When similar molecules disagree

Imagine two compounds with nearly identical structures whose potency differs by an order of magnitude. Such an activity cliff is where structure-activity intuition gets interesting: a small structural change coincides with a disproportionately large change in measured activity.

The famous lock-and-key picture illustrates the intuition: two keys may be cut almost identically, while one opens the lock and the other does not. The difference between the keys is worth studying because it may reveal something important about molecular recognition.

But the picture is only an analogy. Protein-ligand binding is more dynamic than a rigid lock and key, and an activity cliff does not by itself tell us whether the cause is a specific binding interaction, conformational change, solvation, assay context or something else.

Making the cliff a quantitative definition

Informally, everyone knows a cliff when they see one. Finding them systematically requires a commitment to numbers: structural similarity as Tanimoto similarity over ECFP4 fingerprints above a chosen cutoff, and activity difference as a delta above a threshold on the pChEMBL scale.

ChEMBL’s pChEMBL is a negative logarithmic representation of standardized molar activity measures such as IC50, EC50, Ki, Kd and potency. A difference of one pChEMBL unit corresponds to a tenfold difference in the underlying molar value.

The precise thresholds are tunable and depend on the use case. The important point is that once the definition is explicit enough, cliff detection turns into a reproducible query.

SARVault computes fingerprints into the marts layer of the warehouse, forms candidate pairs within the same target from standardized potency records and scores them. “Show me the cliffs for this target” becomes a filter over a table that is re-derived from the source on every build.

For ADC payload chemistry, where small structural changes can affect multiple biological properties, such a list can serve as a reading guide to a target’s most instructive compound pairs - provided the activity endpoint and assay context remain comparable.

Many apparent cliffs are noise

Pitfalls can be mundane. Two measurements of the same compound in different assays might differ enough to fabricate a cliff from experimental scatter. Duplicate records, censored values and careless aggregation across assay types can do the same.

A cliff rests on the comparability of its two endpoints - which is why the aggregation rules live upstream in the warehouse, where they are versioned and tested.

The working order is therefore fixed: define the cliff quantitatively, compute it from curated and comparable pairs, and treat every result as both a data-quality signal and a chemistry insight.

That order holds well beyond molecules. Any analysis of surprising pairs - whether churned twins in customer data or divergent batches in manufacturing - starts with definition, deduplication and only then interpretation.