How to invent novel algorithm ?

Question

[Yean]

Hi all,

I am wondering what is the basic idea of modifying the original algorithm in order to develop novel algorithm.

For example in genetic risk score calculation, the original equation is GRS <- log(odd ratio) * (No. of risk allele).

But, the extended ones can be found in LDpred, metaGRS, lassosum as example where the various features was added into the original GRS equation in many way such as Bayesian approach (LDpred).

For my rough understanding (I dont know if it is correct), the modified one can be developed from the observation in that problem and try to add new variable into the original one in certain ways. After that the new one is subsequently validated.

What is concept behind this process or any detailed guide ?

Thanks a lot.

Yean

Answer 1

[Asaf]

It's called computer science.

Most of the times algorithm follows theory meaning someone thinks about the problem, do the math and comes up with an algorithm that follows the math. In other cases the intuition comes first and the math follows either way it should be based on math.

[Jean-Karim Heriche]

I'd add that implementations of algorithms also make heuristic trade-offs for practical reasons and some new algorithms are developed/derived from others because the originals are not practical under certain circumstances or to account for cases not originally considered. Take for example the various sequence alignment algorithms. We first got Needleman-Wunsch for global alignment then we got Smith-Waterman for local sequence alignments then we got FASTA and BLAST because the others were too slow.