This is 100% not the type of post I envisioned starting this blog off with. I kind of just “lucky dipped” my draft backlog of projects and posts. Anyway, here we are. The prediction of secondary RNA structures is an interesting and important problem in bioinformatics. This post will introduce the **Nussinov-Jacobson** algorithm for predicting secondary fold structures of RNA sequences. It is a dynamic programming algorithm and was one of the first developed for the prediction of RNA structures as RNA combinatorics can be quite involved and expensive.

## Dynamic Programming

Before looking at the Nussinov algorithm, it may be worthwhile recapping some dynamic programming (DP). DP consists of programming in such a way where one uses past knowledge of the problem, to make further solving easier, and as a result, reducing complexity. More concisely, we say that DP is a technique used to avoid computing the same thing multiple times for an exponential number of times by breaking a task down into sub-problems. It’s a great technique for powerful algorithm design, path-finding problems, and reducing time complexity. My professor once called it “careful bruteforce”. As a side-note, I hate the name.

An example of DP can be shown when computing the Fibonacci sequence. If I asked you to compute fibonacci to the 1 millionth integer… that’s a struggle. What if I gave you the two previous integers? Now it’s just a matter of a single, large, addition.

A common way of computing fibonacci is by implementing it recursively:

def fib(n): if n <= 2: return 1 return fib(n - 1) + fib(n - 2)

Even though the above implementation works, it’s not at all efficient as `n`

gets bigger. Also, imagine having to recurse \(100000\) times. Its time complexity is also less desirable since it is exponential. We can represent `fib`

as an equation, \(T(n-1) + T(n – 2) + \mathcal{O}(1)\), meaning the time to compute for `fib(n)`

is equal to the sum of time taken to calculate `fib(n - 1)`

and `fib(n - 2)`

plus a time constant. This produces an upper bound of \(\mathcal{O}(2^n)\), or more specifically \(\mathcal{O}(2^\phi)\), where \(\phi\) is the golden ratio.

Delving into DP, we can implement the same function above using **memoization**. The idea of memoization is a simple one, especially for our `fib`

use case. Whenever we compute a number in the sequence, we store it in a cache type variable – here, in Python, we use a dictionary. So now, when we *are* computing our `fib`

sequence, we can check our cache variable for numbers that have already been computed so as to not have to re-compute them. Memoization is a great “little” technique to speed up and optimize some programs and computations. Let’s rewrite our `fib`

function using it:

mem = {} def fib(n): if n <= 2: return 1 if n in mem: return mem[n] mem[n] = fib(n - 1) + fib(n - 2) return mem[n]

Ok, so, let’s do a quick time complexity analysis. Any non-memoized calls, i.e, when we call `fib(n)`

, only costs us \(n\), and any memoized calls costs us \(\mathcal{O}(1)\). Therefore we have \(\mathcal{O}(n)\). What an improvement already! Take note how we’re also breaking down our task into smaller little sub-problems by introducing more variables and processes, rather than just one recursive function.

Let us finalize our DP approach and implement `fib`

** bottom-up**. When we program recursively, we usually start thinking about how to code it from the top of what we want to solve, and work our way down. With a bottom-up approach we… well, we do the opposite: we work our way to the top from the bottom. Instead of having the recursive call to `fib`

, we implement a `for`

loop. Here it is:

def fib(n): f2 = 0 f1 = 1 if n <= 2: return 1 for i in range(1, n): f = f1 + f2 f1, f2 = f, f1 return f

As you can see, a much nicer, more optimized, faster, and efficient solution to our simple fibonacci problem! We end up needing only the last two numbers to calculate the next number, as opposed to having to recalculate the entire fibonacci “tree” for each \(n\).

For empiricism, here are run times with all three methods of computing Fibonacci:

- Recursion:
`fib(40)`

– \(\sim\) 21.5s - Memoziation:
`fib(250)`

– 258 µs \(\pm\) 4.88 µs per loop (mean \(\pm\) std. dev. of 7 runs, 10000 loops each) - Bottom-up:
`fib(250)`

– 137 µs \(\pm\) 11.2 µs per loop (mean \(\pm\) std. dev. of 7 runs, 10000 loops each)

## RNA & Nussinov

RNA, DNAs less famous cousin, is in charge of carrying out and transferring the genetic messages set out by DNA to your body for tasks such as protein synthesis. DNA codes and acts as a blueprint for your genetic traits, while the various types of RNA convert that genetic information into useful proteins by transferring the DNA blueprints out of the nucleus and into the ribosome. Both DNA and RNA are *pretty cool*.

Formally now, we can define RNA as a sequence \(S\), made up of four (nucleotide) bases; \(S \in \{G, C, A, U\}\). These 4 bases come in pairs, known as Watson-Crick pairs, and couple as follow: \(G \leftrightarrow C, \; A \leftrightarrow U\). These pairs are the “strong/stable” base pairs, however, in reality, we also have the coupling of less stable pairs, known as *canonical pairs*, namely \(G \leftrightarrow U\).

Some types of RNA can form secondary structures by re-pairing up with itself. These secondary structures change the RNA properties drastically. This happens because, unlike DNA, which has a double helix structure, RNA has a single strand structure. Because of this, the RNA sometimes folds in on itself creating these secondary structures. The process of folding can also generate specific patterns or shapes in the RNA structure, such as helixes, loops, bulges, junctions, and pseudoknots. These can occur because of the nature of the fold, or because a nucleotide has had a mismatch pairing. For example, in the image below, \(U\) can not couple with itself, hence creating a small bulge.

There are several approaches to solve the problem of predicting RNA secondary structures, especially with recent developments in the field of bioinformatics and computer science. However, in 1978, a simple yet very effective technique was proposed by Ruth Nussinov et al. which involved a dynamic programming algorithm to predict this secondary structure of an RNA sequence. An extra step is involved with Nussinov’s dynamic programming approach, and that is **backtracking**. Backtracking allows us to abandon partial solutions as soon as it’s determined that our “in progress” solution can not become a complete solution. If this doesn’t make sense, it soon will upon implementation. The algorithm allows for computing a maximum matching size of \((\sim10^3-10^4)\).

I mentioned that these secondary structures can fold into shapes and patterns, one being a pseudoknot. Pseudoknots are ignored by the Nussinov algorithm as they break down dynamic programming algorithms due to being too complicated to handle and perform backtracking on. In order to enforce this, we will set a minimal loop length in our code so that we do not result in nonsensical base-pairs and loops. Novel algorithms can however support them.

Formalizing the problem once more, we have:

- An RNA molecule as a string, \(S\), where \(S \in \{G, C, A, U\}\) of length \(L = |S|\)
- A secondary RNA structure for \(S\), is a set \(P\) of ordered base pairs, written as \((i, j)\)
- \(j – i > 3\) so as to not have bases too close to each other and avoid pseudoknots

The Nussinov algorithm solves the problem of predicting secondary RNA structures by maximizing base pairs. This also “solves” the problem of predicting non-crossing secondary structures/pseudoknots. This is achieved by assigning a score to our input structure within an \(L \times L\) matrix, \(N_{ij}\). To do this, for every paired set of nucleotides, we give it a score of \(+ 1\), and for others, \(0\). We then attempt to maximize the scores and backtrack on the nucleotides which maximize our overall score. To maximize our base pairs, Nussinov states only 4 possible rules we may use when comparing nucleotides.

The main Nussinov calculations are recursive and calculate the best secondary structures for small subsequences of nucleotides until it reaches larger ones.

*Idea:*

- Add unpaired position \(i\) onto best substructure for subsequence \(i + 1, j\)
- Add unpaired position \(j\) onto best substructure for subsequence \(i, j – 1\)
- Add paired bases \(i – j\) to the best substructure for the subsequence \(i + j, j – 1\)
- Combine two optimal substructures \(i, k\) and \(k + 1, j\)

Therefore, we now have:

\(\gamma(i, j) max \begin{cases} \gamma(i + 1, j) \\ \gamma(i, j – 1) \\ \gamma(i + 1, j – 1) + \delta(r_i, r_j) \\ max \{\gamma(i, k) + \gamma(k + 1, j) \} \end{cases}\),

where

\(\delta(i, j) = \begin{cases} 1, \text{if } x_i – x_j \text{is a pair} \\ 0, \text{else.} \end{cases}\)

As we’ve now formalized everything, we can begin putting the ideas into practice and writing the relevant code. Let’s get the basic boilerplate functionality going:

import numpy as np def couple(pair): pass def fill(nm , rna): pass def traceback(nm, rna, fold, i, L): pass def dot_write(rna, fold): pass def init_matrix(rna): M = len(rna) # init matrix nm = np.empty([M, M]) nm[:] = np.NAN # init diaganols to 0 # few ways to do this: np.fill_diaganol(), np.diag(), nested loop, ... nm[range(M), range(M)] = 0 nm[range(1, len(rna)), range(len(rna) - 1)] = 0 return nm if __name__ == "__main__": rna = "GGUCCAC" # our RNA sequence nm = init_matrix(rna) nm = fill(nm, rna)

Ok so we now have some code that returns a matrix of size \(M \times M\), where \(M\) is the length of our RNA sequence, with the first two diagonals set to \(0\) and the rest are empty NaN values. Take note we use `np.empty`

rather than `np.zeros`

to save memory, as `np.zeros`

would initialize every value in the matrix to an `int`

. From here, we can begin filling our matrix using Nussinov’s 4 rules. This step is \(\mathcal{O}(L^2), \mathcal{O}(L^3)\) in memory and time respectively.

def couple(pair): """ Return True if RNA nucleotides are Watson-Crick base pairs """ pairs = {"A": "U", "U": "A", "G": "C", "C": "G"} # ...or a list of tuples... # check if pair is couplable if pair in pairs.items(): return True return False def fill(nm, rna): """ Fill the matrix as per the Nussinov algorithm """ minimal_loop_length = 0 for k in range(1, len(rna)): for i in range(len(rna) - k): j = i + k if j - i >= minimal_loop_length: down = nm[i + 1][j] # 1st rule left = nm[i][j - 1] # 2nd rule diag = nm[i + 1][j - 1] + couple((rna[i], rna[j])) # 3rd rule rc = max([nm[i][t] + nm[t + 1][j] for t in range(i, j)]) # 4th rule nm[i][j] = max(down, left, diag, rc) # max of all else: nm[i][j] = 0 return nm

Running the code together now, we generate a Nussinov matrix given our RNA sequence. Visually, it works something like the GIFs below. There remains only one step now, and that is the traceback using backtrace (that’s like, 50% of a palindrome). Before I explain how the traceback works, let me explain why we need it. Given our now complete matrix, we need to find the optimal secondary structures. This is based, if you remember, on our base pair maximizations. Essentially, we will be looking at the pairs which contribute the most value to our RNA. To do this, we start at the top right position in the matrix, \(N_{0S}\), where \(S\) is `len(rna)`

, and traceback through our maximized base pairs pushing them onto a stack. This traceback is linear in both time and memory.

def traceback(nm, rna, fold, i, L): """ Traceback through complete Nussinov matrix to find optimial RNA secondary structure solution through max base-pairs """ j = L if i < j: if nm[i][j] == nm[i + 1][j]: # 1st rule traceback(nm, rna, fold, i + 1, j) elif nm[i][j] == nm[i][j - 1]: # 2nd rule traceback(nm, rna, fold, i, j - 1) elif nm[i][j] == nm[i + 1][j - 1] + couple((rna[i], rna[j])): # 3rd rule fold.append((i, j)) traceback(nm, rna, fold, i + 1, j - 1) else: for k in range(i + 1, j - 1): if nm[i][j] == nm[i, k] + nm[k + 1][j]: # 4th rule traceback(nm, rna, fold, i, k) traceback(nm, rna, fold, k + 1, j) break return fold

We now have our backtraced, optimal secondary RNA structure! There’s only some extra code left which is for output purposes. The general method of output from these algorithms, and to represent RNA sequences, is by writing them out using the dot-bracket notation. Essentially, we have open/closed parenthesis, and dots. An open parenthesis indicates the nucleotide is paired with another ahead of it, a closed parenthesis indicates that a nucleotide is paired with another behind it, and a dot is for unpaired bases.

def dot_write(rna, fold): dot = ["." for i in range(len(rna))] for s in fold: dot[min(s)] = "(" dot[max(s)] = ")" return "".join(dot)

$ python rna_nussinov.py G G G A A A U C C G 0.0 0.0 0.0 0.0 0.0 0.0 1.0 2.0 3.0 G 0.0 0.0 0.0 0.0 0.0 0.0 1.0 2.0 3.0 G NaN 0.0 0.0 0.0 0.0 0.0 1.0 2.0 2.0 A NaN NaN 0.0 0.0 0.0 0.0 1.0 1.0 1.0 A NaN NaN NaN 0.0 0.0 0.0 1.0 1.0 1.0 A NaN NaN NaN NaN 0.0 0.0 1.0 1.0 1.0 U NaN NaN NaN NaN NaN 0.0 0.0 0.0 0.0 C NaN NaN NaN NaN NaN NaN 0.0 0.0 0.0 C NaN NaN NaN NaN NaN NaN NaN 0.0 0.0 GGGAAAUCC .((..()))

Done! This was a super fun mini project and I learned a lot about RNA, DNA and other bioinformatics areas since I kept going deeper and deeper every time I researched something. I’d like to be able to now produce a nice graphviz plot with the dot-bracket output given and maybe even try and calculate base-pair fold probabilities. Who knows, I may even take a look at some of the more novel algorithms with support for pseudoknots… If you’d like to play around with this kind of stuff, I’d recommend checking out “RNAStructure” from MathewsLab research group. It’s not FOSS, but it is free to play with.

You can grab the code directly from here: rna_nussinov.py

**Resources and References:**

Virginia Tech University, CS382

MIT

University of Wisconsin

Algorithms in Bioinformatics

RNA structure analysis

Freiburg RNA Tools

RNA secondary structure prediction