A Great Algorithm Needs an Ecosystem

On Wall Street, a trading firm with a better algorithm exploits its edge for fantastic yield in weeks or even days, before the competition identifies the advantage and closes the gap. In Silicon Valley, a startup with an algorithmic innovation has months until an incumbent like Alphabet or Meta sees the opportunity, value that then manifests as an acquisition price.

However, in biotech, timelines of therapeutic development make it very hard for an algorithm to be the source of return. AlphaFold is a good example: DeepMind’s algorithm is Nobel-winning and indisputably useful for drug development. But turning that breakthrough into value required standing up an entire company, raising capital, and entering drug discovery to generate therapeutic assets.

If the algorithm must become its own company to justify the investment, then the bar for building it is enormous. You need the science, the team, the capital, and a defensible market position, all at once.

You could, instead, look to service models to capture value without needing to generate an asset. Schrödinger, for example, is an admirable team leading innovation in physics-based methods for molecular simulation, generating hundreds of millions in annual revenue, at a multiplier that makes it a $1B+ public company. Bootstrapped from a self-funded team, theirs would be a success story of entrepreneurial dreams. However, in the realm of venture creation, these value creation models simply don’t return enough value for the dollars invested. Consider that a single strong oncology asset can anchor several times the market cap of Schrödinger!

The challenge, then, is not simply inventing better algorithms, but finding an organizational model that allows their value to be realized without requiring each breakthrough to justify an entirely new company. The ideal environment is one where algorithmic advances can be deployed repeatedly across many therapeutic programs, generating value through a portfolio rather than a single entity, and where the cost of translating an algorithm into impact is measured in weeks or months instead of years.

Pioneering Intelligence is built to operate in precisely this way. We develop foundational algorithms and immediately apply them across multiple companies and programs, capturing their value without first needing to create a new venture around each innovation.

So the portfolio solves how to capture value from an algorithm. But it raises a sharper question about which algorithms to build in the first place.

When I was a PhD student, prediction of protein structure from primary sequence was the sort of problem you worked on because it was hard, not because you expected to solve it. Then, we made incredible progress in protein prediction, spawning a dozen well-funded efforts almost overnight.

This type of adjacent innovation, meaning the next obvious step from where the field already is, can be extraordinarily valuable. But it is also fiercely competitive, and success often depends on timing and luck as much as on insight.

You can see the same dynamic outside biotech. Uber and Instacart proved that on-demand logistics worked; DoorDash refined food delivery into a category. Then, Uber Eats arrived, winning on the strength of a logistics network and a user base it already had. In a crowded market, the edge often goes to the incumbent advantage, not the innovator.

As a result, Flagship Pioneering invests in non-adjacent innovation. It is not required that we conceive of an idea before anyone else in the world, only that we are willing to invest in reducing the uncertainty around ideas others see as too far-flung to pursue. For example, Cellarity pioneered modeling the behavior of whole cells with AI, an early bet on what the field now calls virtual cells. Generate:Biomedicines pioneered generative models for antibody and protein design when “generative model” was still jargon of academic conferences.

In both cases, the value came from being willing to sit with the uncertainty long enough to resolve it.

Since we build companies in whitespaces where the uncertainty is high and the competition is thin, the tools everyone else relies on don’t necessarily apply. The further our companies get from the mainstream, the more they require new computational solutions based on an entirely different view of biology.

Consider our team’s work led by my colleague Andrew Liu predicting very large protein complexes (NeurIPS Flash IPA paper, github link). The field’s models were built to fold a single protein domain of about a few hundred amino acids. Those limits are enough for many companies to advance the design of binders and optimize antibodies.

Our companies, however, are developing large heteromultimeric complexes and drugging dynamic targets that are empirically unobservable. Our team asked, “What if computational structure prediction was not limited to a few hundred residues?” The subsequent work to resolve that uncertainty led to innovation not just in how far prediction can scale, but also the inference time. This work gave our companies a tool no one else would have built, expanding their ability to resolve uncertainty and create value.

A fundamental assumption of my argument is that the time window to exploit an algorithmic advantage in the life sciences is not long enough to realize value. We should acknowledge, however, that pre-clinical timelines are accelerating thanks to AI, and they will continue to shrink. In a faster world, investing directly in a breakthrough algorithm may finally pay off.

But a shorter window is still a window, and a reopened race is still a race, one where the edge goes to whoever gets there first and biggest. Flagship’s advantage comes from embracing the uncertainty that produces more diverse data and methods than a faster race would ever generate. That diversity gives us more ways to build and apply AI than anyone chasing a single window. A faster field doesn't close that gap. It widens it.