Axiom is a seed-stage biotech and AI startup based in San Francisco with a direct pitch: predict a drug’s toxicity before you test it. The company has raised roughly $15M so far, in seed rounds of $7M in 2023 and $8M in late 2024. It sits at the intersection of AI, drug discovery and predictive toxicology, and it is aiming at one of pharma’s biggest and most expensive problems: preclinical safety.
Axiom says its main goal is to replace slow, expensive and outdated toxicity testing, especially animal tests, with in-silico AI models that are more accurate, more scalable and more relevant to humans.
Roughly 90% of drugs fail, and 40-55% of those failures are linked to toxicity, especially liver toxicity. Axiom claims that widespread methods such as animal testing predict human liver injury poorly, and that they are slow (weeks to months) and expensive (thousands to millions per program). The US Congress and the FDA are also beginning to phase out the legal requirement for animal testing, and have explicitly called out AI-based computational toxicity models as an alternative. That gives Axiom an opportunity to be at the forefront of the transition.
The dataset
Axiom’s biggest edge is its dataset, which it describes as the “atlas of human cell toxicity”. By the company’s own figures, it currently includes:
- 115,000+ unique small molecules
- 394 million+ individually labeled cells
- 9 billion+ labeled mitochondria
- 5,200+ clinical DILI labels
- 38,400+ liver enzyme elevation datapoints
- a public dataset (June 2025) of 130,000+ compounds
Axiom runs high-content imaging, or “cell painting”-style assays, to capture morphological and organelle-level features. Those features are used to train ML/AI models on its liver dataset. The models predict clinical liver injury, both as a classification (toxic vs non-toxic) and by severity. They also cover dose-response and exposure, meaning how a molecule’s toxicity changes with concentration or exposure, and phenotypic profiles, meaning how each molecule affects cell morphology.
Inside the toxicology engine
The process starts with a small-molecule structure, supplied as a SMILES or structure file. Axiom’s internal wet lab screens the molecule in primary human hepatocytes, and high-content imaging captures cell and organelle morphology. Feature extraction converts those images into high-dimensional phenotypic profiles, which are then linked to outcomes: clinical DILI labels, liver enzyme elevations and exposure data. ML/AI models are trained on this atlas of human liver toxicity data and deployed in Axiom’s app and API.
When a user uploads a new molecule, the model predicts clinical liver injury risk (DILI), dose-response and exposure (AUC, Cmax). It also identifies toxic phenotypic signatures and structural motifs (toxicophores).
Axiom currently has a live demo available. Researchers can upload a small molecule and see its predicted liver risk, exposure profile and phenotypic effects. They can also compare molecules, explore the dataset and generate explanations.
Outlook
Many other biotech companies are racing to complete a reliable AI toxicology system. So far Axiom has an impressive primary human hepatocyte dataset, which may give it an edge if it continues on a positive trajectory.
There are caveats. Although the FDA is vouching for software like this, it may be slow to accept AI toxicity models as a primary gatekeeper until more solid results are proven. Larger companies such as Eli Lilly and Novo Nordisk already have tons of proprietary data and could train similar models of their own if needed. And producing novel data for every drug toxicity scenario requires expensive new lab infrastructure and technological equipment for AI training and usage, which in turn means more funding and more time before Axiom can develop further.
For now, Axiom’s biggest advantage is the dataset. If the company keeps scaling it, it may build an AI tool no one else can match. Further updates to its AI datasets, or partnerships that expand its data, are the things to watch.