ADMET Hit Triage

Skill workflow-skill many-tools-to-one-skill Proprietary (Paramus) — guidance only; execution performed by Paramus tools under their own license gating.
provenanceai generated statuspartial
Triage a post-assay hit list - dozens of SMILES with measured activity already in hand - into keep / review / drop tiers by joint activity-plus-ADMET ranking, using descriptor and rule-based tools today and the 31-endpoint admet-ai ML panel when it lands.When to use: a screen or assay has produced active compounds and the question is “which hits are worth advancing?”.Multi-tool WORKFLOW with activity-ADMET trade-off judgment.
Procedure8 steps · 7 orchestrated tools
1
Discover tools via search/get_schema ("molecular descriptors", "lipinski",
2
Parse the hit list; pair every SMILES with its measured activity. Quarantine
3
Compute descriptors in bulk with calculate_properties_batch (one call for the
Toolscalculate_properties_batch
4
Apply rule-based gates: reactions_check_lipinski and drug_likeness_profile
Toolsreactions_check_lipinskidrug_likeness_profile
5
Census structural liabilities with count_functional_groups - reactive or
Toolscount_functional_groups
6
Run the ML ADMET panel - in implementation (admet-ai): the 31-endpoint model
7
Check redundancy with calculate_similarity: cluster near-identical hits and
Toolscalculate_similarity
8
Rank jointly: cross the activity tier with the ADMET tier into keep / review /
You get
keep / review / drop tier per compound with reasons; the descriptor and

Overview

After a screen, the question changes from "which molecules are generally clean" to "which of my actives are worth advancing". This skill triages a hit list - dozens of SMILES with measured activity in hand - into keep / review / drop by ranking activity and ADMET jointly: batch descriptors and rule flags today, the 31-endpoint admet-ai ML panel when it lands. The value is the joint ranking and the explicit reasons per compound, not any single descriptor.

Guidance vs execution: every descriptor, flag, and similarity number comes from a tested Paramus tool. Do not hand-roll RDKit calls or a scoring model in ad-hoc code - route to the tools.

When to Use

  • A screen or assay has produced a hit list with measured activity data and the question is "which hits do we advance?".
  • Hit-list pruning before hit confirmation re-tests or hit-to-lead.

Do not use for: liability triage of a discovery library before any assay data exists - that is the published sibling paramus-admet-liability-triage, which ranks a library on ADMET/developability alone with an overlapping tool set. The overlap is real and borderline: both skills reach for calculate_logp, drug_likeness_profile, reactions_check_lipinski, and cloud_rowan_descriptors. The triggers are disjoint on one fact - this skill starts only when measured activity is in hand and performs a joint activity-ADMET ranking; the sibling ranks without activity data. If your question has no activity column, use the sibling. Also do not use for final regulatory ADMET reporting.

Domain Judgment

  • Flags are gates, predictions are signals: a hard structural flag (reactive group) drops regardless of activity; a soft descriptor signal routes a potent compound to review, not drop.
  • Activity is a column, not a filter: the whole point of post-assay triage is that a marginal ADMET profile on a highly potent hit is a medicinal-chemistry problem to fix, not a reason to discard.
  • Near-duplicate hits carry no extra information: redundancy demotion keeps the keep-list chemically diverse, which matters more than list length at this stage.
  • When the admet-ai panel is absent, triage on descriptors and rules and mark the ML dimension open - do not substitute general knowledge for the missing endpoints.

Failure modes

  • A hit list arrives without activity data -> redirect to paramus-admet-liability-triage; joint ranking is impossible and pretending otherwise produces the sibling's output under the wrong name.
  • Parse failures or missing activity values -> quarantine and report; never silently drop a hit.
  • A descriptor or rule tool is unavailable -> triage on the remaining signals and mark the missing dimension; no surrogate models, no hand-rolled calculations.
  • The admet-ai app absent -> descriptor/rule triage stands, the ML dimension is marked in implementation (admet-ai), and borderline keeps are labelled as decided on partial evidence.

Example

Asked: "Our biochemical screen returned 48 confirmed actives (SMILES plus IC50 per compound). Which do we advance?" The skill batches descriptors across the 48 SMILES with calculate_properties_batch, adds calculate_logp, applies reactions_check_lipinski and drug_likeness_profile, and censuses functional groups with count_functional_groups. cloud_rowan_descriptors fills the electronic and ADMET-relevant descriptor gaps; the admet-ai panel step is marked in implementation (admet-ai) and skipped today. calculate_similarity collapses three clusters of near-duplicates to their most potent representatives. The joint ranking crosses the activity tier with the ADMET tier: compounds potent and clean go to keep, potent but flagged go to review with the flag named, weak and flagged go to drop. Every tier carries its reason and every number comes from the tool calls at run time.

Input

hit list (SMILES) with measured activity per compound (assay, readout,

Output

keep / review / drop tier per compound with reasons; the descriptor and