This program obtains 1D arrays as its numerical input and combines them with categorical input which is nucleotide bases of RNA. Then, 1D-CNN accepts mixed input to produce a single numerical outcome, which is conservation score for each nucleotide of interest. One example of these type of models are shown below. As it looks like a graph with 2 branches, one brach (longer) is associated with numerical input and the other brach (little) is designed for accepting categorical variables.

Now, let's go through what the numerical input was: local context for each nucleotide. From the model, the longer brach was designed for taking numerical input with lenght 676. Here, we'll see where 676 comes from. This figure shows what we mean by "local context of a specific nucleotide" in 3D. It contains nucleotide itself with neighboring nts where each nt is made up of multiple atoms. This figure shows what a machine could understand/see when it has the local context of a nucleotide. Now, in this format, we have 15X15 (2D) array which includes cells and numbers indicating the number of atoms falling into each cell. Since we're in 3D, as you may guess that a unique 15X15 (2D) array would be sufficient to represent full information. Instead, we ended up with three 2D arrays which are shown below. Each 2D array refers a layer: bottom (-1), base (0), top (+1) where a base layer is where the nucleotide of interest itself belongs. Now, the idea we break down analzing local context in 3D turned into analyzing three 2D arrays. Then, we reduced the dimension from 2D to 1D and we ended up with a sequence of lenght 675 (3X15X15=675).

The way how we view 675-dimensional input can be given from different perspectives:

From the geometric point of view: Since we work with RNA 3D %structures which are made up of atoms in which each atom is represented by three coordinates (x, y, z). If we take 2D projections, one of these coordinates, z, should be removed, the coordinates of the atoms are reduced to 2 dimensions and the lost coordinate is compensated by 3 layers. Then, each 2D layer is projected onto 1D as sequence. Finally, 225 (15X15)-dimensional sequences in 1D from each layer is added to make 675 (15X15X3)-dimensional input in 1D.

From the biological point of view: 675-dimensional vector includes the number of nearby atoms falling into that region.

From the statistical point of view: 675 different features or variables or predictors.

By adding the total number of atoms found in the local context, we got 676.