One of the most important tasks that computational chemists can perform is to tell experimental chemists what *not* to make, sometimes based on simple sterics or electronics-based arguments and sometimes on statistical analysis that points to affinity confounders. Intelligently constraining the space of purchasable or synthesizable compounds is quite valuable, especially when confronted with hard-to-access building blocks for lead optimization or challenging scale-up processes.

But it's not easy to find good data-based metrics on how much effort these kinds of "negative suggestions" save. That's why it was valuable to find an assessment in this 2010 report on the evolution of molecular modeling at Merck, a longtime pioneer in the field.

"It turns out that predicting ‘‘not active’’ is a lot easier than predicting ‘‘active’’ since many compounds are not active in counter-screening. Today, we are counter-screening 30–40% fewer compounds per program because they are predicted to be ‘‘not active’’ which saves the company hundreds of thousands of dollars a year."

This approach speaks more generally to the utility of computational approaches which are still sometimes regarded as either completely marginal or able to provide the one, true, winning chemical structure. In practice they are almost never the former, and rarely the latter. In most cases modeling is valuable for narrowing down the chemical space, focusing on certain parts of a molecule that may or may not be modified, suggesting molecules within property, cost or other constraints and occasionally revealing insights like dynamic motions of proteins or unusual binding modes that may not be apparent from static structural data or limited SAR data.

One way to think about computation in drug discovery is as a relatively cheap (compared to experiments) “gambling” machine that improves the odds of success by providing multiple shots on goal than what you would have had otherwise. But when you think about it, that’s basically the value of any approach in the statistically improbable process of acquiring a lead or optimizing it into a clinical candidate or marketed drug. What you are trying to do is make the dice more loaded so that you get desirable outcomes more frequently than random chance. And computation can help load the dice.