Why we flag withdrawn drugs, whatever the AI says

The first time we ran a full AI brief for ALS on production, the top-ranked candidate was AMX0035, the sodium phenylbutyrate and taurursodiol combination sold as Relyvrio. On paper that looks reasonable: it was approved for ALS in 2022. But in 2024 its Phase 3 trial, PHOENIX, failed to show benefit, and the company withdrew it from the market.

For a research tool, presenting a withdrawn drug as the best lead is a credibility failure. So we went looking for why it happened.

Three sources, three blind spots

• The model's memory. Language models learn from a snapshot of the world. Anything that happened after that snapshot, like a withdrawal, simply isn't there.
• The trial registry. ClinicalTrials.gov still listed the PHOENIX trial (NCT05021536) with an unknown status when we checked, not as completed with a negative result.
• The compound database. ChEMBL, which records many withdrawals, didn't mark this one as withdrawn in the release we query.

Each source was a little out of date, and the AI had no way to know.

A check the model can't talk its way around

We now run a rules-based safety check after the AI step, on every candidate, every time:

1. We check a curated list of withdrawn and discontinued drugs, each with its source, for cases like this one that the databases miss.
2. We look up each candidate's withdrawal flag and reason in ChEMBL. Withdrawn drugs are marked as withdrawn and moved to the bottom of the list.
3. We match each candidate against the trials we retrieved. A trial that was terminated, withdrawn or suspended adds a visible flag with the registry's stated reason, and the candidate's evidence grade is capped.
4. If every matching trial has an unknown status, we flag that too, because an outcome that was never confirmed isn't support.

Drug names are matched as whole words, so a trial of esketamine doesn't flag ketamine. After the check, AMX0035 carries a warning and no longer ranks first.

What's still missing

A curated list only knows what we've added to it, so we keep extending it and are working on checking the literature for withdrawal notices automatically. The flags tell you what the data says, and our limitations tell you what it might not.