Early ADMET Filters and Their Effect on Late-Stage Attrition

In our experience running virtual screening campaigns for specialty CROs and early-stage biotech teams, the most expensive mistake we see isn't bad docking. It's routing compounds to synthesis before anyone has looked at ADMET. Applying absorption, distribution, metabolism, excretion, and toxicity scoring at the hit identification stage, not at lead optimization, is one of the highest-impact changes a small discovery team can make. The cost of one failed synthesis batch averages $4,000 to $8,000 when you factor in reagents, FTE time, and instrument overhead. Multiply that by the 40 to 60 percent of batches that yield no binding activity above 10 µM, and the math on early filtering gets compelling fast.

Why the Rule of Five Is a Starting Point, Not a Finish Line

Lipinski's rule of five is still the most cited filter in drug discovery. Molecular weight under 500 Da, log P under 5, hydrogen bond donors under 5, hydrogen bond acceptors under 10. Most computational chemists have these numbers memorized. The problem is that rule-of-five compliance is necessary but nowhere near sufficient for oral drug-likeness, and teams sometimes treat it as a pass/fail gate that replaces the harder analysis.

Modern variants extend the framework. The Veber rules add rotatable bond count and polar surface area to predict oral bioavailability in rats. The Egan egg plots log P against PSA using slightly different boundaries. Gleeson's analysis of in-house Pfizer ADMET data produced tighter logP cutoffs than Lipinski, particularly for hepatotoxicity risk. None of these rules agree perfectly. Real talk: a compound that clears all of them can still fail metabolic stability screening. What they collectively do is shrink the space of likely failures early, which is exactly the point.

The practical move is to apply multiple orthogonal filters and score compounds on how many they clear, rather than treating any single rule as a hard gate. A compound failing two out of five lightweight filters deserves closer attention before routing to synthesis. One that fails four can probably be deprioritized without further analysis.

LogP and LogD: Where Most Teams Are Too Permissive

LogP is the octanol-water partition coefficient at pH 7 for the neutral form. LogD is the same measurement at a specified physiological pH, accounting for ionization state. For oral CNS candidates, logD at pH 7.4 is the number that matters. For peripherally acting targets, the cutoff can shift. The distinction matters.

In our tracking of compound short-lists from structure-based campaigns, a large fraction of early hits cluster in the logP range of 4 to 6. Scaffolds in that range dock nicely into hydrophobic pockets. They look great on paper. They also have elevated CYP3A4 inhibition risk, poor aqueous solubility, and are more likely to trigger hERG channel blockade. The sweet spot for oral small molecules, based on data from the DMPK literature and our own model outputs, is logD between 1 and 3. Compounds above 4 warrant an explicit solubility check before synthesis decisions.

Soft filter. Hard filter. The distinction is worth making explicit. A hard filter rejects compounds unconditionally when they fail. A soft filter flags them and adjusts their ranking. For logP and logD, we treat values above 5 as a hard reject only when combined with molecular weight above 450 Da. Otherwise, they're soft flags. Discarding a promising scaffold because it sits at logP 5.2 with an otherwise clean ADMET profile is a mistake. Sending a logP 6.8, MW 490 compound to synthesis without comment is a bigger one.

hERG Liability: The Filter Teams Skip Until It's Too Late

hERG channel inhibition causes QT interval prolongation, which in rare cases triggers fatal arrhythmia. The FDA has pulled drugs for this. The channel is also notoriously promiscuous: it has a large, lipophilic binding site that accommodates structurally diverse molecules with high affinity. Positively charged nitrogen-containing scaffolds, which appear constantly in medicinal chemistry, bind it readily.

This is one we've seen teams deprioritize at hit identification on the assumption that it's a lead optimization concern. That assumption is expensive. If the scaffold that generates your best binding hit against the target of interest has a predicted hERG IC50 below 1 µM, you need to know that before committing to six analogues for SAR. Not after.

Predictive hERG models vary in quality. ADMET Predictor's hERG classification has been validated against in-house assay data from multiple pharma programs with reasonable AUC values in the 0.80 to 0.85 range. QikProp includes hERG activity predictions based on logP and charge state. SwissADME does not include hERG, which is one of its meaningful gaps for this use case. For open-source alternatives, the DeepTox and CardioTox models available on GitHub have been published with benchmarks on the hERG dataset, though we recommend treating them as screening-level tools, not definitive assessments.

CYP450 Inhibition: The Metabolism Flag That Compounds Attrition

The cytochrome P450 enzyme family handles a large fraction of phase I drug metabolism. CYP3A4, CYP2D6, and CYP2C9 together account for metabolizing roughly 70 percent of marketed drugs. A compound that potently inhibits CYP3A4 will cause drug-drug interactions. A compound that is itself rapidly metabolized by CYP3A4 has poor oral bioavailability. Both failure modes are common. Both are partially predictable in silico.

CYP inhibition prediction models classify compounds as likely inhibitors or non-inhibitors for each major isoform. The better implementations, including those in ADMET Predictor and Schrödinger's QikProp, include separate models for each isoform rather than a generic CYP flag. That level of granularity matters. CYP2C9 inhibitors create warfarin interaction risks. CYP2D6 inhibitors affect codeine metabolism and dozens of psychiatric medications. The relevant risk depends on the indication and the patient population.

Fact: in the ChEMBL dataset, roughly 15 to 20 percent of compounds with drug-like properties flag as probable CYP3A4 inhibitors at concentrations relevant to therapeutic dosing. That fraction is high enough that running CYP inhibition predictions on every compound in a short-list is not paranoia; it's basic due diligence.

Aqueous Solubility and PAMPA/Caco-2 Surrogates

Aqueous solubility sets a practical ceiling on bioavailability regardless of permeability. A compound with excellent predicted permeability but solubility below 10 µg/mL in aqueous buffer will not achieve therapeutic exposure via oral dosing without formulation work. That formulation work is not a given at the hit identification stage.

Predicted solubility from descriptors, the approach used in SwissADME and QikProp, correlates reasonably well with experimental intrinsic solubility for drug-like compounds but is less accurate for compounds near the edge of the applicability domain of the training set. Kinetic solubility and thermodynamic solubility are different measurements; predicted models generally target thermodynamic solubility. Teams should be aware of which they're comparing against when correlating predictions to assay results.

PAMPA (parallel artificial membrane permeability assay) and Caco-2 cell line data both measure transcellular passive permeability, which is the main permeability route for small molecules without active transporters. In silico surrogates for both exist as part of the QikProp and ADMET Predictor toolkits. Caco-2 predictions are less reliable than PAMPA surrogates for most compound classes, largely because Caco-2 permeability integrates both passive diffusion and active transport, which is harder to model purely from structure. Use predicted Caco-2 values as a triage signal, not a quantitative prediction.

Tool Choices: When to Pay and When to Use Open Source

The honest answer is that the commercial tools are better, and for teams running 2 to 10 screening campaigns per year, the cost is usually justified. Here's how we think about the landscape:

Tool	Strengths	Gaps	Best for
QikProp (Schrödinger)	Fast, integrated with Glide workflows, covers hERG, CYP, Caco-2, solubility, logP/logD	License cost; limited customization	Teams already on Schrödinger Suite
ADMET Predictor (Simulations Plus)	Best-in-class model depth; separate CYP isoform models; hERG with inhibitor classification	Steeper learning curve; expensive at enterprise tier	CROs running high-volume ADMET triage
SwissADME	Free; web-based; fast for small sets; good druglikeness filters	No hERG; no CYP isoform detail; batch limits	Fast sanity checks; resource-limited teams
pkCSM / ADMETlab 2.0	Free; web APIs available; covers reasonable breadth	Accuracy lags commercial tools; no local deployment	Academic settings; preliminary screens
RDKit + open models	Fully scriptable; integrates into any pipeline; free	Requires assembly; model quality varies by property	Teams with cheminformatics engineers in-house

Our default recommendation for teams running fewer than 50 compounds per campaign is SwissADME for a first-pass druglikeness check, followed by QikProp or ADMET Predictor for hERG and CYP isoform flags before final short-list selection. For larger libraries going through automated pipelines, ADMET Predictor's batch processing and programmatic API make it the most practical choice.

Translating Predictions Into Short-List Decisions

ADMET predictions are probabilistic, not deterministic. Every practicing computational chemist knows this. The question is how to translate a collection of probability scores into a synthesis decision when you have 30 candidates and budget for 8 synthesis slots.

Here's the thing: the goal isn't to eliminate all ADMET risk. It's to eliminate the easily avoidable ADMET failures. A compound flagged for likely hERG inhibition, predicted CYP3A4 inhibition, and solubility below 5 µg/mL has three compounding red flags. Route that one to synthesis last, if at all. A compound with one soft flag, strong binding scores, and a clean scaffold is a better use of an early synthesis slot.

In practice, we score each short-listed compound on four to six ADMET properties and compute a simple flag count. Compounds with zero flags go to the front of the synthesis queue. Those with one or two flags are reviewed by the medicinal chemist. Those with three or more are deprioritized unless the binding score is exceptional and there's a clear structural path to improving the flagged property.

Applied consistently, this approach reduces the fraction of synthesis batches yielding no useful binding data. In our data, teams that run ADMET pre-filtering at the short-list stage before synthesis decisions see a meaningful reduction in wasted synthesis slots compared to teams that do hit confirmation first and ADMET later. The exact numbers vary by target class and library, but the direction is consistent across every campaign type we've run. Early filtering does not eliminate late-stage failures. It does make late-stage failures less common and, more importantly, less surprising.