The regulation of metabolites is typically controlled by multiplicative effects (1,2). It is therefore reasonable to look for fold-differences rather than for absolute differences in such concentrations. On the log-scale, concentrations often show a quite symmetric distribution, so standard methods of analysis using the log-concentrations are likely appropriate. Hence, you can calculate the mean difference in the log-concentrations. This difference of logs is the same as the the log of the ratio, so it is a log-ratio. The confidence interval for this estimated log-ratio should also be calculated on the log-scale, using the t-distribution. It gives you an upper and lower limit for the log-ratio. If you later wish to provide ratios rather than log-ratios, you can antilog the mean and the limits of the confidence interval determined on the log concentrations.

I am not sure why you want to test and what you want to test. You might test the null hypothesis that the mean log-ratio is zero. Then the (unpaired) t-test would be the correct tool. However I don't see why this null hypothesis should be reasonable and needs to be tested. Why should both species have exactly the same amount of this metabolite? I would consider this as wrong a priori. There is no test required to convince (me) that the concentrations might be different. The more interesting research question -to my opinion- would be: how much different are these concentrations, and this is an estimation problem (as explained in the first paragraph above).

(1) PRL 97, 168302 (2006)

(2) Mathematical Biosciences 68:187-212 (1984)

* The papers actually demonstrate that such data is gamma-distributed, so one may use a gamma model to analyze such data, but typically analyzing the log data with a "normal" model is sufficiently good.

A paired t-test might be appropriate if particular pairs of measurements (one for E.coli and one for Salmonella) share a common source of variance. For instance when two such preparations were performed in parallel and measured parallel. Then the next pair is prepared and measured in parallel. If the preparations ore measurement runs will give varying results, then it would be beneficial to use a paired t-test.

The pairing structure can also be used to increase the precision of the estimate (of the log-ratio): first compute the pair-wise differences of the log-concentrations, and then determine the mean and the confidence interval for these differences.

If you wish to account for more covariables you can fit a linear model (or then better a generalized linear model of the gamma-family with log-link). But this is presumably going too far here.