Let’s not call it AI

How machine learning changes the story of what’s possible in biology.

A conversation with Vytas SunSpiral on his quest to push the Zymergen envelope in technology and biology.

This article originally appeared here.

Vytas SunSpiral

Vytas SunSpiral thinks the field of artificial intelligence is kind of cursed. Not because it can’t solve big problems across many domains — it’s already done that — but because of its name.

“‘Intelligence’ is a loaded word,” he says. “AI has come to mean anything that resembles human intelligence that we don’t yet know how to do.”

As an example, Vytas says at one time there was a quest in the field to find good search algorithms for very large databases — something we take for granted today but back then might have been called artificial intelligence.

“We figured out search and it was like, yeah, that’s just an algorithm,” he says. It’s a story that has played out time and again in the field, and it has given rise to some funny if cynical tropes about how we define AI.

At Zymergen, Vytas and others tend to use the less-hyped term “machine learning” to describe how they build models based on sample data to make predictions without being explicitly programmed to do so. The truth, however, is more complicated. Zymergen’s Data Sciences team uses a wide range of tools — from simple linear regression to very complex neural networks — to do everything from designing genetic pathways to predicting how well small-scale fermentation will work at commercial scale.

One of the teams that Vytas leads is the Computational Product Development team, which develops computational tools for material innovation at Zymergen. He finds places where technologies like computational chemistry and machine learning can be brought together to make better products in a better way.

Exploring the frontiers of biomaterials — in silico

Computational modeling of materials is one such area, developed partly out of need and partly out of foresight.

First, a little background. Most materials these days — from phone screens to bubble wrap — are made from a half-dozen or so petrochemical building blocks. This has seriously stifled materials innovation: it’s the reason your phone screen breaks when you drop it and there’s no such thing as flying cars. Biofacturing companies, like Zymergen, break out of the limitations of conventional manufacturing in two ways. First, we scan nature’s vast molecular catalog for molecules with properties that are exquisitely tailored to a specific application. Second, we use nature’s microbial factories to ferment those chemicals in facilities that look more like wineries than refineries.

Some of these biomolecules are easy to make in cells, but they haven’t been pursued in the past because they are complex and difficult to make with classical chemistry. Because of this, they are not available “off-the-shelf” from chemical catalogs to inexpensively order up in amounts needed to make test products in a lab, the way a conventional materials process might do.

As Vytas puts it, “We’re opening up a whole new space of chemistry by looking at the molecules you can build from biology. All sorts of molecules that no one else works with because they’re difficult and expensive to make with traditional chemistry.”

So how do you test materials when the chemicals to make them don’t yet exist?

The solution: Vytas’s team helped implement predictive models that simulate how molecules will perform in a product long before scale-up. For Zymergen’s family of electronic films, the cluster of servers has predicted the thermal and mechanical properties of hundreds of candidate polymer films that can be made with biology. The team has also started deploying software that will help with the much trickier task of predicting the optical properties of these films, using sophisticated modeling that goes all the way down to the electron level.

Image showing Computational Simulation of film properties

Using computational simulations like this one can usefully predict the mechanical properties of films before they are ever made

“It can take weeks and months to synthesize the molecules, formulate a film, and then test it,” Vytas says. “Our goal is to make this process faster and more repeatable, and this approach is one way in which Zymergen embodies the ethos of a Silicon Valley technology company in its drive to disrupt the materials industry.”

The scale-up challenge: churning out good ideas in commercial quantities

Simulating how a material will perform in the real world is just one important piece of the computational puzzle when it comes to designing and scaling bio-inspired products. Even if you find the perfect biomolecule and cajole a microbe to make it in small amounts, there’s no guarantee that you can make it in large amounts at an economical cost. The challenge of scale-up is one of the key barriers that has prevented the early pioneers of synthetic biology from realizing its potential.

Machine learning helps overcome this challenge. Vytas points to a suite of algorithms developed by various teams at Zymergen that help design host microbes. They can generate thousands of variants of a molecule-producing organism. These tools are particularly good at identifying changes in parts of the genome whose relationship to the molecular pathway is unexplainable by humans. Such unexplained changes may account for more than half of the performance-improving changes in the microbe.

Once these microbes are created and tested in small-scale fermentation facilities, another machine learning tool expertly predicts how they will perform in large-scale commercial fermentation tanks. And yet another tool can identify the best solvent for extracting the desired product at the end of fermentation — helping scientists determine whether the downstream processing (DSP) and economics are going to work out at all.

Vytas says that machine learning’s greatest value doesn’t come from any single superstar tool. “It’s these everyday tools all along the value chain that combine to make machine learning a real force designed to help accelerate our material innovation pipeline,” he says.

The role of culture in technology

Beyond the Computational Product Discovery Team, Vytas also leads the Advanced Technologies Program, a three-person team that pushes the technology envelope at Zymergen by prototyping systems geared for big, long-term R&D innovations.

In his work, Vytas says introducing a new technology is only one part of the challenge. Another big part is about changing the story of what’s possible within the company. He says his projects often bridge existing departments and workflows — in places where project ownership is fuzzy and everyone’s busy with their real jobs.

“If a person from our team can dedicate themselves to driving a certain idea forward, other people will collaborate with them because they have a part of it. They couldn’t own the whole thing, but they’ll dive in and become part of it.”

When people see the possibilities of a new way of doing things, he says, that’s when it changes from a story about early adoption to one about working toward putting better ideas into use.

That same adoption story is playing out now with Zymergen’s biofacturing approach. As industries as diverse as electronics, agriculture, and consumer care begin to realize that biology provides better designs and better ways of making things, we expect that more and more will come into the fold, leading to a virtuous cycle that yields not just better products, but a better world. In time, we may look back and wonder why we all didn’t dive in sooner.