This is only true if the response was log2-transformed prior to running the model.

The easiest way to think through is probably a toy example.

 set.seed(123)
 x <- c(rpois(50,100), rpois(50,75))
 y <- rep(factor(c("A", "B")), each=50)

That's 50 observations for each of 2 phenotypes, with the "true" fold difference between B and A being 75/100 = 3/4 ~ 2^-0.415.

If you fit a linear model the betas are going to tell you the mean of each group:

(betas <- lm(x~y)$coefficients)
# (Intercept)          yB 
#     100.06      -23.68 
(means <- tapply(x,y,mean))
#    A      B 
#100.06  76.38 
betas[1] + betas[2]  == means[2]
# TRUE

Obviously, the betas are not he same as the log2 fold-change. To get that you can either transform the ratio of the estimates

log(sum(betas)/betas[1], 2) 
# -0.3895985

Or perform the transformation before the model-fitting

 lm(log(x,2)~y)$coefficients
 # (Intercept)     yB 
 # 6.639029   -0.389812

EDIT

I forget to explain the math-sy reason why this works, which might not be immediately obvious. As we've seen, when we do a linear regression with a categorical predictor, the Beta values reflect difference in the mean value between groups. If we first log-transform the response values then, of course, we'll end up with a difference of logs. In the example that's log(100) - log(75) which, thanks to the magic of logs, is the same thing as log(100/75): the log fold-difference.

EDIT 2

To clarify about the small difference between taking the log of the ratio of betas, rather that first log-transforming the values. This arises because the log-transform also changes the shape of the distrbution and therefore the mean value. With the toy example, the means of the log transformed x (tapply(log(x,2), y, mean)) is slightly different than log-transform of the means of x (log(tapply(x,y,mean),2))

Thanks for your answer. So If I understood, you have 2 ways of calculating log2 fold change with betas:

1) In your exemple, the fold change is FC= mean( x( B ) ) / mean( x( A ) ). Though FC = 76.38/100.06 = 0.763342. And log2( FC ) = -0.3895985 = log(sum(betas)/betas[1], 2) .

2) If you log2 transform x before model fitting which is the case of gene expression usually after normalisation and summarization:

log2x <- log2( x ),
mean( log2x( B ) - log2x( A ) ) = -0.389812 = yB.

So in that case you have the log2( FC ) = beta.

How do you explain the small difference between -0.3895985 and -0.389812 ? Are they caused by some approximation ?