Two drugs abandoned after large clinical trials are getting a second look — regulators in the US and Canada just approved new study designs after an AI re-analysis found the original trials measured the wrong endpoints entirely.
In plain English: both drugs were tested as pain drugs, and the trials measured whether patients reported fewer pain crises. But the molecules appear to be doing something different at the biological level — the new designs test for that actual mechanism instead.
The method is called text embedding, and it works by converting drug descriptions into numerical coordinates that can be plotted against disease gene profiles to surface patterns humans miss. Biossil applied it to the literature on senicapoc, a J&J molecule that failed a Phase III trial in 2015. The original study measured whether patients reported fewer pain crises. A 2021 re-analysis of the same trial data using text embeddings found something the researchers had not looked for: 34% of patients on senicapoc showed a hemoglobin response versus 6% on placebo, per a re-analysis published in the British Journal of Haematology. Health Canada accepted the argument and approved a new trial design in 2025, according to the Globe and Mail.
The rivipansel story is similar. Pfizer's RESET trial — large, rigorous, terminated in 2019 — measured time-to-readiness-for-discharge as its primary endpoint. Pinpointing when a pain crisis actually begins is clinically difficult, and the trial allowed late enrollment to account for this. The drug failed, per Pfizer's Phase III press release. Biossil's re-analysis, applying the same text-embedding method to the clinical literature, suggested the drug works within a narrower window: 24 hours of pain onset. Pfizer never tested that window. The FDA approved a confirmatory trial design with a tighter enrollment criterion, according to the Globe and Mail.
Both cases are endpoint failures, not drug inefficacy. The molecules were doing something; the trials simply were not measuring the right thing.
The financial and regulatory details — the $43M Series B co-led by Founders Fund and OpenAI, the valuation above $100M, the FDA and Health Canada approvals — trace to reporting by the Globe and Mail. None has been independently corroborated. The text-embedding method itself has no peer-reviewed validation outside the case studies in this article. The approach is proof-of-concept in one therapeutic area.
The competitive landscape: Isomorphic Labs, which uses protein structure prediction to design drugs from scratch rather than find new uses for existing ones, raised $600M in 2025, led by Thrive Capital with GV and Alphabet participation. BioXcel's NovareAI uses knowledge graphs. Ignota's SAFEPATH combines cheminformatics and bioinformatics to identify toxicity problems. Biossil's text-embedding approach to failed drugs appears distinct from all three.
Biossil was founded in 2023 by Anthony Mouchantaf, a lawyer who ran RBC's venture strategy arm, and Dr. Alexander Mosa, an MD-PhD. Seed funding of $3.7M came from Staircase Ventures, Golden Ventures, and Panache Medical Ventures. A $22M Series A was led by Founders Fund in 2024. The Series B co-led by Founders Fund and OpenAI in fall 2025 lifted the valuation above $100M per the Globe and Mail — that figure has not been independently confirmed. Dr. Kevin Kuo, a sickle cell specialist at Toronto General Hospital and co-lead of the Red Blood Cell Clinic, joined Biossil as head of medical last year. Mouchantaf estimates roughly $1B was spent developing these molecules before they were set aside.
Three molecules are in the pipeline, with conditional approval filings underway. The senicapoc readout is expected in 2026.
What to watch next is the same as what made the drugs interesting in the first place: whether the endpoint switches hold in Phase III. If they do, the framework extends to every failed trial in the literature — a large and mostly untapped archive. If they do not, the embedding method is wrong for this context, and the molecules go back on the shelf.
