insitro

Machine learning is the mechanism of the company. insitro combines more than 20 petabytes of automated cellular experiments with population scale genetics, and its two named engines are models: Virtual Human, a causal AI engine built on human clinical and cellular data to identify disease targets, and TherML, a modality agnostic design engine for small molecules, oligonucleotides and complex biologics. Its programs and partnerships (Bristol Myers Squibb in ALS, Eli Lilly) come out of those engines.

Models propose and experiments decide. Virtual Human identifies genetic drivers and targets, using reasoning models and agentic orchestration to integrate insights, and each hypothesis is tested in purpose built cellular disease models (induced pluripotent stem cells, genome editing and high content phenotyping) before a program advances, so predictions meet experimental evidence. Partners make their own calls: Bristol Myers Squibb nominated ALS-2 and ALS-3 from targets the Virtual Human identified.

What evidence a target needs before insitro commits a program to it is not published.

insitro's method papers include the POSH self supervised cell painting CRISPR platform (Nature Communications, 2025), Channel Vision Transformers, the Sparse Additive Mechanism Shift variational autoencoder for cellular perturbations, deep learning on iPSC derived neurons, and the KinDEL dataset (ICML 2025). The company describes its training data as more than 20 petabytes of automated cellular experiments plus population scale genetics, and TherML as a physics informed design engine.

What is not published is which models and versions run in Virtual Human and TherML, or how changes are tracked.

Named outside sources feed the platform. Genomics England keeps its genomes and images inside its own research environment, where insitro's search engine runs; UK Biobank imaging and clinical datasets underpin the liver and brown fat work; Lilly's multispecies preclinical data trained the pharmacokinetic models, which run in Lilly TuneLab under a third party host; INSIGHT at Moorfields Eye Hospital supplies eye images; and CombinAbleAI's biologics design engine came in house through the 2026 acquisition.

No subprocessor or cloud provider list is published.

CTRO-1013, the lead program, is a liver targeted siRNA against IRS1, chosen from 480 genetic signals for liver fat. A single dose cut hepatic IRS1 expression by 85% and de novo lipogenesis by 59% in diet induced obese mice, it lowered the fibrosis and liver injury biomarkers TIMP-1 and CK-18 by 37% and 68% in a fibrogenic mouse model, and target reduction held over 90 days in cynomolgus monkeys, data presented at the American Diabetes Association 86th Scientific Sessions in June 2026. A brown adipose tissue study built a machine learning phenotype from UK Biobank MRI in 69,598 people, and knockdown of the prioritized target, BAT-01, cut body weight by 15% over four weeks in obese mice. In ALS, targets from the Virtual Human rescued neurite growth in iPSC derived motor neurons, and the POSH pooled cell painting CRISPR platform was published in Nature Communications in December 2025.

None of these candidates has yet been dosed in people; the liver fat and fibrosis programs are expected to enter the clinic in early 2027.

insitro builds its targets from human cohort data, and its Genomics England work keeps that data in place. Its embedding search engine runs inside the secure Genomics England Research Environment for that organization's research partners, over almost 150,000 whole genomes with phenotypic data from NHS patients with rare diseases and cancer plus histopathology images, under Genomics England's own consent and governance framework for deidentified data; the models went to the data. The liver and brown fat discovery work uses UK Biobank imaging and clinical datasets.

insitro has no published policy of its own on consent, retention, secondary use or model training rights for the human and partner data it holds.

The company develops drugs with pharmaceutical partners rather than serving covered entities, and its human data work draws on research cohorts such as UK Biobank and Genomics England. Its privacy policy applies only to its website.

No statement on protected health information, business associate status or how HIPAA applies to its data is published.

The company holds human genetic and clinical data and partners' confidential programs, and its website privacy policy says it has implemented reasonable precautions to protect the information it collects from loss, misuse and unauthorized access.

No security certification, trust center or security page is published.

The platform is an internal research engine: it identifies targets and designs therapeutics, makes no diagnostic or treatment recommendation, and its candidates follow the drug pathway rather than any device route. The lead program, CTRO-1013, is in IND enabling studies, and Hideo Makimura, who has more than a dozen investigational new drug and clinical trial application filings behind him, became chief medical officer in September 2026 to lead insitro's transition to a clinical stage company.

Among its published methods are tools for checking genetic discovery itself: EmbedGEM, a framework to evaluate how useful learned embeddings are for genetic discovery, and a paper on the pitfalls of genome wide association studies on ratio traits. The models learn from human cohorts such as UK Biobank and from cell lines carrying their donors' genetic backgrounds, so cohort ancestry and composition shape which targets look causal.

No model card, governance framework or analysis of model performance across ancestry groups is published.

Wrong predictions are meant to fail in the lab before they cost a program: targets from the Virtual Human go through cellular disease models and then in vivo studies before programs advance, and partners such as Bristol Myers Squibb choose which targets to take forward. Partner economics are tied to progress through milestone payments.

No accuracy figure for target calls, and no warranty or remedy for a partner when a platform derived target fails, is published.

Data reaches the platform through research environments rather than hospital record systems. The embedding search engine runs inside the Genomics England Research Environment for that organization's research partners, liver and brown fat discovery draws on UK Biobank imaging and clinical datasets, a collaboration with INSIGHT at Moorfields Eye Hospital uses its bioresource of eye images as the basis of a new AI foundation model, and the in vivo pharmacokinetic models built with Lilly are available to certain biotech companies inside Lilly TuneLab.

No connector to a partner's electronic lab notebook, laboratory information system or research data platform is named.

The discovery platform runs in house, and TherML is directly integrated with insitro's automated labs; teams work in the United States, Poland and Malaysia, with a research center in Israel for large molecule design that came with CombinAbleAI. Two arrangements place its models in other organizations' environments: the embedding search engine inside the secure Genomics England Research Environment, where the data stays put, and the pharmacokinetic models inside Lilly TuneLab, which runs on federated learning infrastructure hosted by a third party provider.

Hosting providers and regional data residency for the core platform are not stated.

insitro earns through discovery partnerships while keeping its own programs. Its five year Bristol Myers Squibb collaboration in ALS and frontotemporal dementia carried $50 million in upfront cash with a total potential deal value over $2.1 billion, followed by $25 million in milestone payments for discovery milestones and the first target, and $10 million when Bristol Myers Squibb selected ALS-2 and ALS-3. Lilly is a modality partner on the liver fat and fibrosis programs, with insitro retaining full global rights.

No price or standard terms for a new partner are published.

The pipeline spans four disease areas: liver fat and fibrosis (CTRO-1013, CTRO-1029 and CTRO-1035), obesity (OBS-1 and OBS-2), ALS (an ALS-1 small molecule, CTRO-2018, ALS-2 and ALS-3) and ophthalmology (three dAMD programs). The company chooses diseases where it can mine rich clinical data or build tractable cellular models, and TherML covers small molecules, oligonucleotides and complex biologics including multispecific antibodies and T cell engagers.

The ophthalmology programs and the biologics work have no published results yet.