2
$\begingroup$

I'm reading Trifinov et al. (2013) and I am trying to understand the designed framework. I list some excerpts to focus on the passage that is not clear to me:

A lesion $M$ is a result of a random process with a given distribution $P^m(M)$, which specifies the probability of occurrence of the lesion $M $ due to $m$utation.

[...] To model selection we assume that the genotype resulting from a disjoint set of such lesions $S = \{M_1, ⋯, M_t \}$ expresses the phenotype when a lesion in the sequence contains a driver gene. We consider two possibilities for the posterior probability that a given gene $D$ is a driver of the phenotype, given that the genotype of the set $S$ expresses the phenotype: a global posterior and a local posterior. To obtain the two posteriors we define first the likelihood that a lesion $M$ is a driver lesion given that $D$ is a driver gene as

$$L(M|D) \quad \underline{deff} \quad P^m(M)·δ(D ∈ M)$$

where $δ(D ∈ M)$ is $1$ if the lesion $M$ includes the gene $D$ and $0$ otherwise, with corresponding posterior

$$P(D|M)= \frac{δ(D∈M)·P^d(D)}{∑_{G∈M}P^d(G)}$$

where $P^d(D)$ is the prior probability that $D$ is a $d$river of the phenotype.

I am trying to understand how the posterior has beed computed, i.e. how to pass from the first to the second equation.

I have little experience on bayesian statistics, except for the basic understanding required by the Wikipedia pages (1, 2). For instance, following from the Wikipedia's likelihood definition, first definition translates to:

$$L(M|D) = P(D|M) = P^m(M)·δ(D ∈ M)$$

So, the probability that $D$ is a driver gene given that $M$ is a driver lesion is equal to the probability of lesion $M$ due to mutation if $D$ is in $M$, $0$ otherwise. Is it correct?

Otherwise, could it be that $L(M|D)$ and $P(D|M)$ are both probabilities, in particular the first defining the probability of $M$ given $D$, and the second defining the posterior distribution of $D$ given $M$?

If this is the case, it unfolds to:

$$P(D|M)=\frac{P(M|D)·P(D)}{P(M)}$$ $$=\frac{P^m(M)·δ(D ∈ M)·P^d(D)}{P^m(M)}$$ $$=δ(D ∈ M)·P^d(D)$$

But in this case, the sum at the denominator of the second equation is missing.

How can I reduce the calculation of the posterior? What can I read to have a solid understanding of it?

Edit

I am now pretty sure that the derivation of the second equation should follow simply using the Bayes theorem, i.e.:

$$posterior = \frac{likelihood·prior}{marginal\ likelihood}$$

with

$likelihood=P^m(M)·\delta(D \in M)$

$prior=P^d(D)$

but I am unable to compute the marginal likelihood correctly. How can it be done?

$\endgroup$

1 Answer 1

1
$\begingroup$

You have the right idea in your second attempt, using Bayes' theorem. However, keep in mind that $P(D \mid M)$ is the probability that $D$ is a driver gene given that $M$ expresses the phenotype. Therefore, the denominator must be the marginal probability that $M$ expresses the phenotype, which we can write as $$P(M) = \sum_G P(M \mid G) P(G).$$ This is different from $P^m(M),$ which is the probability that lesion $M$ occurs due to mutation.

Substitute in the above equation to the denominator, and keep in mind that $\sum_G \delta(G \in M) f(G) = \sum_{G \in M} f(G).$ Then you should get the expression given in the passage.

EDIT: While the above assumes that $L(M \mid D) = P(M \mid D),$ which is how I interpret their expression for the "likelihood", it's possible that they simply mean that $L(M \mid D) \propto P(M \mid D)$ where the proportionality constant is independent of $D.$ If that's the case the math still works out, as you can simply cancel the proportionality constant from the numerator and denominator. I should clarify that I'm simply showing how to get to the result purely from a mathematical perspective. I can't speak for a conceptual understanding here because I don't know this field, nor did I read the paper.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.