After leading the biotechnology practice at Andreessen Horowitz for more than a decade, growing it to manage nearly $4 billion, Vijay Pande left in June of last year to establish the smaller investment firm VZVC with Zach Werner. The new entity relies on a limited number of focused bets each year, while using artificial intelligence and software agents to reduce the need for a traditional investment team.
But the change in the investment model is not the only focus of Pande’s vision. He believes that artificial intelligence is gradually moving biology from a field dependent on scattered discoveries and luck toward one that can be designed and engineered to a greater extent. This includes identifying drug targets, designing drugs, and improving clinical-trial stages, which are among the most expensive parts of drug development.
From Discovery to Design
According to Pande, machine-learning technologies have enabled computers to handle biological complexity that was difficult to analyze using traditional methods. Over the past decade, parallel advances have accumulated in artificial intelligence for biology and chemistry: the former focuses on understanding disease and ways to treat it, while the latter focuses on creating compounds that target specific proteins.
These developments do not mean that drug development has become cheap or guaranteed. Pande says that the time and cost required to reach clinical trials have begun to decline, but conducting a trial can still require hundreds of millions of dollars. He also notes that the probability of a drug successfully progressing from the first trial to the end of the third trial is about 20%, meaning that the failure of eight out of every ten drugs raises the calculated cost per successful drug.
Pande questions whether the use of synthetic data has actually reduced the cost of clinical trials to the extent commonly claimed; he describes that as more of an ambition than an achieved reality. He also links a significant share of drug failures to the limited ability of animal models, such as mice, to predict what will happen in humans. In his view, an artificial-intelligence model will not be perfect, but in some cases it may surpass the predictive ability of animal models.
Precision Medicine Needs Data About the Individual
The vision extends beyond drug discovery toward selecting the most suitable treatment for each patient. Pande explains that medical practice relies heavily on comparing a patient’s test results with population averages, while the more important question may be how unusual the result is for that specific person.
He believes that relying on the genome alone is insufficient; the genome resembles a blueprint of a house when it is being built, but it does not necessarily describe its condition years later. Therefore, other measurements, such as proteomics, are becoming important for understanding the body’s current state. Automation and robotic measurements also provide new quantities of data that can be linked to artificial-intelligence models.
The Problem of Closed Data
Biology differs from text available on the internet in one fundamental respect: companies cannot simply collect their biological data from the web and train models on it. As a result, each company tends to build its own private dataset, limiting the ability to train shared models or easily transfer knowledge from one model to another.
Pande believes that developing broad atlases of biological information, often in the form of foundation models, could change this trajectory. He expects open-source biological models to play a role similar to that played by open large language models in competing with commercial models. But this proposal remains tied to the willingness of companies and researchers to share data of commercial and scientific value.
What Is Changing in Practice?
VZVC focuses on two main areas mentioned by Pande: artificial intelligence for delivering health care, and artificial intelligence for clinical trials. He says the fund plans about five investments per year, not thirty investments, with him and Werner directly participating in the selected companies. It is also looking for founders with integrity who think in terms of a long-term relationship, rather than short-term competition to win a hot funding round.
The most important takeaway from Pande’s remarks is that the main obstacle facing medical artificial intelligence is not the algorithm alone. Even as models improve, data quality, availability, and shareability remain decisive factors. Technology’s success in the laboratory also does not guarantee commercial success; Pande acknowledges that reaching the market may be more difficult than building the technology itself.
Thus, the discussion presents an optimistic but conditional vision: artificial intelligence may improve the selection and design of drug targets and guide treatment, but it cannot create nonexistent biological data or eliminate the risks of clinical trials. Open questions about data sharing, validating models in humans, and turning scientific results into viable products will continue to determine whether the transition from “discovery” to “engineering” becomes a broad practical transformation.