David Ding

I like building and I cannot lie.

[David Ding @ding_david]

Do you need a billion parameters and millions of sequences for #proteindesign? Maybe not! See our paper: https://t.co/tUqcVxKowL

We show that interpretable, non-autoregressive structure-based protein design can work!

Original thread: https://t.co/ylIBfzzXcW

[Djenet Bousbaine @DjenetBousbaine]

Check out our last pub. from @mfgrp lab! We discovered that skin commensal microbes induce systemic and local B cell responses upon colonization and took advantage of this knowledge to create topical vaccines using engineered skin microbes. Please read the thread for details.

[Conor McClune @conor_mcclune]

Excited to share our new preprint on the Taxol pathway!
We developed a strategy to systematically activate and identify gene sets in non-model organisms and used it to find 8 new genes in the Taxol pathway, allowing us to biosynthesize Taxol precursors de novo.
Surprisingly, this required a new protein (FoTO1) that helps the first Taxol P450 oxidase produce the correct product. more here: https://t.co/vyUNRFKwRM congrats to coauthors
@Sattely_lab
@ctliu629 @fordycelab

[Savage Lab @SavageCatsOnly]

What limits rubisco function? Is it the chemical mechanism? Evolution? In an updated pre-print @prywes et. al explore this question by assaying >99% of single amino acid mutants in Form II rubisco (1/7)

[David Ding @ding_david]

@jakob_tr @NatureMicrobiol cool crazy stuff!

[David Ding @ding_david]

So we are blowing past 1.5C, and expecting hundreds of millions of global south folks with forced migration and newly poverty in the coming years. #ipcc

[David Ding @ding_david]

Really enjoyed this summary of recent AI progress and possible implications: https://t.co/6B8spiob1e thanks Carlo amadei

[David Ding @ding_david]

Thanks to the team @ecolienthusiast, @samsinai, @_nathanrollins, @NoamPrywes, @dfsavage, @michael_laub8, @deboramarks. Thanks also to @chloehsu0 + @SteffanPaul3, and 2 anonymous reviewers for great suggestions.

[David Ding @ding_david]

Do you need a billion parameters and millions of sequences for #proteindesign? Maybe not! See our paper: https://t.co/tUqcVxKowL

We show that interpretable, non-autoregressive structure-based protein design can work!

Original thread: https://t.co/ylIBfzzXcW

[David Ding @ding_david]

The 5050 program is giving me the tools to turn my research into a startup, everything from how to move fast to the secrets of fundraising to how to build a team around you. If you want to learn about deep tech startups, and build one, go apply!
https://t.co/kfSsIdvJiH

[Seth Bannon @sethbannon]

The SaaS startup playbook doesn’t work in deep tech. “Move fast and break things” doesn’t apply when building nuclear reactors or bioengineering cell therapies.

That’s why we built 5050, “a cheat code for starting a deep tech startup.” (quote from an alumni)

Applications open!

https://t.co/ucsabh5CyT

5050 is a program to help scientists and engineers start indispensable companies. We’ve distilled everything we know about deep tech startups into a two phase program.

Phase I: Explore
We’ll help you answer: How do I turn my breakthrough science into a business? What do I need to make a startup idea work? How do I recruit a world-class team Am I ready to be a founder?

We'll help you diagnose and mitigate risks and pivot quickly if necessary.

At the end of Explore, those ready to build a startup will be invited to the next phase, Build

Phase II: Build
You’ll join a select cohort of the most high-speed founders solving massive world problems. Build will help you de-risk their technology, hit milestones, raise a first round, and reach takeoff speed faster than you think is possible. You’ll learn everything you need to know about starting a deep tech company and what it takes to become a world class entrepreneur.

At @fiftyyears, we’ve backed over 100 deep tech startups, helped them raise over 5.5 billion dollars, and supported our founders in achieving many firsts:

➜ the first carbon-negative molecule factory
➜ the first cultivated meat approved in the U.S.
➜ the first microgeo satellite for internet connectivity
➜ the first in-orbit space pharma drug factory
➜ the first de novo synthesis of a 1000+ base DNA molecule
➜ and many more

Whether you’re validating an idea or ready to build — 5050 is for you.

Apply! ➜ https://t.co/ucsabh5CyT

[David Ding @ding_david]

@CRGenomica @MafaldaFigDias @Jonnygfrazer If you're looking for academic work: Some of the nicest mentors to have around to work on problems with significant impact!

[Centre for Genomic Regulation (CRG) @CRGenomica]

@MafaldaFigDias & @Jonnygfrazer lab is looking for someone who loves model building and is keen on studying disease from an evolutionary perspective.

Experience in statistics, deep learning, or transcriptomics is desirable but not essential.

[Kasiakready @kasiakready]

New preprint out!

Babies are born to breastfeed.

While 50% of lactating persons struggle to make enough milk, there are no FDA-approved drugs to enhance lactation.

We engineered a long-acting prolactin, Prolactin-XL,
to enhance milk production.

https://t.co/qv2D2NoU6u
1/14

[Andreja Kust @AndrejaKust]

Excited to share our preprint on cyanobacterial communities and their core microbiome, with @Dr__Diamond as corresponding author and incredible help from @jkzorz, @K_BoumaGregson, @catacruanas, @BanfieldLab. #microbiome #Genomics #WomenInSTEM

https://t.co/zhMjA8qslJ

[David Ding @ding_david]

@fkondras Totally, unsure to how many mutations we could go, though works well in datasets with some neighbouring residues mutated + your Somermeyer et al. suggests quite far. Atm, not confident about supervised 'oracle model' beyond >15 muts to assess unsupervised sampled sequences.

[David Ding @ding_david]

Do you need a billion parameter protein language model to make the right mutations?

Turns out: Learning residue mutation preferences from structure enables designing protein variants better than autoregressive methods.

https://t.co/EzhaVSaOXh
1/7

[David Ding @ding_david]

More info here: https://t.co/8jbhFkLIv2

This was great teamwork with @ecolienthusiast, @samsinai, @_nathanrollins, @NoamPrywes, @dfsavage, @michael_laub8, @deboramarks. Thanks also to @chloehsu0 + @SteffanPaul3.

7/7

[David Ding @ding_david]

So maybe most that expensive & uninterpretable sequence models are learning is just the local structural context. Building models directly on structure can be more powerful.
6/7