The Vector Institute hosted a Foundation Models for Science Workshop at MaRS in Toronto, bringing together astrophysicists, biologists, and chemists to address the problem of building AI systems for domains where labeled data is scarce and stakes are high. A recap authored by John Knechtel and published by the institute describes three challenges that recurred across the day’s presentations: the annotation bottleneck, the trust gap, and what it called the lab-to-production leap.
The session’s premise, as described in the recap, is that internet-scale pretraining does not transfer cleanly to scientific domains. Chemical signatures, galaxy images, and clinical measurements cannot be obtained by scraping the web; they are produced through expensive, slow experimental processes.
The annotation bottleneck was illustrated by Nolan Koblischke, a doctoral student in astrophysics at the University of Toronto, who described the challenge of finding gravitational lenses — rare celestial phenomena appearing in roughly one in every ten thousand telescope images. Koblischke’s team fed 140 million galaxy images to GPT-4o-mini to generate textual descriptions, then used contrastive learning to build a shared semantic space between image and text, producing a foundation model called AION-1. The system allows astronomers to search for image types using natural language queries — including query types that were never explicitly trained — a technique the recap calls “zero-shot” search.
The trust gap was addressed by Justin Donnelly of Axiom Bio, who described the problem of predicting drug-induced liver injury — a toxicity outcome for which clear labels exist for approximately 2,000 compounds, according to the source. Donnelly’s team trained a model on around 116,000 compounds tested in laboratory assays to predict a set of biological features from molecular structure. Those predicted features were then fed to a more transparent second model. When a drug called Lexipeptide showed toxicity in development, the system attributed the risk primarily to mitochondrial stress and high drug concentration, the source states. The system also flags uncertainty explicitly when it lacks sufficient data on a mechanism. The recap characterizes this as turning AI from an opaque scoring system into a “decision companion with an audit trail.”
The source excerpt cuts off before the account of the third challenge, the lab-to-production leap. The recap does not include quantitative benchmarks comparing AION-1 or the Axiom Bio model against prior approaches, and the workshop date is not specified in the available text.
The Vector Institute’s science-facing AI programs run alongside its applied and industry-facing work. The Foundation Models for Science Workshop is one in a series of specialist convenings the institute has organized around specific application domains.