You are on page 1of 2

f fG (1)

𝐅𝐂𝐆𝐑𝟏 (𝐆) = ( C ),
fA fT

where 𝐅𝐂𝐆𝐑𝟏 (𝐆) represents the first-order FCGR matrix and f is the frequency of each 1-mer (A, T, C, and G).

𝐟𝐤 − minf (5)
𝐧𝐟𝐤 = × 255,
maxf − minf

where nfk and fk represent the normalized and original frequencies of each k-mer, respectively, and maxf and minf

are the maximum and minimum frequencies, respectively.

H (p, q) = ∑ 𝑝(x) log q(x) (6)


𝑥𝑐

where p(x) is the actual probability distribution, c is the classes number and q(x) is the predicted probability

distribution

𝜃 = 𝜃 −  𝛻𝜃 𝐽(𝜃; 𝑥 (𝑖) ; 𝑦 (𝑖) ) (7)

where θ is the model weights,  is the learning rate, and ∇θ J(θ; x (i) ; y (i) ) is the gradient of the cost function.

|value(f, S1) − value(f, S2)| 8


diff (f, S1, S2) = ,
max(f) − min(f)

where f represents a feature, S1 represents a RO, and S2 represents an H or M.

k k 9
diff (f, ROI , HJ ) diff (f, ROI , MJ )
𝐰[𝐟] ∶= 𝐰[𝐟] − ∑ +∑ ,
n×k n×k
J=1 J=1

where H and M represent the observations from the same and opposite classes, respectively.

NF

Ypredi = β0 + β1 fi1 + β2 fi2 + ⋯ + βNF fiNF = β0 + ∑ βj fij , 10


j=1

where Ypredi is the label at observation i, β0 is the intercept of Ypred, fij is the jth feature at observation i, βj is
the coefficient of the jth feature, and NF is the number of features.

N
11
MSE(β) = ∑(Yactuali − Ypredi )2 ,
i=1

where N is the number of training observations, Yactuali and Ypredi are the actual and predicted responses at

observation i.
N NF

MSE(β) = ∑(Yactuali − Ypredi )2 + λ ∑|βj |, 12


i=1 j=1

where λ is a nonnegative regularization parameter, βj is a coefficient vector of length NF.

Tp + Tn
Accuracy = × 100%,
Tp + Fp + Tn + Fn

Tp
Precision = × 100%,
Fp + Tp

Tn
Specificity = × 100%,
Tn + Fp
Tp
Sensitivity = × 100%,
Tp + Fn
2Tp
F1 − score = ,
2Tp + Fp + Fn
(Tp × Tn) − (Fn × Fp)
MCC = ,
√(Tp + Fp)(Fp + Tn)(Tn + Fn)(Fn + Tp)
where Fn and Tn denote false-negative and true-negative values, and Fp and Tp denote the false-positive and true-

positive values, respectively.

You might also like