The authors use
Kapetanovic IM (2008) Computer-aided drug discovery and development (CADDD): in silico-chemico-biological approach. Chem Biol Interact 171: 165–176.
This generates the descriptors, which are normally used in small molecule interactions and they obtained a moderate (in reality low for ML) correlation. That is the innovative part of the study. "Decision trees" used alone are not recommended because it will over tighten onto the data. The reason they used them was to maximise the fit - the risk is over tightening. When its says "the algorithm successfully ...", well thats using a Decision Tree, any other ML algorithm will be lower.
The very nice part of the paper is the combination of wet-lab and analytical work, that balance is often not logistically trivial to achieve.
What is not clear is they talk about 6 descriptors out of 60 and also 27 descriptors out of 60. Which is the one they use? I would suggest the paper was not rigorously reviewed/edited to permit this sort of confusion for the reader. It could be (suspect) they performed two different analyses, course grain (6) and fine grain (27), but I don't think that is described very well.
I will not critique further.
Summary The paper is commendable for using a small molecular physiochemical/ thermodynamic description algorithm onto RNA data. Thats cool. Wrapping it around a Decision Tree alone, I wouldn't agree with - nothing wrong with DT's but they can't be used in isolation.
QSAR Model
In order to find the mathematical relationship between aptamers’ structure and activity in the training set, two functions of MASS packages of R program were used (R 2.14.1 2011 The R Foundation). Firstly, a stepwise-selection with the function “stepAIC” (direction = both) focused on the most relevant 27 descriptors out of the 60 that can contribute to the prediction of the relative activity. Then, a second step with multiple-linear-regression function “lm” was applied on the selected descriptors to generate appropriate coefficients for the linear correlation equation. The quality of the output was evaluated according to the p-value of the null hypothesis, and multiple R2 indicated the correlation between actual and predicted values.
QSAR
Sixty descriptors quantifying structural and thermodynamic properties of the aptamers were produced and gradually the most relevant descriptors were selected to form a six descriptor- based algorithm, which can predict the binding of aptamers to influenza virus and thus anticipate its anti-viral activity. The algorithm successfully predicted the binding affinity in the training set of aptamers (R2 = 0.702) and the was validated with a test set (R2 = 0.66); moreover, the model’s sensitivity and specificity in selecting aptamers with enhanced relative binding were high, 89% and 87%, respectively.
Personally, I don't think that is the question the OP is asking - I think they are asking about a different technique in RNA secondary structure analysis to the technique authors performed here. That however is simply a guess. It's an interesting paper, there is no question about that - it's not how I personally would do it - but it's interesting. The issue is the thermodynamics has been rigorously worked out historically for RNA folding - it's a large body of work - and I'm sure the one or two reviewers would have pointed this out. It's complicated and if the authors are using their predictions to recycle back to the wet-lab .... thats what it's all about.
Normally, folding is calculated using graph theory ... FYI.
From there comments .... whats the program ... its Kapetanovic (2008) as cited above, which will be a classic small molecular binding algorithm the authors called QSAR
.
In cases there's any misunderstanding - we wouldn't use small molecule interaction based calculations (MD) to understand RNA folding/binding because there's an alternative body of theory for this specific calculation, partly thermodynamics, partly a separate factor. ML is heavily used but not via an MD style calculation - which is what the authors are doing. Thus if you use the method there is a risk of criticism.
It's complicated - I understand the authors perspective, they are saying classical style hybridisation its just not that simple. It might be complicated but hybridisation shouldn't be ignored as the empirical assumption.