Releasing 100,000+ enzyme sequence - Fitness Data Points — and the method behind them

Now published: over 100,000 enzyme sequence-performance data points to the AI-guided enzyme engineering community, and the method behind its collection so that you can do it too.

‍

1. The problem

AI has immense promise for revolutionizing enzyme engineering, but this promise is fundamentally constrained by data scarcity.

We've become great at building river channels (learning algorithms), but are short on water source (data).

‍

2. The solution

To solve the data drought, we developed MillionFull for collecting enzyme sequence-fitness data, with massive throughput.

‍

Typical automation platform per run: 10^2 - 10^4 data points.

MillionFull output per run: 𝟏𝟎^𝟓 - 𝟏𝟎^𝟕 𝐝𝐚𝐭𝐚 𝐩𝐨𝐢𝐧𝐭𝐬.

‍

This is 𝟏𝟎𝟎𝟎𝐱 more data.

‍

Typical cost on a highly optimized robotic platform: ~$30 per data point.

MillionFull cost: $𝟎.𝟑-𝟎.𝟎𝟎𝟑 𝐩𝐞𝐫 𝐝𝐚𝐭𝐚 𝐩𝐨𝐢𝐧𝐭

‍

.This is 𝟏𝟎𝟎 - 𝟏𝟎,𝟎𝟎𝟎𝐱 lower cost.

‍

In our published demo project, we identified, in one go:

>10,000 sequences with improved performance;

>400 point mutations and >18,000 epistatic interactions with positive effects;

>2000 point mutations and >50,000 epistatic interactions with negative effects.

‍

This is a lot of "water" for data-thirsty AI river channels.

‍

Find the paper in the comments:

"MillionFull enables massive, full-length enzyme sequence-fitness data collection at low cost for machine learning-guided enzyme engineering"

‍

3. Why publish

From the start, our plan was to share both the method and the first dataset.

‍

We stand on the shoulders of giants. The open-source ML community has given us amazing protein models and algorithms, freeing us to focus on what we are best at solving: data collection.

‍

We hope this dataset helps sharpen the next generation of AI for enzymes.

‍

(Open for non-commercial use. Patent filed Dec 2023; ENZIDIA is finalizing exclusive commercial licensing. For commercial interests in data use, data collection partnership, or sub-licensing, please contact us.)

‍

4. The team

MillionFull is highly multidisciplinary, so we had a highly multidisciplinary team to get it done, Bjarke, Simon, Sonia, Kenan, Soeren, Lei, Alex, with generous help from: Bruce, Jason, Le, Feiran, Jonathan, Paul, Viji, André, Arsenios, Linda, Adrian, Scott, Troels, Tue, Zofia, Keyan, Se Hyeuk, Christoffer, Christina, Lena, DTU Biosustain.

‍

5. What's next

This manuscript is a story in progress. We are collaborating closely with:

‍

‍Sonia Yuan (already an incredible contributor to the current version) and Jason Yang in Prof. Frances Arnold's group at Caltech (Nobel Prize in Chemistry, 2018), andTyler Korman, Bastian Vögeli, Michael Heltzen from eXoZymes (Nasdaq: EXOZ),

‍

- to find out what surprises could emerge when massive data, cutting-edge AI, and rapid variant prototyping are combined.

‍

‍

‍

‍

‍

‍

‍

‍

‍

‍

‍

‍

‍

‍

‍

‍

‍

‍

‍

‍

‍