Traditional cheminformatics tooling answers one narrow question at a time. You query a compound database, get raw rows back, and then spend hours assembling the actual answer: Is this molecule soluble? Does it cross the blood-brain barrier? Which off-targets should worry us? The retrieval was never the hard part — the synthesis is.
Why retrieval is the wrong unit of work
A search box optimises for recall. A discovery team optimises for decisions. Between those two things sits an enormous amount of undifferentiated labour: exporting hits, normalising identifiers, running descriptor calculators, pulling percentile baselines, plotting permeation, formatting a slide.
That work is deterministic, repetitive and expensive — which is precisely the profile of work an agent should own. The scientist's judgement belongs on the interpretation, not on the tool-hopping that precedes it.
From retrieval to reasoning
The Cyberia Agentic Platform treats molecule search as an agent task, not a lookup. You send a SMILES string — or simply name a compound in plain language — and the BioMed agent orchestrates the full analysis pipeline:
- ADME profiling — water solubility, pharmacokinetics and drug-likeness rules scored in one pass
- Toxicity forecasting — ADMET-AI endpoints benchmarked against DrugBank percentiles, so "high risk" is quantified, not vibes
- BOILED-Egg permeation — intestinal absorption and blood-brain barrier behaviour on a single interpretable plot
- Target and category breakdown — absorption, distribution, metabolism, excretion and toxicity organised as structured data
- Interactive 3D conformer — generated in-line and downloadable directly from the report
What the agent actually does per run
- Resolve the input — plain name, SMILES, InChI or an internal compound ID — to a canonical structure.
- Plan the toolchain against the question asked; a permeability question and a tox-triage question do not need the same passes.
- Execute descriptor, ADMET and conformer tools in parallel, retrying and substituting when a model is unavailable.
- Benchmark each endpoint against reference distributions so every number carries a percentile, not just a value.
- Compose a typed report with charts, tables and the 3D viewer already rendered.
Structured output, not raw text
Every run returns typed, schema-validated output that renders directly in the workspace: charts, distribution tables, radar plots and 3D viewers. There is no copy-paste from a chat window into a slide deck — the report is the deliverable.
Schema validation matters more than it sounds. Because the output is typed, a downstream agent can consume it: a triage agent can rank fifty candidates by predicted hERG risk, a reporting agent can assemble the weekly portfolio review, and an alerting agent can flag any new compound that breaches a solubility threshold. Free text cannot be composed; typed reports can.
Instructions, skills and memory
Agents are configured, not prompted ad hoc. Each agent carries:
- Instructions — the operating contract: scope, tone, refusal rules, how to present uncertainty.
- Skills — reusable markdown-defined procedures, such as how to format image outputs or how to structure an ADMET section.
- Memory — prior runs, project context and the team's conventions, so the tenth report matches the first.
Because agents are wired to data products and ML feature pipelines as first-class tools, the same pattern extends to any domain: swap the molecule engine for your own models and the agent becomes an analyst for your data, not a search box.
Traceability and governance
Every figure in the report names the model that produced it, the version, and the reference set used for percentile ranking. Runs are logged with their inputs, tool calls and outputs, so a result can be reproduced or audited months later — a requirement, not a nicety, when the output feeds a regulatory dossier or an investment decision.
What used to take a computational chemist hours of tool-hopping now arrives as one report, in minutes, with every number traceable to the model that produced it.
Why it matters
Drug discovery teams drown in candidate compounds. When every search returns a complete, comparable, decision-ready profile, triage stops being a bottleneck and becomes a filter you can run at portfolio scale.
The change is not that the models are new — most of these endpoints exist in public tooling. The change is that an agent runs all of them, every time, in a consistent shape, so comparisons across compounds are actually valid and the scientist spends their time deciding rather than assembling.
Getting started
Point an agent at your compound source, attach the model endpoints you trust, and define the report shape once. From then on, every molecule your team looks at — from a single hypothesis to a full library sweep — returns the same decision-ready profile.