Industries / Health / Pharmaceuticals
Preclinical testing
Testing a candidate in cells, tissues and animals to judge whether it is safe enough to try in people, and at what dose.
7
Milestones achieved, from 1959 to 2025.
1
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1959
Replace, reduce, refine
Russell and Burch, scholars of the Universities Federation for Animal Welfare (UFAW), published The Principles of Humane Experimental Technique. It set out the Three Rs: replacement, reduction and refinement of animal use. The UK's National Centre for the 3Rs says the principles are now embedded in national and international law on animal research.
Credit
Russell and Burch, for UFAW
2006
Six men in intensive care
In March 2006 a first trial in people of TGN1412, an antibody meant to treat autoimmune diseases, put six men in intensive care at Northwick Park Hospital in London. Tests in monkeys had shown no such reaction, though monkeys carry an almost identical receptor for the antibody. A UK expert panel made 22 recommendations that December, including dosing volunteers one at a time, consulting outside experts on high-risk drugs and building a database of animal results.
LessonPassing animal tests does not prove a first human dose is safe, above all for drugs that act on the immune system.
Credit
the UK expert scientific group that reviewed the trial
2010
A lung on a chip
D. Huh, D. E. Ingber and colleagues built a small device that copies the boundary between the air sacs and the blood vessels of the human lung. It stretched like a breathing lung, and that motion increased the toxic and inflammatory responses to silica nanoparticles. Devices of this kind are called organ-chips.
Credit
D. Huh, D. E. Ingber and colleagues
2012
Few landmark results repeated
Scientists at the biotech company Amgen tried to confirm 53 "landmark" preclinical cancer papers before building on them. The findings held up in 6 cases, or 11%. A Bayer team had reported the year before that about 25% of published preclinical studies could be validated.
LessonA published result is not a validated drug target until someone repeats it.
Credit
C. G. Begley (Amgen) and L. M. Ellis
2022
Animal-test wording dropped
The FDA Modernization Act 2.0, passed in December 2022, replaced US legal wording that pointed drug makers to tests on animals with the broader term "nonclinical tests", and defined that term to include human-relevant methods such as cell-based and computer models.
Credit
US Congress
2023
530 drugs, one database
M. Evangelisti and colleagues released a database of animal-study results for 530 approved drugs, taken from FDA pharmacology reviews, with effects in people taken from the drugs' labels. It links each observed or not-observed effect to its dose level, so animal and human findings can be compared.
Credit
M. Evangelisti and colleagues
2025
FDA plans fewer animal tests
On 10 April 2025 the FDA announced a plan to reduce, refine or replace animal testing for monoclonal antibodies (lab-made antibodies used as drugs) and other drugs. It named computer models, human cell lines, organoids and organ-chips, which it groups as new approach methodologies (NAMs), and said implementation would begin immediately. Its roadmap states that "no comprehensive database containing animal and human toxicity data currently exists".
Credit
FDA
Watch for
An organ-chip nears qualification
Emulate submitted its full qualification package for its Human Liver-Chip to the FDA's ISTAND programme. Once the FDA qualifies a tool for a stated use, any company can use it with confidence that the FDA will accept the data. The stated use here is judging drug-induced liver injury. Emulate reports reproducible results across three laboratories and three donors of liver cells. An earlier study of 870 chips reported that they flagged 87% of the liver-toxic drugs and wrongly flagged none of the safe ones, in a blinded set of 27 drugs.
Achieved whenthe FDA qualifies the chip for that use.
Claimed by
Emulate, September 2026
Open
01
Beat the Liver-Chip on paper
A computer model, trained only on public data, that sorts the same 27 benchmark drugs used in the Liver-Chip study into liver-toxic and safe, scored the same way the study scored the chip.
The briefCredit · unclaimed02
Grow the toxicity database
A pipeline that reads FDA pharmacology and toxicology reviews and drug labels, and fills the same fields as the 530-drug database: effect, species, dose level, severity and whether the effect went away. Use it to add the drugs approved since.
The briefCredit · unclaimed
Open 01 · the brief
Beat the Liver-Chip on paper
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A computer model, trained only on public data, that sorts the same 27 benchmark drugs used in the Liver-Chip study into liver-toxic and safe, scored the same way the study scored the chip.
Why it is still open
The Liver-Chip's 87% sensitivity and 100% specificity on 27 blinded drugs is the result the FDA cites in its roadmap for cutting animal tests. I found no computer-only model scored on the same 27 drugs.
Built with what exists
the FDA's DILIrank 2.0 dataset, which sorts 1,336 approved drugs into four classes of liver-injury concern; RDKit for chemical features; scikit-learn for the model; the list of 27 drugs from the Liver-Chip paper.
First step
Copy the 27 drugs and their labels from the Liver-Chip paper and remove them from a copy of DILIrank 2.0.
Done looks like
The model is frozen and published before it is scored, with the 27 drugs and their close chemical relatives removed from training. Its sensitivity and specificity on the 27 are reported beside the chip's, and an outside person reruns the code and gets the same numbers.
Open 02 · the brief
Grow the toxicity database
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A pipeline that reads FDA pharmacology and toxicology reviews and drug labels, and fills the same fields as the 530-drug database: effect, species, dose level, severity and whether the effect went away. Use it to add the drugs approved since.
Why it is still open
The FDA's April 2025 roadmap says no comprehensive database of animal and human toxicity data exists, and lists building an open one as a step it plans to take. The 2023 database stops at 530 drugs.
Built with what exists
FDA review documents from Drugs@FDA; drug labels from DailyMed; a language model that reads scanned PDFs page by page and cites the page for each value, such as Claude or Gemini; the 530-drug database as the answer key. The note "Seven more flagged" covers a Claude-based reader of scanned drug-making records.
First step
Pick 10 drugs from the 530-drug set, run the pipeline on their FDA reviews and compare the results field by field.
Done looks like
On 50 drugs held back from the 530-drug set, at least 90% of extracted effect-and-dose entries match the curated values. Then 200 newer drugs are added, with a person checking 1 in 10 entries.