Of these approaches, artificial neural

networks (ANN) combined with GAs and a genetic algorithm for local search (GLSDC)

performed the best (Sîrbu, Ruskin, and Crane 2010). The ANN-GA made the most

accurate predictions for small networks, and also was able to arrive at a solution with the

least amount of runtime. In a test case for noisy expression data, ANN-GA and GLSDC

continued to outperform other methods and was still able to provide accurate results

(Sîrbu et al. 2010). These methods not only provide a framework for the discovery of

gene regulatory networks, but also showcase the ability for hybrid GA methods to

outperform their constituent parts.

RNA Editing Site Prediction

A final regulatory process to consider is the possibility of post-transcriptional

modification of RNA. Of these possible alterations, RNA editing is currently not very

well characterized; in fact, many known instances of this process were discovered by

accident (Thompson J. and Gopal S. 2006). Consequently, there is opportunity for

computational methods to aid in the search for possible editing sites and direct

29
experimental validation efforts. In order to discover new RNA editing sites, Thompson

et al. (2006) implemented a GA that predicts C → U editing sites in plant mitochondria.

In order to make these predictions, candidate sites had fitness evaluated based on

comparison with known cytosine editing sites (Thompson and Gopal 2006). The

resulting RNA Editing site prediction by Genetic Algorithm Learning (REGAL)

algorithm was able to successfully predict C → U editing sites with 87% accuracy in

testing, outperforming existing methods (Thompson and Gopal 2006). While this is not

perfect, this tool could be applied to discover new RNA editing sites, which in turn could

further improve the algorithm’s predictions. Collaboration between computational and

experimental research in this field could prove vital in learning more about this biological

process.

Proteomics

Another major goal of genetics and bioinformatics research is to determine the

resulting protein from coding sequences in the genome. While it is simple to translate

coding DNA into amino acid sequences, the process of predicting the resulting protein

structure is an incredibly complex problem. This stems from the fact that the number of

possible rotation angles in an AA chain results in a combinatorial explosion of

conformations. Of all these possible options, only a few will be chemically stable, and of

these even less may be present in vivo.

30
Protein Structural Prediction

To eliminate complications that stem from variability in chemical stability, a free

energy function, such as Shannon entropy, is often used to evaluate how viable a protein

structure would be. However, to calculate this for all possible structures would be

infeasible due to the immense computational cost. In order to work around this

complexity, some methods have used simplifications of protein surfaces, side chains, and

even main chains to accelerate the process, but this comes at a great cost to accuracy

(Pedersen J. T. and Moult J. 1997). A separate strategy called correlated mutation

analysis (CMA) limits the number of comparisons by focusing on residues at conserved

positions (Bywater 2016). While this is a leading approach in protein structural

predictions, it has several drawbacks. CMA assumes coevolution between the compared

sequences based only upon homology which can be incorrect, as similar structures may

have evolved independently. An additional problem is that sequence similarity does not

always indicate structural similarity (Bywater 2016).

In order to maintain high accuracy and minimize assumptions, Pedersen and

Moult (1997) used genetic algorithms to reduce the search space of the problem. By

doing so, fewer protein structural evaluations would be required, mitigating

computational costs while maintaining accuracy. The authors used a fitness function

comprised of a free energy model, such that the GA’s population would automatically

evolve towards structures that were chemically viable (Pedersen and Moult 1997). This

eliminates a massive number of calculations on unstable protein structures that could not

31
exist long enough in vivo to have a meaningful biological effect. While the resulting

algorithm was unable to run faster than a state of the art Monte Carlo Chaining (MCC)

approach, it outperformed MCC in structural prediction as resulting structures had a

lower free energy (Pedersen and Moult 1997). Furthermore, the authors mention that

parallelization of the algorithm is a possibility, suggesting there is a potential for GAs to

outperform existing methods in runtime as well (Pedersen and Moult 1997).

A later approach developed by Rashid et al. (2013) sought to further improve

existing methods by mixing low- and high-resolution energy models in a GA. By

including the lower resolution Hydrophobic-Polar model, the algorithm was able to

explore conformations with hydrophobic cores. This is a problem previous approaches

have struggled with, as hydrophobic cores are often rejected early on due to lower initial

free energy before the rest of the structure is found and considered (Rashid, Newton,

Hoque, and Sattar 2013). Later on, a high resolution model is used to improve the

accuracy of larger structures. This allowed the approach to explore more possible

conformations without sacrificing accuracy. Although computationally intensive, this

approach was able to outperform existing methods by making lower free energy

structural predictions with lower root mean square deviation values (Rashid et al. 2013).

These GA-based in-silico structural predictions provide an opportunity to explore

possibilities for pharmaceutical compounds without costly laboratory work.

Finally, a GA-based approach to predicting protein structure implemented by

Huang et al. (2004) further improved upon previous approaches. The authors pointed out

32
that hydrophobic-hydrophilic models (HP model) of protein folding had implicit

assumptions caused by thermodynamics. Due to the fact that free energy is lower for

hydrophobic cores surrounded by hydrophilic amino acids, there is a tendency for these

HP models to gravitate toward these structures without first exploring other possibilities

(Huang, Yang Chang-Biau, Tseng, and Yang Chia-Ning 2004). To mitigate this bias, the

authors decided to use a 3D lattice model representation to predict protein structure, using

GA chromosomes to represent different possible conformations of the protein. To further

broaden the scope of possible structural predictions, the authors included many different

chemical properties based on sequence homology to inform their predictions (Huang et

al. 2004). In order to do this, a fitness function was implemented to consider many

different properties that evaluate candidate protein conformations. These properties

include secondary structures, such as alpha helixes and beta sheets, side chains, disulfide

bridges, and electrostatic interactions, and were all used to score the chemical viability of

possible structures (Huang et al. 2004). By using homology to discover and include these

properties, the resulting algorithm was able to make more accurate predictions than those

of other HP model based methods (Huang et al. 2004).

Drug Discovery

Protein structural predictions are important for discovering the biological effects

of genes, but are also critically important for drug development. Searching for new

compounds with desirable bioactivity is of great medical and economic importance.

However, screening for new bioactivity is an extraordinarily costly and time consuming

33
process, meaning experimental investigation of every compound is infeasible (Katsila,

Spyroulias, Patrinos, and Matsoukas 2016). As such, computer-based methods to

determine which compounds to screen experimentally are extremely important for the

field. Still, the exponentially large number of possible combinations and problem of

local optimums requires a systemic search in order to discover the best drug candidates

(Mandal et al. 2007). A recent approach developed by Mandal et al. (2007) uses GAs in

order to solve this problem. The authors chose to represent the problem with a

chromosome comprised of chains and substructures of a possible compound and

evaluated fitness using a simulation of activity for the desired target (Mandal et al. 2007).

Due to the generational improvement in GAs, this also greatly reduces the search space

by improving off existing chains and combinations in the population rather than creating

all compounds from scratch. The integration of computer-based predictions can greatly

facilitate drug design cut research costs (Katsila, et al. 2016). By reducing search space,

GAs provide a novel, efficient way to screen drug candidates.

Discussion

Advantages of Genetic Algorithms

Genetic algorithms are a powerful, flexible tool to address many challenges in

bioinformatics. Any problem where solutions can be ranked in a meaningful way to

measure fitness can be solved through the evolutionary aspect of GAs. This is

particularly helpful as many other machine learning methods are reliant on the use of

34
training data. These methods often fall victim to biases in the data, making selection of

the training sets a problem itself (Kubat and Matwin 1997). GAs only require a fitness

function and therefore avoid these problems; these functions can often be designed with

limited prior knowledge of a given bioinformatics problem. This makes GAs ideal for

problems where the search space is poorly understood, as they can tolerate some noise in

the fitness function (Manning et al. 2013). Due to these traits, GAs can be rapidly

applied to new areas of bioinformatics research where knowledge is imperfect.

This flexibility also leads GAs to be easily hybridized with other approaches. In

many cases, combining pre-existing methods with GAs can result in even better

algorithms, as demonstrated by the work of Hwang et al. (2013), Notredame et al. (1997),

Sîrbu et al. (2010), and Yang et al (2004). In addition to overall improvements through

combined methods, hybrid approaches may also be used to focus on particular aspects of

a solution that may be desirable. In this sense, specialized methods that may be weak on

their own can be improved with the strength of GAs while maintaining their mastery in

particular niches.

Another particularly useful aspect of GAs is their ability to reduce the

computational complexity of a given problem. Due to the fact that they explore only

some, but not all solutions for a given problem, the number of calculations, and therefore

runtime, required is greatly diminished. Complexity can be reduced even further through

the use of variable reduction strategies; these incorporate some domain knowledge to

greatly improve the performance of a GA (Wu et al. 2013). This is extremely helpful in

35
bioinformatics, as most problems in the field involve complex statistics to arrive at a

solution. This advantage can be even further improved through parallelization; the

population-based system of GAs can easily be distributed over multiple processors or

cores greatly reducing runtime requirements (Manning et al. 2013).

Perhaps the greatest and most straightforward advantage of genetic algorithms is

their ability to learn solutions with little knowledge of the problem beforehand. Even

simple definitions of fitness may result in extremely complex and effective solutions to a

problem (Angeline, P. J. 1994). Owing to GAs’ roots in evolution, emergent solutions

can arise, and just as in nature the sum of individually evolved parts may result in a

greater overall complexity than is apparent for each piece. As a result, the selective

process in GAs often results in robust solutions that would not have been foreseen by

human-directed research. These approaches also limit assumptions by researchers, as the

learning process can take place without providing the algorithm with explicit knowledge

(Angeline, P. J. 1994).

Weaknesses of Genetic Algorithms

While genetic algorithms are a very effective tool in bioinformatics, no method is

perfect and GAs have some drawbacks. While GAs often use less a priori knowledge

than other approaches, they do require researchers to design a fitness function (Spears

and De Jong 1990). Some problems may require complex fitness functions, meaning

substantial knowledge of the problem may still be required. The fitness function is also

instrumental in guiding the algorithm to a solution; biases can easily be introduced to the

36
function that alter the program’s output. If fitness is poorly defined or approximations

are introduced, the selection process will still follow those rules resulting in error (Sastry

and Goldberg 2002). This will result in a poor solution that may only work for a test

case, and fail in “real world” applications. Decisions on population size, number of

generations, and crossover and mutation rates may also influence learning and require

tuning for the best results (Manning et al. 2013). Unlike the fitness function, these

decisions must all be made arbitrarily, resulting in a painstakingly trial-and-error process

for tuning parameters. Optimizing GAs can be a problem in itself, and it can require the

algorithm to be run several times to find the best solution (Manning et al. 2013).

Another concern is that while GAs can be used to find strong solutions, they are

generally only “answers.” In other words, their output does not directly teach scientists

the rules for the problem. As discovering underlying genetic methods and discovering

how genomes are organized is a principal topic of research, this is one of the greatest

shortcomings of GAs and machine learning in the field. GAs can be used to guide this

type of research by providing examples, but only provide answers and never an

explanation as to why the algorithm arrived at it. In this sense, GAs are a “black box”

with an input and output (Manning et al. 2013).

In the scope of bioinformatics methodology, GAs are relatively fast for making

accurate predictions. However, the fact that a large population size and many generations

are required by GAs means they still take a substantial time to run; they are generally

slower than directed search methods (Manning et al. 2013). These runtime constraints

37
mean GAs are not the best option for simple problems that do not require substantial

machine learning to solve. In cases where the rules of the problem are well established, a

search algorithm incorporating pre-existing knowledge may be more efficient (Manning

et al. 2013). An “educated guess” solution may also suffice in some cases, meaning the

improved accuracy of a GA may not be worth the runtime cost. In these situations GAs

would take an extremely long time to run compared to other algorithms, making

traditional methods more suitable. The runtime costs of GAs can be mitigated by

parallelization, but they can still be considerably slow (Manning et al. 2013).

Conclusion

Genetic algorithms provide an incredibly diverse and effective set of tools for

bioinformatics analysis. Their ability to solve major problems in genomics,

transcriptomics, and proteomics may greatly expedite bioinformatics research. This is

demonstrated by the consistently higher performance of GA methods over existing

methods in these categories. GA methods are also highly adaptable, and can often

overcome weaknesses through the use of hybrid approaches. Finally, additional sequence

data will continue to improve GAs as more learning examples become available. The

accuracy, efficiency, and potential for growth in GA-based methodology provides a

robust solution to data analysis in bioinformatics.

38
Bibliography

Angeline, P. J. 1994. Genetic Programming and Emergent Intelligence. Advances in

Genetic Programming. Ed. K. E. Kinnear. Cambridge, MA. MIT Press.

Azad A. K. M., Shahid S., Noman N., and Lee H. 2011. Prediction of plant promoters

based on hexamers and random triplet pair analysis. Algorithms for Molecular

Biology 6:19.

Beasley D., Bull, D. R., and Martin R. R. 1993. An Overview of Genetic Algorithms:

Part 1, Fundamentals. University Computing, 15(2) pp. 58-69.

Bhaskaran, M., & Mohan, M. 2014. MicroRNAs: History, Biogenesis, and Their

Evolving Role in Animal Development and Disease. Veterinary Pathology, 51(4),

759–774. http://doi.org/10.1177/0300985813502820

Bhandari D., Murthy C. A., and Pal S. K. 2012. Variance as a stopping criterion for

genetic algorithms with elitist model. Fundamata Informaticae 120:145-164. doi:

10.3233/FI-2012-754

Blickle T. and Thiele L. 1996. A Comparison of Selection Schemes Used in Evolutionary

Algorithms. Evolutionary Computation, 4-4:361-394. Dec. 1996. doi:

10.1162/evco.1996.4.4.361

39
Bojarski M., Yeres P., Choromanska A., Choromanski K., Firner B., Jackel L., and

Muller U. 2017. Explaining how a Deep Neural Network trained with End-to-End

Learning steers a car. arXiv:1704.07911 [cs.CV]

Brain, Damien & Webb, Geoffrey. 2000. On the Effect of Data Set Size on Bias and

Variance in Classification Learning. Proceedings of the Fourth Australian

Knowledge Acquisition Workshop.

Burge, C. and Karlin, S. 1997. Prediction of complete gene structures in human genomic

DNA. Journal of Molecular Biology 268:78-94.

Bywater, Robert P. 2016. Comparison of algorithms for prediction of protein structural

features from evolutionary data. PLoS ONE 11(3):e0150769.

doi:10.1371/journal.pone.0150769

Chowdhury B., Garai A., and Garai G. 2016. An optimized approach for annotation of

large eukaryotic genomic sequences using genetic algorithm. bioRxiv preprint,

posted online Oct. 25, 2016.

Corpet F. 1988. Multiple sequence alignment with hierarchical clustering.

Nucl. Acids Res., 16 (22), 10881-10890.

Delcher A., Bratke K., Powers E., and Salzberg S. 2007. Identifying bacterial genes and

endosymbiont DNA with Glimmer. Bioinformatics 23:673–679.

doi:10.1093/bioinformatics/btm009

40
Devi Arockia Vanitha C., Devaraj D., Venkatesulu M. 2015. Gene Expression Data

Classification using Support Vector Machine and Mutual Information-based Gene

Selection. Procedia Computer Science 47:13–21. doi:

10.1016/j.procs.2015.03.178

Dharwal R. and Kaur L. 2016. Applications of Artificial Neural Networks: A Review.

Indian Journal of Science and Technology, 9(47).

doi:10.17485/ijst/2016/v9i47/106807

Diaz-Gomez P. A. and Hougen D. F. 2007. Initial population for genetic algorithms: A

metric approach. International Conference on Genetic and Evolutionary

Methods, 2007. Las Vegas, Nevada, USA.

Dowell R. and Eddy S. 2006. Efficient pairwise RNA structure prediction and alignment

using sequence alignment constraints. BMC Bioinformatics 7:400.

doi:10.1186/1471-2105-7-400.

Eskin E., and Pevzner P. 2002. Finding composite regulatory patterns in DNA sequences.

Bioinformatics 18:354-363.

Frankel WN and Schork NJ. 1996. Who’s afraid of epistasis? Nature Genetics 14:371-

373. [Pubmed 8944011]

41
Galperin M. and Koonin E. 2010. From complete genome sequence to “complete”

understanding? Trends in Biotechnology 28(8):398–406.

doi:10.1016/j.tibtech.2010.05.006.

Gardner P. P. and Giegerich R. 2004. A comprehensive comparison of comparative RNA

structure prediction approaches. BMC Bioinformatics 5:140.

https://doi.org/10.1186/1471-2105-5-140

Ghaheri A., Shoar S., Naderan M., and Hoseini S. S. 2015. The Applications of Genetic

Algorithms in Medicine. Oman Medical Journal 30(6), pp. 406–416.

http://doi.org/10.5001/omj.2015.82

Gómez-Ramos E. and Venegas-Martínez F. 2013. A Review of Artificial Neural

Networks: How well do they perform in Forecasting Time Series? Journal of

Statistical Analyis 6(2): 7-15.

Guzman C. and D'Orso I. 2017. CIPHER: A flexible and extensive workflow platform

for integrative next-generation sequencing data analysis and genomic regulatory

element prediction. BMC Bioinformatics 18. doi:10.1186/s12859-017-1770-1.

Haupt, R. L. 2000. Optimum Population Size and Mutation Rate for a Simple Real

Genetic Algorithm that Optimizes Array Factors. Applied Computational

Electromagnetics Society Journal 15: 1034-1037. 10.1109/APS.2000.875398.

42
Haupt R. L. and Haupt S. E. 2004. Practical Genetic Algorithms. Second Edition.

Hoboken, NJ: John Wiley & Sons, Inc. Chapter 2 pp. 27-50.

Holland, John. 1992. Genetic Algorithms. Scientific American July 1992:66-72

Horner D. S., Pavesi G., Castrignanò T., De Meo P. D., Liuni S., Sammeth M., Picardi E.,

and Pesole G. 2010. Bioinformatics approaches for genomics and post genomics

applications of next-generation sequencing. Briefings in Bioinformatics 11:2, pp.

181–197. https://doi.org/10.1093/bib/bbp046

Hwang K., Ha B., Ju S., and Kim S. 2013. Partial AUC maximization for essential gene

prediction using genetic algorithms. BMB Reports 46:41-46. doi:

10.5483/BMBRep.2013.46.1.159

Huang Y., Yang Chang-Biau, Tseng K., and Yang Chia-Ning. 2004. Protein Folding

Prediction with Genetic Algorithms (Master’s Thesis). Retrieved from Research

Gate: https://www.researchgate.net/profile/Chia-

Ning_Yang/publication/240836700_Protein_Folding_Prediction_with_Genetic_A

lgorithms/links/004635296a5ec9a5b0000000/Protein-Folding-Prediction-with-

Genetic-Algorithms.pdf

Jayaram N., Usvyat D., and Martin A. C.R. 2016. Evaluating tools for transcription factor

binding site prediction. BMC Bioinformatics BMC Series 1298.

doi:10.1186/s12859-016-1298-9.

43
Jürgen Schmidhuber. 2015. Deep learning in neural networks: An overview. Neural

Networks 61:85-117. http://dx.doi.org/10.1016/j.neunet.2014.09.003

Katsila T., Spyroulias G. A., Patrinos G. P., and Matsoukas M. 2016. Computational

approaches in target identification and drug discovery. Computational and

Structural Biotechnology Journal 14, pp. 177-184.

https://doi.org/10.1016/j.csbj.2016.04.004

Kim I. Y., and Weck O. L. 2005. Variable chromosome length genetic algorithm for

progressive refinement in topology optimization. Structural Multidisciplinary

Optimization 29:445-456. doi: 10.1007/s00158-004-0498-5

Krause, L., McHardy, A. C., Nattkemper, T. W., Pühler, A., Stoye, J., and Meyer, F.

2007. GISMO—gene identification using a support vector machine for ORF

classification. Nucleic Acids Research 35(2), pp. 540–549.

http://doi.org/10.1093/nar/gkl1083

Krogh, A. 1997. Two methods for improving performance of an HMM and their

application for gene finding. Proceedings of the 5th International Conference of

Intelligent Systems for Molecular Biology. pp. 179-186. AAAI Press, Menlo

Park, CA.

Kubat M. and Matwin S. 1997. Addressing the Curse of Imbalanced Training Sets: One-

Sided Selection. Proceedings of the Fourteenth International Conference on

Machine Learning. pp. 179-186, San Francisco, CA, July 8-12 1997.

44
Leung Y. and Wang Y. 2001. An orthogonal genetic algorithm with quantization for

global numerical optimization. IEEE Transactions on Evolutionary

Computation, 5-1:41-53. Feb 2001. doi: 10.1109/4235.910464

Levitsky V. G., Ignatieva E. V., Ananko E. A., Turnaev I. I., Merkulova T. I., Kolchanov

N. A., and Hodgman T.C. 2007. Effective transcription factor binding site

prediction using a combination of optimization, a genetic algorithm and

discriminant analysis to capture distant interactions. BMC Bioinformatics 8:481.

doi:10.1186/1471-2105-8-481.

Mandal A., Johnson K., Wu, C. F. J., and Bornemeier D. 2007. Identifying promising

compounds in drug discovery: genetic algorithms and some new statistical

techniques. Journal of Chemical Information and Modelling 47:981-988.

Manning T., Sleator R. D., and Walsh P. 2013. Naturally selecting solutions: The use of

genetic algorithms in bioinformatics. Bioengineered 4:266-278.

http://dx.doi.org/10.4161/bioe.23041

Manzoni C., Kia D. A., Vandrovcova J., Hardy J., Wood N., Lewis P., and Ferrari R.

2018. Genome, transcriptome and proteome: the rise of omics data and their

integration in biomedical sciences, Briefings in Bioinformatics, 19-2:286-302.

https://doi.org/10.1093/bib/bbw114

Mathé C., Sagot M., Schiex T., and Rouzé P. 2002. Current methods of gene prediction,

their strengths and weaknesses. Nucleic Acids Research 19:4103-4117.

45
Montaseri S., Zare-Mirakabad F., and Moghadam-Charkari N. 2014. RNA-RNA

interaction prediction using genetic algorithm. Algorithms for Molecular Biology

9:17. http://www.almob.org/content/9/1/17

Moore J. H., Hahn L. W., Ritchie M. D., Thornton, T. A. and White, B. C. 2004. Routine

discovery of complex genetic models using genetic algorithms. Applied Soft

Computing 4:79-86. doi:10.1016/j.asoc.2003.08.003.

Nizam A., Ravi J., and Subbaraya K. 2011. Cyclic genetic algorithm for multiple

sequence alignment. International Journal of Research and Reviews in Electrical

and Computer Engineering ISSN: 2046-5149.

Notredame C., O’Brien E. A., and Higgins D. G. 1997. RAGA: RNA sequence alignment

by genetic algorithm. Nucleic Acids Research 25:4570-4580.

Oliveto P. S., and Witt C. 2014. On the runtime analysis of the Simple Genetic

Algorithm. Theoretical Computer Science 545:2-19. ISSN 0304-3975.

https://doi.org/10.1016/j.tcs.2013.06.015.

Ooi C.H., and Tan, P. 2003. Genetic algorithms applied to multi-class prediction for the

analysis of gene expression data. Bioinformatics 19:37-44.

Pedersen J. T. and Moult J. 1997. Protein folding simulations with genetic algorithms and

a detailed molecular description. Journal of Molecular Biology 269:240-259.

46
Rashid M. A., Newton M. A. H., Hoque Md. T., and Sattar A. 2013. Mixing energy

models in genetic algorithms for on-lattice protein structure prediction. BioMed

Research International Volume 2013. http://dx.doi.org/10.1155/2013/924137.

Reuter J and Mathews H. 2010. RNAstructure: software for RNA secondary structure

prediction and analysis. BMC Bioinformatics 11:129.

http://www.biomedcentral.com/1471-2105/11/129

Rogers A. and Prugel-Bennett A. 1999. Genetic drift in genetic algorithm selection

schemes. IEEE Transactions on Evolutionary Computation, 3-4:298-303, Nov.

1999. doi: 10.1109/4235.797972

Roth F. P., Hughes J. D., Estep P. W., and Church G. M. 1998. Finding DNA regulatory

motifs within unaligned noncoding sequences clustered by whole-genome mRNA

quantitation. Nat Biotechnol. 16(10):939-45.

Ruffalo M., LaFramboise T., and Koyutürk M. 2011. Comparative analysis of algorithms

for next-generation sequencing read alignment. Bioinformatics 27-20:2790-2796.

https://doi.org/10.1093/bioinformatics/btr477

Havgaard J. H. and Gorodkin J. 2014. RNA structural alignments, part I: Sankoff-based

47
Sastry, K., and Goldberg, D.E. (2002). Genetic algorithms, efficiency enhancement, and

deciding well with differing fitness variances. Proceedings of the Genetic and

Evolutionary Computation Conference, 528–535. (Also IlliGAL Report No.

2002002).

Sîrbu A., Ruskin, H. J., and Crane M. 2010. Comparison of evolutionary algorithms in

gene regulatory network model inference. BMC Bioinformatics 11:59.

http://www.biomedcentral.com/1471-2105/11/59

Spears, William M. 1995. Adapting crossover in evolutionary algorithms. Retrieved

from: https://www.researchgate.net/publication/2263702_Adapting_Crossover

_in_Evolutionary_Algorithms

Spears, W. M. and De Jong K. A. 1990. Using genetic algorithms for supervised concept

learning. Proceedings of the 2nd International IEEE Conference on Tools for

Artificial Intelligence. pp. 335-341. Herndon, VA, USA. doi:

10.1109/TAI.1990.130359

Stanke M. and Morgenstern B. 2005. AUGUSTUS: a web server for gene prediction in

eukaryotes that allows user-defined constraints. Nucleic Acids Research, 33(Web

Server issue), 465–467. http://doi.org/10.1093/nar/gki458

Tahmasebipour K. and Houghten S. 2015. Disease-gene association using a genetic

algorithm. Conference paper presented at 2015 IEEE Conference on

48
 Computational Intelligence in Bioinformatics and Computational Biology,

held at Niagara Falls, Canada in August 2015. doi:10.1109/CIBCB.2015.7300331

Taneda A. 2008. An efficient genetic algorithm for structural RNA pairwise alignment

and its application to non-coding RNA discovery in yeast. BMC Bioinformatics

9:521. doi: 10.1186/1471-2105-9-521

Thompson J. and Gopal S. 2006. Genetic algorithm learning as a robust approach to RNA

editing site prediction. BMC Bioinformatics 7:145. doi:10.1186/1471-2105-7-145

Thompson J. D., Higgins D. G., and Gibson T. J. 1994. CLUSTAL W: improving the

sensitivity of progressive multiple sequence alignment through sequence

weighting, position-specific gap penalties and weight matrix choice. Nucleic

Acids Research, 22(22), 4673–4680.

Umarov R. K. and Solovyev V. V. 2017. Recognition of prokaryotic and eukaryotic

promoters using convolutional deep learning neural networks. PLoS ONE, 12(2):

e0171410. http://doi.org/10.1371/journal.pone.0171410

Xie X., Wu S., Lam K., and Yan H. 2006. PromoterExplorer: an effective promoter

identification method based on the AdaBoost algorithm. Bioinformatics 22:2722-

2728. doi:10.1093/bioinformatics/btl482

49
Wall D. P., Pivovarov R., Tong M., Jung J.-Y., Fusaro V. A., DeLuca T. F., and

Tonellato P. J. 2010. Genotator: A disease-agnostic tool for genetic annotation of

disease. BMC Medical Genomics, 3, 50. http://doi.org/10.1186/1755-8794-3-50

Winters-Hilt, S. and Merat, S. 2007. SVM clustering. Proceedings of Fourth Annual

MCBIOS Conference. BMC Bioinformatics 8(Suppl. 7):S18. doi:10.1186/1471-

2105-8-S7-S18

Wu G., Pedrycz W., Li H., Qiu D., Ma M., and Liu J. 2013. Complexity Reduction in the

Use of Evolutionary Algorithms to Function Optimization: A Variable Reduction

Strategy. The Scientific World vol. 2013, Article ID 172193.

https://doi.org/10.1155/2013/172193.

Yang C., Moi S., Lin Y., and Chuang L. 2016. Genetic algorithm combined with a local

search method for identifying susceptibility genes. JAISCR 6:203-212.

doi:10.1515/jaiscr-2016-0015

Zara-Mirakabad F., Ahrabian H., Sadeghi M., Hashemifar S., Nowzari-Dalini A., and

Goliaei B. 2009. Genetic algorithm for dyad pattern finding in DNA sequences.

Genes and Genetic Systems 84: 81-93.

50