Maybe We Just Needed More Time to Think?

In 1905, Einstein described the theory of special relativity to the world. Thirty-five years later in the mountains of Colorado, physicists Bruno Rossi and D.B. Hall made clever measurements of subatomic particles to show the empirical evidence of time dilation implied by Einstein’s theory.

The Higgs boson was predicted by mathematical theory in 1964 — work for which Peter Higgs and François Englert later shared the Nobel. Forty-eight years later, it was empirically observed at the Large Hadron Collider.

Maxwell’s equations predicted that electromagnetic waves move at the speed of light, confirmed empirically decades later with radio waves.

It is widely appreciated in physics that theory precedes empirical understanding of nature. Yet, ask most biologists how they know something is true, and they’ll direct you to data. Question, experiment, observe, and question some more. We go to the lab to see if a hypothesis is real. In many cases, this approach works, incrementally propelling us toward understanding.

But it raises a question worth sitting with: What if biology’s slow march has been limited — not by how much data we have — but by our capacity to reason through it? Reasoning is how the biggest leaps have always happened. The opportunity now is to scale it.

Of course, theory isn’t new to biology. Like physics, many substantial breakthroughs in the history of biology have been theoretical, followed by empirical verification.

Mendel theorized a mechanism of genetic inheritance decades before Morgan’s experiments on fruit flies provided the physical evidence that genes reside on chromosomes.

Darwin’s theory of natural selection predicted intermediate forms, but not until systematic reconstruction of molecular phylogenies did we see the empirical evidence of the fundamental mechanism of that selection.

Of course, many theories are wrong, and empirical evidence proves it. Crick conceived of a brilliant model of the genetic code that required self-punctuating syntax of triplet codons, only to have it falsified by the observation that UUU coded for phenylalanine by Nirenberg.

Without theory and reasoning, empirical observation in biology constrains us to incremental progress. At Flagship Pioneering, we pursue bigger leaps in science, and that’s why at Pioneering Intelligence, we are betting on the promise of AI for scientific reasoning.

Prediction vs. Reasoning

Betting on AI to reason means being clear about what today’s most celebrated tools actually do — and don’t. AlphaFold, for example, predicts protein structures by learning to generate posterior samples from known protein structures. But AlphaFold is not reasoning. The reasoning comes from the researchers who look at the output and revise their beliefs about whether the predicted structure is right.

Tools like AlphaFold are remarkable at prediction. But prediction was never the real promise of AI. The promise is in its ability to reason abstractly and counterfactually, faster and more coherently than humans.

We’re already living with this reality: the tools millions of us reach for daily don’t just retrieve answers but think alongside us. The shift from prediction engine to reasoning engine is already here, and the proof is in your hands and mine.

Reasoning at Scale Demands the Right Infrastructure

Seen another way, Crick wasn’t limited by the evidence in front of him, but by his own ability to see through competing models and theories to their logical conclusions. Perhaps, given more time, he would have weighed the redundancy of a degenerate code, the steric hindrance in base pairing, and the constraints of amino acid biosynthesis — all of this visible in data he already had. AI reasoning enables the unprecedented ability to generate thousands of competing models and follow them to their deep and non-obvious conclusions.

This will invert how we use AI. Today, these models mostly observe data and help us interrogate hypotheses we’ve already formed. Instead, we could use them to develop deep conjectures — thousands of competing theories, reasoned through to their conclusions. The lab doesn’t go away; it is where we test optimally discriminative AI-generated hypotheses.

A conjecture-first approach doesn’t need more data — it needs the right data. We need breadth rather than depth. Instead of collecting more of the same readouts, we will need broader, heterogenous inputs, beyond the easily collected modes of contemporary -omics data.

Pioneering Intelligence is bringing this vision to fruition, developing AI reasoning that leverages the broad and diverse modalities that Flagship’s ecosystem is uniquely positioned to create as we pioneer new landscapes of biology.

Toward Theoretical Biology

Progress often comes from abandoning frameworks that once made our understanding possible. We build them to organize and communicate, but those same structures can turn into constraints. Paradigms shift by deconstructing predecessors, dropping constraints to explore new space.

We have this opportunity as we partner with machine intelligence to decipher nature’s complexities. Protein language models rely on our 20-letter amino acid alphabet, but why? Nature operates with electron density, not alphabets. Why should AI be constrained to our abstraction? Such constructs have been quite useful, but they also could be limiting us.

With the appropriate breadth of data, AI models could invent their own representations of nature. We should empower AI to reason through nature in its own ways.

And we need those approaches desperately. The world is facing a wide range of challenges for which incremental solutions are insufficient. If reasoning is how scientific breakthroughs originate, we can scale that beyond what Crick, or Darwin, or Einstein, were able to do with their remarkable human intellect.

If we embrace this scaled reasoning, our frontier is revealing nature’s verifiable, if probabilistic, fundamental principles. That is our path to compressing centuries of discovery into years.

Matt Davis is Senior Partner and Head of Pioneering Intelligence at Flagship Pioneering, where he leads the application of AI to accelerate innovation in the life sciences and beyond. He has been an AI and technology leader for over two decades and has authored numerous papers and patents spanning AI, biology, and human-computer interaction.