A useful ADMET profile for a series does not require a cheminformatics team. It requires knowing which endpoints actually change your next decision, insisting that every predicted value arrives with an uncertainty and a domain flag, and treating the output as a ranking that tells you which two assays to run. The common failure is the opposite: a spreadsheet with eighty predicted columns, no uncertainties, and no decision attached to any of them.
Start from the decision, not from the endpoint list
For a series in lead optimisation, three decisions usually matter: which compounds to synthesise next, which to put into the first in vivo study, and which liability to engineer out. Each decision needs a different, small set of endpoints.
| Decision | Endpoints that move it |
|---|---|
| Which to synthesise next | Solubility, lipophilicity, pKa, permeability, metabolic stability proxy |
| Which to take into the first animal study | Metabolic stability, plasma protein binding, CYP inhibition, hERG, mutagenicity structural alerts |
| Which liability to engineer out | The endpoint that fails, plus the properties that will move with the change |
Everything else is context. Predicting eighty endpoints costs the same as predicting eight, but reading eighty costs attention you then do not spend on the two that matter.
Four properties that carry most of the early decisions
Lipophilicity and solubility.
Together they set the ceiling on formulation and dose. A series drifting towards higher logP and lower solubility is drifting towards a problem whatever else improves.
pKa.
It determines charge state at the relevant pH, which in turn drives solubility, permeability and the ionisation-dependent part of binding. A predicted pKa is enough to sort an ionisable series; where the value decides, measure it.
Metabolic stability.
The most common reason an otherwise good compound fails in vivo. Predictions identify likely sites of metabolism, which is actionable: block the site, or move the substituent.
Structural alerts for mutagenicity.
Cheap to check, expensive to discover late. A hit is not a verdict; it is a reason to look at the mechanism and, if the series continues, to plan the test.
How to read a prediction
Four habits make predictions usable rather than decorative:
1. Require an interval. A single number implies precision that does not exist. A value with a stated interval, and a statement of how often that interval contains the truth on held-out data, is a result you can reason about. 2. Require a domain flag. Predictions for compounds unlike anything the model was built on are guesses. A flag that says "outside the chemistry I know" is more valuable than a confident number. 3. Read differences, not values. Within a congeneric series, predicted differences between analogues are more reliable than absolute values. "B is worse than A by more than the interval" is a usable conclusion; "B is 1.8" often is not. 4. Ask how the accuracy was measured. Any performance figure should state the metric, the sample size and the split rule. A figure from a random split on chemical data is optimistic, because near-duplicates land in training and test. Measurement on a scaffold-split held-out set is the question you care about.
A workable sequence for a 40-compound series
1. Prepare the structures. Consistent, standardised, with salt forms and stereochemistry explicit. Most disappointing profiling exercises are input problems. 2. Profile all 40 on the short endpoint list. Keep the full output including intervals and flags. 3. Separate three groups: clearly acceptable, clearly problematic, and uncertain. The uncertain group is the one the predictions cannot resolve, and it is where the experiments go. 4. Run two assays, not six. Pick the two endpoints that most constrain the next decision, measure them on a spread of the series including the extremes. 5. Calibrate. Compare measured against predicted for those compounds. The offset and spread tell you how much weight the remaining predictions deserve for this chemotype. 6. Then rank and decide. With a calibration in hand, the predictions for the unmeasured compounds have a known reliability, which is the condition for using them.
Step 5 is the one most often skipped and the one that makes the difference. Twenty measurements convert a generic model into a calibrated instrument for your series.
Limits of the statement
- Predicted ADMET values are triage. They prioritise experiments and prevent obvious mistakes; they do not substitute for the assay where the number enters a decision with consequences.
- Models are weakest exactly where new chemistry lives. For an unusual scaffold, expect the domain flags to be set and plan measurements accordingly.
- A profile says nothing about efficacy. A compound can be excellent on all eight endpoints and inactive.
- Any assessment that goes into a regulatory context, for example a mutagenicity evaluation, has its own requirements and belongs with a qualified assessor.
Frequently asked questions
Which ADMET endpoints should a small team predict first?
Solubility, lipophilicity, pKa, a permeability estimate and a metabolic stability indicator for synthesis decisions; metabolic stability, plasma protein binding, CYP inhibition, hERG and mutagenicity alerts before an in vivo study. Choose by the decision you are about to make, not by availability.
Can predicted ADMET values replace assays?
No. They rank candidates and decide which assays to run on which compounds. Where the value enters a decision with real consequences, it needs measurement.
What does an applicability domain flag mean in practice?
That the compound lies outside the chemical space the model was built on, so its prediction is an extrapolation. Treat a flagged prediction as a reason to measure, not as a value.
Why are predicted differences within a series more reliable than absolute values?
Because systematic bias largely cancels between closely related compounds. A model can be consistently 0.5 units off and still order analogues correctly, which is what a synthesis decision needs.
How do I judge whether a published accuracy figure is relevant to my compounds?
Check the metric, the sample size and the split rule, and whether your chemistry resembles the test set. A figure from a random split overstates performance on genuinely new compounds; a scaffold-split figure is the closer analogue of your situation.