The model is based on a Multiple Logistic Regression (MLR) taking four variables from the 5’UTRs 
to predict a score: Length, GC percentage, upstream open reading frames (uORFs) and similarity to 
Atf4 positive control (Atf4like). Moreover, despite the score, we only consider positive targets 
containing at least one uORF, so the previously described biological mechanism for p-eIF2α-dependent 
translation can be possible. Then, before performing MLR analysis, we filtered our training dataset 
for transcripts having at least one uORF. 

For our MLR model, we obtain a value for each beta coefficient of the following formula: 
Score= exp(Prescore) / 1+exp(Prescore) 
Prescore=β0+β1*A+β2*B+β3*C+β4*D+β5*A*B+β6*A*C+β7*A*D+β8*B*C+β9*B*D+β10*C*D
Where: A = uORFs, B = Atf4like, C = Length, D = %GC. 

The first step to train our model was to obtain the A, B, C and D values for each 5’UTR in order to 
be able to obtain beta values. 
  -A (uORFs) is a discrete variable that counts all possible uORFs. 
  To achieve this, we consider each of the three possible reading 
  frames and count the AUGs with an “adequate” Kozak consensus sequence 
  within them (18). We exclude the ones either too close to the previous 
  one (less than 30 nucleotides of distance, considering a minimum space 
  for the ribosome to leaky-scan the first uAUG and try to translate a second 
  uAUG (19)) or within an already opened reading frame by a previous uORF that 
  did not found a STOP codon. 

  -B (Atf4like) is a categorical variable (0, 1, 2 or 3) that counts the reading 
  frames in which we find an uORF which has still not found a STOP codon and was 
  preceded by at least one already closed uORF (not being necessary that the already 
  closed uORF is in the same reading frame). 

  -C (Length) is a continuous variable that counts the number of nucleotides present in the 5’UTR.

  -D (%GC) is a continuous variable that takes the percentage of G or C nucleotides from the total length. 

Once having the input variables for each 5’UTR, we split the filtered training dataset in four subsets, 
containing the closest to the same ratio of positive to negative targets. Then, we clustered the subsets 
in the four possible groups of three, in order to use these clusters to train the model. For each cluster, 
a MLR model was trained using sklearn.linear_model.LogisticRegression module from Python3. We tested each 
MLR model using the remaining subset for each case as training.