How a brief is made, and where it can be wrong.

1. Retrieval

For a disease name, the engine queries four sources in parallel, each with its own timeout:

• PubMed through NCBI E-utilities: matching publications, titles and abstracts.
• ClinicalTrials.gov API v2: trials registered for the condition, with phase, status, interventions and stop reasons.
• ChEMBL: drug indications that match the disease, enriched with each molecule's mechanism, target and withdrawal record.
• UniProt: reviewed human proteins annotated with the disease.

Common abbreviations are expanded before searching (ALS becomes amyotrophic lateral sclerosis, for example). A source that fails or times out is reported in the brief as failed, and the analysis continues with the others rather than guessing.

2. Synthesis

The retrieved records are condensed and given to a large language model (currently an OpenAI GPT-4o-class model) with strict instructions: rank candidates by combined evidence strength, only name drugs that appear in the retrieved data, and treat terminated, withdrawn or suspended trials as negative evidence. The model returns structured JSON, which the engine validates before display.

3. Safety check

Language models can be confidently out of date. In our first live run, the model ranked a drug that had been withdrawn after a failed Phase 3 trial as the top ALS candidate. So after the AI step, a rules-based check that doesn't depend on the model runs on every candidate:

• A curated, sourced list of withdrawn and discontinued drugs is checked first, for cases the databases miss. Then ChEMBL's withdrawal flag and reason are looked up. Withdrawn drugs are marked and moved to the bottom.
• Matching trials that were terminated, withdrawn or suspended add a flag with the registry's stated reason, and lower an A grade to B.
• Drugs whose only matching trials have an unknown status are flagged, because the outcome was never confirmed.

Flags explain themselves in plain language and link to the record behind them. Read the story behind the check.

4. Historical approval odds

Each candidate shows how often drugs at its clinical stage have gone on to be approved, using industry-wide rates from BIO, Informa Pharma Intelligence and QLS Advisors' Clinical Development Success Rates 2011–2020:

These are base rates across all diseases. Rates differ a lot by therapeutic area, and a flagged candidate's real odds are likely lower. They are not a probability that a drug cures a disease or works for any one person.

5. Evidence grades

Grades are assigned by the model from the retrieved data and are a triage aid, not a formal systematic-review rating.

6. Grounding check and memory

After synthesis, an automated check scores how well the brief's candidates are grounded in the retrieved records. The brief, its score and any researcher rating are saved, and similar past runs are recalled as hints for future analyses.

Limitations

• AI can be wrong. Language models can misread, overstate or mis-rank evidence. Treat every candidate as a hypothesis and read the cited records.
• Retrieval is partial. Each brief reads a sample of the most relevant records, not the entire literature.
• Sources lag reality. Registries and databases update on their own schedules. A trial's status or a drug's approval may have changed since the source last updated, and some withdrawals are recorded by neither ChEMBL nor the trial registry.
• Name matching is imperfect. Diseases with many synonyms or subtypes can return incomplete or mixed results. Specific names work best.
• Not for clinical use. CITADEL does not provide medical advice and must not be used to make decisions about patient care.

Data sources and attribution

CITADEL is not affiliated with or endorsed by any of these providers.