Drug development faces a significant hurdle as nearly nine out of ten clinical programs fail. Many of these setbacks occur during efficacy testing, often because the chosen biological mechanism fails to drive the disease in question. Industry leaders have long known that targets supported by human genetics are two to four times more likely to succeed in clinical trials. However, the application of this knowledge remains limited, with only 3.6% of genetically supported targets ever pursued for the specific indications their genetics would suggest.
Addressing the Limitations of Current Genetic Research
The current challenge lies in the ambiguity of human genetic data. Mapping a disease frequently reveals dozens or even hundreds of associated genes, yet most display only muted effects. This occurs because natural selection often filters out genetic variants that exert a massive impact on human health. Daphne Koller, founder and CEO of insitro, describes this as a foggy window. Researchers struggle to distinguish the causal drivers from the noise in these large-scale genomic datasets.
To overcome this, insitro has developed specialized machine learning models that transform high-content measurements into precise, actionable data. These models integrate diverse inputs, including whole-body imaging from large human cohorts and genome-scale perturbation screens in human cells. By creating a library of what Koller terms precision phenotypes, the company aims to sharpen the view of human genetics. In studies of metabolic dysfunction-associated steatohepatitis, or MASH, these models surfaced more than 30 times the number of genetic associations compared to traditional clinical staging methods.
The Role of Causal AI and the Virtual Human
At the center of this strategy is a causal artificial intelligence model, which the company calls the Virtual Human. This model serves as an integration layer, drawing evidence from different data modalities, disease biologies, and physical scales to build higher conviction in target selection. The hypothesis is that this integration will lead to higher success rates in the clinic. To test this, the team ran a retrospective analysis using historical phase 2 trial data. Without exposing the model to prior trial outcomes, the system successfully identified target-indication pairs with lower failure rates.
In the model's top decile of predictions, the false-positive rate for metabolic disease reached 10%, while cardiac conditions reached 21%. These figures contrast with the historical failure rate of 57% for similar trials. While a retrospective benchmark does not guarantee future results, it provides a measurable baseline for how AI might filter out doomed drug programs before they reach patients. Koller maintains that the value of this approach is two-fold, offering both economic savings and a path to therapeutic options for previously untreatable conditions.
Translating Data into Physical Experiments
The utility of these models extends beyond target identification. Because the defined phenotypes span from the patient level down to individual cells, a genetic hit becomes an experiment researchers can execute in the lab. The platform guides scientists on where a gene acts, which pathways it disrupts, and whether a drug should activate or inhibit that specific gene. This cycle creates a direct link between the discovery of a mechanism and the development of the drug itself.
Assays used to validate a target now serve as the foundation for optimizing molecules. This consistent thread of data flows from initial biobank samples through to the final clinical candidate, as demonstrated in the company's work on MASH. By generating causally credentialed targets, the goal is to shift the conversation from individual program success to the volume of diseases that can be addressed at once. This methodology represents a wager that a systematic, data-driven approach to human biology will finally break the cycle of failure that has plagued drug development for decades.