-The clinical classification of a person's genetic vari- that affect the interaction of genes. Also, based on models of ant can lead to conflicting classifications. The presence of conflicts ensemble methods, the most important variables affecting the is determined manually by laboratory methods. If there is a genome were identified. But the authors led to the fact that tchoinsflwicot,rkt,hewniththethree hiselap doifffmicauclthyinien leinatrenripnrgetainlggortihthemrse,suitlt.wIans there were some limitations that allow us to consider the use of possible to train the neural network to predict conflicts with an machine learning methods as an addition to existing laboratory accuracy of 77%, and also to determine which parameters are tests to identify the basis of complex genetic diseases [6]. most important in classification. A study was conducted in [7], during which it was proved
The constant discoveries in biology provide us with a huge
amount of data for analysis, studying this data we can find
more and more patterns and relationships between the various
characteristics of living organisms, which leads us to a greater
understanding of how the world works and how we can
improve it. One of the studies that triggered a push in the
development of genetics was done by the biologist Gregor
Mendel. His works contain information on the established
relationship between the presence of certain genes and various
morphological and physiological characteristics of individuals,
as well as on such a key property of organisms that the genetic
code is able to be transmitted hereditarily, with preservation
of signs from parents to descendants [
Starting with the study of simple organisms, research has
reached the human genome. One of the directions in this
environment is a genome-wide search for associations related
to the study of associations between genomic variants and
phenotypic characters. The main goal is to predict a predisposition
to a disease by identifying genetic risk options, which is based
on a comparison of the alleles of healthy people and people
with diseases [
The studies mentioned above are carried out empirically
by specialists in this field. T he d evelopment o f information
technologies allows us to simplify and automate the same type
of work, so using machine learning methods, the researchers
in [
A. Feature engineering of the current issue
Machine learning methods can work only with data
presented in numerical form. Also, based on work [
Non-numerical features were processed. The number of unique values in them influenced the choice of method for processing.
The PolyPhen (Polymorphism Phenotyping) characterization consists of several descriptions of the possible effects of amino acid substitution on the structure and function of the human protein. For this data set, the Label Encoder method was applied, which assigns a certain number to each state. Such a characteristic as EXON (Exon), a part of the gene encoding amino acids, was presented in the source data as a part of exons of their total number in the body, which required additional processing using methods for working with strings from the Python library.
For characteristics such as cDNA position (position of the gene pair in the sequence of additional DNA), CDS position (position of the base pair of genes in the coding region), Protein position (position of the amino acid in the protein), the data presented as ranges of positions; therefore, their median values expressed scalar.
Characteristics: REF (comparison allele), ALT (alternative allele), CHROM (chromosome variant), Allele (allele), Consequence (consequence type) contain a number of different values from 24 to 866. Because of this, we can consider the applicable encoding method for these values The Feature hasher method, which vectorizes features into a certain number of columns that can represented as elements of Boolean algebra.
Such characteristics as: CLNVC (variant type), IMPACT (impact modifier for the kind of consequence), BIOTYPE (biotype) have a relatively small set of different values. Therefore, the One Hot Encode method is well applicable for them, creating a vector for each characteristic value.
Studies also conducted on the remaining columns; in some of them, there were no values, which would only complicate the work of the Random forest method.
In some, like CLNHGVS (a characteristic containing a description of the level of genome location), all values were unique, it follows that the correlation between them is zero. The establishment of relationships that play a role for the classifier is impossible. Therefore, part of the information has been removed.
This forecasting problem can be solved using binary
classification methods [
A. Description of the mathematical model of the selected methods
The decision tree described as follows: let the training
vectors xi 2 Rn; i = 1; : : : ; l , and the label vector y 2 Rl
given: the decision tree divides the space recursively so that
the samples with the same labels grouped. Let the data at
node m represented by Q. For each candidate, the
separation = (j; tm), consisting of the characteristic j and
the threshold tm, divides the data into subsets Qleft( ) and
Qright( ), where Qleft( ) = (x; y)jxj tm; Qright( ) = Qn
Qleft( ). Then, the function G(Q; ) = nleft H(Qleft( )) +
Nm
nright H(Qright( )) is calculated, where H( ) is the measure
Nm
of entropy given by the formula: H(Xm) = Pk pmklog(pmk).
Next, the parameters selected according to the following
criterion: = argmin G(Q; ). The subsets Qleft = ( ) and
Qright = ( ) are determined until the maximum available
depth is reached:Nm < minsamples or Nm = 1 [
The interaction of decision trees in the Random Forest
algorithm carried out using the bagging approach - the creation of
independent models for assessment, and then the averaging of
their forecasts using the following formula: Sl = 1l PL
l=1 wl;
where L is the number of independent base models, and wl is
the received dataset by each model [
In the Gradient boosting method, interactions between
decision trees carried out according to the principle of boosting,
based on the fact that the family of models is combined to
create the strongest of the basic ones. Several weak models
are adaptively selected, and based on their results. A stronger
value is attached to those objects in the dataset that were
poorly processed by previous models, thus reducing the bias
of the estimate even with a decrease in the spread [
Gradient boosting considers additive models of the
following form:F (x) = PmM=1 mhm(x); where hm() are the
basic models, and m is the step size calculated by the
one-dimensional optimization process. This method, like the
Random Forest, uses the decision tree as a simple model.
Thanks to such features of the ”tree” as processing mixed-type
data and modeling complex functions, it is ideal for optimizing
the step. GB constructs the additive model in a ”greedy” way:
Fm(x) = Fm 1(x)+ mhm(x), where the recently added tree
hm tries to minimize the loss of L, given the previous
ensemble Fm 1: hm = argmin Pn
i=1 L(yi; Fm 1(xi) + h(xi))
h
[
used, such as: Precision, Recall, F-score. These metrics are
calculated based on the following criteria: True Positives (TP)
- the correctly predicted positive value of the source class;
It is worth noting that for binary classification, the positive
result is one, and the negative is zero; True Negatives (TN)
- the correctly predicted negative value of the source class,
False Positives (FP) - the case when the result in the original
class is represented by negation, while the classifier returned a
positive result; False Negatives (FN) - the case is the opposite
of FP. Precision is the ratio of correctly predicted values to the
total number of attempts to give a positive result, presented by
the formula: T PT+PF P . Recall is the ratio of correctly predicted
values to their total number: T PT+PF N . F-score is the most
accurate measure, which contains the two scores listed above,
given by the formula: 2 RReeccaallll +xPPrreecciissiioonn [
Consider how the results of the models depend on the number of trees. For Random Forest, the result of the dependence of F-score, Precision, Recall on the number of trees in the ensemble is clearly shown in Figure 1. As you can see, the Precision value actively increases with the number of trees. Still, after 64, the result remains practically unchanged in the Recall, and The F-score also has such a tendency with a large number of trees. However, the graph also shows that they take the most excellent value with five trees, in which case the Precision, Recall, F-score measures take the following values, respectively: 0.46, 0.37, 0.41. But, unfortunately, a model with so many trees is not suitable for correction using the predict proba method, since the essence of this method is that you can control the adoption of the tree’s voice by increasing the weights, then with a small number of trees the result of such a change is rather weak. Therefore, it will be entirely objective to take a model with 128 trees.
The result of such a model recorded in table 1; the table also contains the result after applying predict proba.
The same study done for Gradient boosting, the metrics for a different number of trees shown in Figure 2. Because each of the following trees of this method is trained based on previous results, it found experimentally that less than 14 trees are not enough to predict results. But after 14, an apparent increase in the classification accuracy is visible, so it would be reasonable to take the one that has the largest number of trees as the main model and modify it by changing the weights. Still, practical measurements have shown that for this task, the classifier models based on Gradient boosting with the number trees 128, 256, 512 have the same result after applying the predict proba method. In this case, for a more objective comparison with the Random Forest method, we take a model with 128 trees.
The result of the best classifier m odel b ased o n Gradient boosting presented in Table 2. Comparing the results from Tables 1 and 2, it becomes clear that the results of the metrics before modification are identical for the two classifiers; after applying the predict proba method, the Precision and Recall metrics are different, but for a positive case, the F-score metric is the same. For an additional comparison of methods, to find the optimal one, Table 3 shows the time during which the classifiers a re t rained a nd a lso p redict t he v alues o n t he test
The training time for the two methods turned out to be comparable, while the classification time for the model based on Gradient Boosting is ahead of the model based on Random Forest by order of magnitude. Thus, the Gradient Boosting classifier is most preferred for determining whether a patient has a genetic conflict.
Also, Figure 3 shows a graph of the ROC-curve for two
models, which allows one to evaluate the quality of the binary
classification [
Based on Figure 2, the reading area is estimated at 0.77, the quality of the classifier i s d etermined b y h ow m uch this indicator is high.
The work was partially funded by the Russian Foundation for Basic Research under grants No. 19-29-01235 and 19-29and Higher Education within the government project of the FSRC Crystallography and Photonics RAS under grant No. additional medical research, and will also predict the need for screening for patients with suspected illness.