Bioisosterism and activity cliffs

You have a group that is causing a problem — it is metabolised too fast, it is too polar to get in, it is a liability. A bioisostere replaces it with something that keeps what you need and changes what you do not. This module closes the course on the limits of that idea.

The original notion was narrow: classical bioisosteres are groups with the same number of valence electrons and similar size. Fluorine for hydrogen, oxygen for sulfur or NH, a carbonyl for a sulfonyl.

Non-classical bioisosteres abandon the electron-counting and keep only the functional equivalence: they mimic the properties that matter — the charge, the geometry, the hydrogen bonding pattern — while looking quite different. A tetrazole for a carboxylic acid is the standard example, and these are where nearly all the practical value is.

Work through these. For each one, the question to ask is the one that organises the whole module: what is it meant to preserve, and what is it meant to change?

Four of them are worth commentary.

Carboxylic acid to tetrazole. Preserves the negative charge at pH 7.4 and roughly the geometry of the anion. Changes the lipophilicity — the tetrazole is more lipophilic despite being charged — and the metabolic stability. It is bigger. The sartans all carry one.

Ester to amide. Module 3's example. Preserves size and shape; changes a group cleared in minutes for one that survives, and adds a donor.

Methyl to trifluoromethyl. Preserves rough steric bulk. Changes everything else: strongly electron-withdrawing, more lipophilic, and metabolically inert where the methyl was a hydroxylation site.

Phenyl to bicyclo[1.1.1]pentane. A modern one. Preserves the distance between the two substituents — the 1,4-vector. Removes the aromatic ring entirely, raising fraction sp³ and improving solubility, at the cost of any stacking the ring was doing. This is module 13's advice made concrete.

Bioisosterism applied to the core rather than the periphery. Replace the central framework with a different one that presents the same substituents in the same directions.

The distinction matters because the risks differ. Changing a peripheral group is a local edit with a local effect. Changing the core changes every vector at once, and the compound may simply not bind. Against that, a successful scaffold hop can escape a patent, fix a whole-molecule property problem, or get past a metabolic liability built into the framework — none of which peripheral edits can do.

A matched molecular pair is two compounds differing by one well-defined change. Collect enough of them from a large dataset and you can ask: on average, what does this transformation do?

The answers are genuinely useful. Across thousands of pairs, replacing a hydrogen with a fluorine changes logP by a predictable amount; replacing a phenyl with a pyridyl lowers it by about one unit; adding a methyl raises it by about half a unit. These are the numbers behind every medicinal chemist's intuition, made explicit and checkable.

The limitation is equally important. A matched pair analysis gives you the average effect across many different contexts. For your compound in your binding site, the effect can be anything at all — and the variance around those averages is large. Use them to form expectations, not predictions.

Almost everything in cheminformatics rests on one assumption: similar molecules have similar properties. Similarity searching, clustering, QSAR, machine learning on molecular fingerprints — all of it assumes that a small structural change produces a small property change.

Most of the time it holds. When it fails, it fails spectacularly.

An activity cliff is a pair of molecules that are structurally very similar and differ in potency by orders of magnitude. One atom, three log units.

The causes are usually specific and, in hindsight, sensible:

• The change completes or breaks a specific interaction — a hydrogen bond, a halogen bond, a salt bridge.
• It causes a clash, and module 15's repulsive wall is very steep.
• It shifts a pKa across physiological pH, changing the charge — module 10's imidazole problem.
• It changes the conformational preference, so the bioactive conformation becomes expensive — module 8.
• It causes a different binding mode entirely. The two compounds are not doing the same thing at all.

Notice that this list is a summary of the whole course. An activity cliff is where a molecule's chemistry stops averaging out and one specific thing matters — which is why understanding the chemistry is not optional, however good your models get.

Two practical consequences:

• Cliffs are where the information is. A pair that differs by one atom and three log units tells you exactly which interaction matters. They are the most informative compounds in any dataset.
• Cliffs are where models fail. A QSAR or machine-learning model trained on similarity will confidently mispredict a cliff, because its central assumption is violated there. When a model's prediction disagrees sharply with an assay, a cliff is one of the first things to look for.

You began by learning to read a skeletal formula. You can now open an unfamiliar structure and say what it is made of, what shape it can adopt and at what cost, what it will do in water at pH 7.4, what it will stick to and why, and how to triage it against the rest of a hit list.

You can also say, specifically and with reasons, what a computational method is likely to get wrong about it: the tautomer it silently chose, the protonation state it assumed, the water it deleted, the strain it did not count, the sigma hole it cannot represent, and the desolvation it approximated.

That last list is the point of the whole course. Computer-aided drug design is next, and the people who get the most out of it are the ones who know where its models stop.