Rank by what you can see
Virtual screening produces a lot of numbers for every candidate. A docking score, the number and type of interactions with the target, how often a binding mode recurs across poses, how buried the ligand is. The usual next step is to combine them into one composite score with a set of weights and sort by that.
I have stopped doing this in my own pipelines. I sort on the raw metrics one after another instead, in an order I decide and write down.
What goes wrong with a weighted score
- The weights are a choice that looks like a result. Someone picked them, often by feel, and the final ranking depends on them as much as on the chemistry.
- Trade-offs disappear. A candidate with an excellent docking score and almost no specific interactions can land next to one with the opposite profile. The single number hides that these are very different molecules.
- Nobody can retrace it. When a reviewer asks why one candidate beat another, the honest answer is a weighted sum of five things on different scales.
What sequential sorting looks like
- Sort by the metric I trust most for the question at hand.
- Within ties or a tolerance window, sort by the next metric.
- Keep going until the list is short enough to inspect by eye.
Every step is a statement anyone can check. The cost is that the order of metrics is itself a decision. I accept that, because it is one visible decision instead of several hidden ones, and I can rerun the ranking with a different order to see how much it matters.
When a composite score is still fine
When the weights come from data, for example a model fitted against measured affinities, a combined score is a result of that fit rather than a matter of taste. For early-stage filtering, where no such data exists, I prefer a ranking I can explain in one sentence per step.