Computing candidate selectivity: how a manual re-check of the direction of action filtered out false hits and brought two methods into agreement

The case in brief

The client was selecting antagonist molecules ("silencers") against a single therapeutic target — a GPCR-class receptor. Automatic classification by experiment type produced 1401 "confirmed antagonists", but a strict manual check against the raw records showed the classifier was noisy: it had filed molecules with the opposite action — agonists — as antagonists. After cleanup, 19 chemically distinct, reliably confirmed antagonists remained. Only on this honest list did two independent affinity-calculation methods agree for the first time in the project (r = +0.44) — whereas on repurposed drugs they contradicted each other (r = −0.38).

The starting situation

The client was looking for candidates able to "silence" a therapeutic target — a GPCR-class receptor. The first pass went through approved drugs (a repurposing strategy): there, activity against the target of interest is incidental and weak in affinity, and two different methods of calculating the fit into the binding pocket contradicted each other (r = −0.38). That did produce a repurposing list, but relying on it was risky. The question remained: do the available chemical libraries contain compounds engineered specifically for this target — potent and, above all, selective in their direction of action — and can the fast automatic "antagonist / agonist" labelling be trusted?

What the analysis showed

We collected from the literature all 6286 measured activities against the target together with the experiment descriptions and classified the direction of action by assay type: suppression of stimulation = antagonist, activation = agonist. The automated pass returned 1401 "confirmed antagonists". After filtering for potency and drug-likeness (pChEMBL ≥ 7, Brenk-clean, QED ≥ 0.4), 391 molecules remained; they were grouped into 47 chemical families (Butina clustering, Tanimoto 0.4) so we could work by one representative per chemotype rather than blindly across the whole list.

• The fast automatic direction labelling proved unreliable. A strict manual check of the representatives against the raw records revealed that the classifier had filed molecules with the opposite functional action — known agonists — as "silencers". This is not a cosmetic error: agonist and antagonist are opposite in effect. After manual cleanup, 19 chemically distinct reliably confirmed antagonists remained — not 1401.
• On the honest list, the two methods agreed for the first time. An independent affinity check by a second method (a neural-network binding calculation) gave these specialised antagonists a positive correlation with measured affinity, r = +0.44 — against r = −0.38 for the "incidental" binders from approved drugs. Two independent signals confirmed each other for the first time in the project — as it should be when a molecule really was designed for this pocket.
• The lead is confirmed by both methods independently. The strongest candidate: measured affinity ~20 pM (pChEMBL 10.7), independent calculation 10.6, binding probability 0.83 (the highest in the set), a confirmed antagonist, representing a family of related molecules.
• Selectivity by site of action. By physicochemical rules of CNS penetration (mass, polar surface area, hydrogen-bond donors), 3 of 19 likely cross the CNS barrier while the other 16 stay in the periphery. This let us split them into two independent solutions on already-existing molecules, without synthesis.

Why this is value, not criticism

Had the client trusted the fast automatic labelling, compounds with the opposite functional effect — agonists instead of antagonists — would have gone into procurement and the lab experiment. This is not a shade of error but a flipped sign: reagents, months of experiments and conclusions would have rested on molecules doing the reverse of what was intended. The cost of such an error at the entry to preclinical work is out of all proportion to the cost of a computational re-check. A separate value is the honest limit fixed in numbers: the methods agree (r = +0.44) only on ligands genuinely designed for the pocket, and diverge on repurposed drugs (r = −0.38). That says where the calculation can be trusted and where you need your own experiment — before the money is spent.

Caveats and data status

We separate what is measured, what is a calculated estimate, and what is taken from the literature.

• The measured affinities (pChEMBL, including the lead's ~20 pM) are taken from published experiments, not generated on the client's bench.
• The direction of action (antagonism) comes from literature experiments; an in-house functional assay on the client's own system is mandatory.
• The second affinity-calculation method, the correlations (r = +0.44 and −0.38) and the binding probability 0.83 are computational estimates.
• CNS-barrier permeability is estimated by physicochemical rules (mass / polar surface area / hydrogen-bond donors), not measured — to be confirmed in PAMPA/Caco-2.
• The compounds are research-grade (bought for research, not pharmacy medicines), have not been through the clinic, and their safety profile is unknown.
• The therapeutic target, indication, specific tissues, client and market are not disclosed. Details — at a diagnostic session under NDA.

The case is anonymised. The therapeutic target, indication, client and market are under NDA. The measured affinities (pChEMBL, ~20 pM) are from published experiments; the second calculation method, the correlations (r = +0.44 and −0.38) and the binding probability 0.83 are computational estimates; CNS-barrier permeability is estimated by rules, not measured. The compounds are research-grade and have not been through the clinic. The numbers are quoted from a real project report.