ADMET-First Filtering: Killing Liabilities Early
The majority of drug candidates that fail in clinical development fail for reasons that were present — and in some cases predictable — at the lead identification stage. ADMET liabilities (absorption, distribution, metabolism, excretion, toxicity) are the most common driver of late-stage attrition. The principle of ADMET-first filtering is simple: find these liabilities as early as possible, when the cost of addressing them is measured in chemistry cycles, not clinical dollars.
The cost of late discovery
When a candidate fails Phase 2 for ADMET reasons — metabolic instability leading to variable exposure, off-target cardiac liability driving hERG concerns, hepatotoxicity signal — the cost is measured in tens of millions of dollars and years of time. The same liability discovered during lead optimization costs one synthesis cycle and a few weeks.
This asymmetry is well-understood. The problem is that traditional drug discovery workflows still apply ADMET analysis after potency optimization. The reasoning: if the compound isn't potent, ADMET doesn't matter. This logic is correct but the conclusion — wait for potency optimization to apply ADMET analysis — generates the late-attrition problem.
ADMET-first in practice
ADMET-first filtering inverts the priority order:
-
Screen for acute ADMET liabilities first. Before investing in potency optimization of a compound, assess predicted hERG binding, CYP inhibition, reactive metabolite formation, and Ames mutagenicity. Compounds with clear acute liabilities are deprioritized regardless of potency.
-
Apply computational ADMET filters to all synthesis candidates. Before any compound is synthesized, AI predicts its key ADMET properties. Compounds with predicted liabilities that can't be structurally addressed are removed from the synthesis queue.
-
Optimize for ADMET and potency simultaneously. During lead optimization, SAR is driven by both target activity and ADMET goals. AI tracks both dimensions simultaneously, flagging when a structural modification improves one at the cost of the other.
-
Flag the liabilities that remain. Not all ADMET liabilities are disqualifying. Some are manageable (formulation, dosing regimen, monitoring). The goal is transparency: enter lead optimization knowing which ADMET challenges remain and whether they have credible solutions.
The ChEMBL advantage
The ChEMBL database contains ADMET data for over 2 million compounds. AI models trained on this data provide accurate ADMET predictions — not perfect, but accurate enough to make meaningful early triage decisions. Prediction accuracy is high enough that routing synthesis resources based on AI ADMET predictions systematically improves lead quality.