Analyzing diseased-based data "backwards"? In my line of work, we have clinical samples from various stages of a pathology. We might measure levels of proteins, RNAs, etc.--we'll call them all "markers". Invariably, the presumption is that I would analyze the data thus:
$$\text{marker} \sim \text{Diagnosis } (\times \text{ covariates})$$
And present the results. Essentially, the model states that the marker is an outcome of the disease state, then we write the papers to try to demonstrate that the disease state is influenced by the marker level (the reverse of the model structure).
Conceptually, is there anything wrong with the following:
$$\text{Diagnosis } \sim \text{marker}_1 + \text{marker}_2 + \text{marker}_n\  (\times \text{ covariates})$$
as a binomial or multinomial model for "Diagnosis" and presenting the effects of markers in concert on the outcome of disease state? In other words, explicitly model what the paper is trying to claim. It also allows me to model several markers as a multiple regression (glm).
Or is this just not biologically valid?
 A: I'm not an expert on causal modelling, but it may be worth investigating what sort of DAG each model implies.  Then, you can compare this against the expert knowledge in your field to decide which model to pursue.
The first model you have written implies a causal model shown below.  Here, the biomarker is the outcome, and the covairates are a common cause of both the diagnosis and the biomarker.  Were we to condition on covariates and modulate the diagnosis, this would neccesarily change the biomarker.

Now, doesn't that seem a bit strange?  I'll repeat what I said: "A change in the diagnosis would cause a change in the biomarker".  If that is the case, your colleagues would have to present the mechanism by which the biomarker can somehow listen to our diagnosis.  I'm being a bit fast and loose here.  I haven't defined what I mean by diagnosis, and this model may be entirely plausible if perhaps the biomarker is indeed caused by some different disease state (e.g. perhaps the disease is effecting a different part of the body than the one the disease is primarily associated with).
Compare this to the alternative model you've presented.  Here the diagnosis is the outcome, and the causal phrase "modulating the bio marker causes a change in the diagnosis" doesn't seem as strange.

This post is not so much an answer of the kind you expect.  I don't think providing your colleagues with references will do you much good (for they could and likely will present papers to you which do it their way.  My experience with clinical work is that people want to do what other people are doing).  The solution will come from thinking about what sort of implications your model is making.  To that end, I think its important you provide us more details about the diagnosis, the study, the biomarkers etc so that we can reason through the model implications together.
