QED: The Quantitative Estimate of Drug-likeness

Two QEDs, One Acronym: A Quick Disambiguation

If you searched for QED expecting Feynman diagrams and photon-electron interactions, you have landed on a different QED entirely. In cheminformatics and drug discovery, QED stands for the Quantitative Estimate of Drug-likeness, introduced by G. Richard Bickerton and colleagues in a 2012 Nature Chemistry paper. It has nothing to do with quantum field theory; the shared acronym is pure coincidence. Drug-likeness QED asks a much more down-to-earth question: given the structure of a candidate molecule, how similar are its physical and chemical properties to those of drugs that have actually made it to market? Instead of answering with a blunt yes or no, QED returns a smooth score between 0 (very unlike a drug) and 1 (very drug-like), making it easy to rank thousands of candidate molecules against each other.

The Problem With Pass/Fail Rules

Long before QED, medicinal chemists leaned on simple heuristics like Lipinski's Rule of Five, which flags a molecule as likely to have poor oral absorption if it violates more than one of a handful of thresholds: molecular weight over 500, calculated LogP (a measure of fat-versus-water solubility) over 5, more than 5 hydrogen bond donors, or more than 10 hydrogen bond acceptors. The trouble is that these are hard cutoffs. A molecule with a molecular weight of 499 and one with a molecular weight of 220 both get a full pass, even though the second is comfortably inside the sweet spot for oral drugs while the first is barely scraping by. Conversely, a molecule that fails by the tiniest margin on one property gets treated exactly the same as one that fails catastrophically on all four. This throws away exactly the information a chemist most wants: not just whether a rule was broken, but by how much, and how that shortfall trades off against everything else that is going right or wrong with the molecule. QED was built to keep that nuance instead of discarding it at the pass/fail gate.

Desirability Functions: Turning Raw Numbers Into Smooth Scores

QED's core building block is the desirability function, a concept borrowed from industrial statistics. For each molecular property that matters, such as molecular weight, LogP, the number of hydrogen bond donors and acceptors, polar surface area, the number of rotatable bonds, the number of aromatic rings, and the presence of flagged structural alerts, a desirability function d_i maps the raw value of that property onto a smooth curve between 0 and 1. A molecular weight near the middle of the range typical for approved oral drugs scores close to 1, while a molecular weight far out in the tails, either too small or too large, scores close to 0, with a gradual, continuous slope in between rather than a sudden cliff. Bickerton and colleagues fit the shape of each desirability curve empirically, using the actual distribution of that property across a large set of known oral drugs, so the curves reflect what real successful medicines look like rather than an arbitrary rule of thumb. The result is eight separate 0-to-1 desirability scores, one per property, each capturing a graded sense of goodness instead of a binary verdict.

Why a Geometric Mean, Not an Average

Once you have eight desirability scores d_1 through d_n for a molecule, you need to combine them into one overall QED value. Bickerton's team deliberately chose a geometric mean over a simple arithmetic average, computed as: QED = exp( (1/n) * sum_{i=1}^{n} ln(d_i) ). Equivalently, this is the n-th root of the product of all the d_i values. The reason this matters is how each combination method handles a single very poor property. Under an arithmetic mean, one property scoring d_i = 0 could still be diluted by seven other properties scoring close to 1, leaving a deceptively respectable overall average. Under the geometric mean, a single d_i near 0 makes ln(d_i) plunge toward negative infinity, dragging the entire sum, and therefore the final QED, sharply down no matter how well the other properties score. This mirrors real medicinal chemistry intuition: a molecule that is otherwise perfect but carries a severe toxicity-linked structural alert, or is wildly too large to cross a cell membrane, should not be rescued by looking good everywhere else. The geometric mean enforces that no single fatal flaw can be averaged away, giving QED a built-in intolerance for one-property disasters that an arithmetic mean would happily paper over.

From Academic Metric to Everyday Screening Tool

Because QED collapses many molecular properties into one comparable, continuous number, it has become a standard way to rank and prioritize candidate molecules in drug discovery pipelines. When a virtual screening campaign or a combinatorial chemistry library produces tens of thousands of candidate structures, QED lets a team sort that entire list by drug-likeness in seconds, focusing synthesis and testing effort on the molecules most likely to behave like a viable oral medicine. QED has also become a popular reward signal in generative chemistry and de novo molecular design: machine learning models that dream up novel molecular structures are frequently trained or fine-tuned to favor outputs with high QED scores, alongside other objectives like predicted binding affinity or synthetic accessibility, so that the generated candidates are not just novel but plausibly developable. It is important to remember what QED is not, though: it is not a prediction of biological activity, safety, or efficacy against any specific target, only a statistical statement of how closely a molecule's bulk properties resemble those of past successful oral drugs. Bickerton's original desirability functions were fitted on hundreds of orally administered drugs already approved for human use, which is precisely why QED captures the physicochemical patterns that correlate with a molecule successfully becoming a marketed medicine, even though it says nothing about whether a specific new molecule will actually work against its intended disease target.

Frequently asked questions

Try it live

Everything above runs in your browser — open QED: The Quantitative Estimate of Drug-likeness and change the parameters while it is running. Nothing is installed, nothing is uploaded, the whole model lives in one tab.