What makes problems ripe for disruption through DL? Why are there limited breakthroughs in Bio x AI so far? In this post we argue that not every problem is equally suited but there are clear common patterns - and to unlock new problems we need to move beyond “white-box” data
Like for so many,
@mmbronstein
's lectures & papers on Geometric Deep Learning really got me into the field. Incredibly thrilled to now join forces on our journey to unlock one of the most promising novel modalities!
When
@sacdallago
pinged us for drug discovery use-cases for an LLM-based end-to-end pipeline for a review on protein design, of course our first question was: can we use this to design molecular glues? Turns out you probably can! 🧵 (1/5)
Our pre-print cataloguing generative deep learning tools for protein design is available on
@biorxivpreprint
: . We aimed it at practitioners wondering how to navigate design tools. Short 🧵 for one of the stories in the manu...
📢 The VantAI team is growing in Europe! A picture of today's team lunch in Zurich with
@mmbronstein
in town. We're hiring for Machine Learning, Engineering, Comp Bio/Chem and all other scientific positions. E-Mail at recruiting
@vant
.ai
What an incredible honor to work with some of the pioneers in our field on this paradigm shifting modality. Was great to sit down with
@Forbes
and chat about what makes this so exciting!
Excited to support this meetup and also give a talk about our recent work on protein surface-based diffusion. Make sure to chat with
@AkdelMehmet
,
@clemensisert
to learn more about some of the exciting science at VantAI if you are there.
📢 Excited to announce the establishment of our scientific advisory board - pioneers of protein design, generative chemistry and structural proteomics. Welcome professors Bruno Correia, Philippe Schwaller and Fan Liu to VantAI! Stay tuned for more today!
If you're a graduate student interested in & have experience with research in GDL, diffusion models, protein & small molecule representation learning,
@vant_ai
has summer & fall graduate internships in our New York and Basel Offices!
#compchem
#compbio
#MachineLearning
Excited to share what
@RebeccaNeeser
has been up to during her internship at
@vant_ai
- with ShapeLinker we introduce a novel, fast attention-based point cloud alignment to allow RL-driven PROTAC generation in 3D. All open source & permissively licensed!
Excited to finally be able to share our ICLR work critically analyzing the capacity of deep learning docking methods to generalize and how to improve this (spoiler scaling, augmentation and RL)! With this, we release a new significantly improved version of DiffDock!
A thread! 🧵
🎷🧠 VantAI will be at
#NeurIPS2023
next week. Meet us to chat about our exciting science, accepted papers and open research positions! Reach out to hi
@vant
.ai or
@NaefLuca
We are excited to present some our work, see below 1/5 🧵
Huge tour de force of GDL in 2023
@michael_galkin
and
@mmbronstein
, with a small contribution from yours truly on exciting things and what’s next for 2024 in our field
📣Two new blog posts - a comprehensive review of Graph and Geometric ML in 2023 with predictions for 2024.
Together with
@mmbronstein
, we asked 30 academic and industrial experts about the most important things happened in their areas and open challenges to be solved.
🧵 1/n
Please join me and an absolutely stellar panel to discuss AI-Driven Drug Discovery at
#GTC24
. I'll be joined by
@DaphneKoller
Gary Glick Edith Hessel and
@NaefLuca
.
It'll be a packed house!
And we're still growing! If you want to join a fast growth startup defining the future of induced proximity using Machine Learning, reach out! We're hiring across
#software
,
#MachineLearning
,
#compchem
and
#compbio
📅 Mark your calendar for this Friday!
Dive into our monthly lecture on Generative AI in Drug Discovery with Prof. Michael Bronstein & Prof. Bruno Correia
Topic: "Harnessing Geometry For Molecular Design"
When: October 6, 11 am - 12 pm ET / 5 - 6 pm CET
How: Join live via
@MichaelRetchin
VantAI () has open positions for Computational Bio, Computational Chemistry, ML and Software - Machine Learning for Targeted Protein Degradation
We've been big fans of
@elizabeth_dlha
's work over at
@DeepnoteHQ
. Interactivity not only much quicker but also allows chemists/biologists working live alongside ML Engineers. Excited for all the future features the team has been adding at breakneck speed!
I'm so excited to share our very first case study - a story of how VantAI, a biotech startup and the team led by
@NaefLuca
uses
@DeepnoteHQ
to tighten the feedback loop and move faster in computational drug discovery.
[🧵 with key highlights]
Excited to announce that
@vant_ai
will be working together with the fantastic scientists at
@Boehringer
to develop protein degraders against some of the most challenging disease targets
VantAI is excited to announce our research collaboration with
@Boehringer
Ingelheim today, which will focus on designing targeted protein degraders against traditionally undruggable targets using our geometric deep learning platform
Search engines suck - very happy to see companies like
@sourcegraph
. Neat for searching code usage examples. Really hope they'll support searching SO soon too. Somewhat surprised that
@TeejDeVries
hasn't provided a nvim plugin yet, so here's a quick hack
New interview with
@NaefLuca
, CTO at
@vant_ai
, an AI-first biotech company about advanced analytics in drug development, the field of "induced proximity" and what he has learned along the way!
#healthtech
#biotech
Incredibly excited to announce
@vant_ai
's multi-year collaboration with
@JanssenGlobal
. Couldn't be happier to work with such a fantastic group of scientists on the discovery of novel hetero-bifunctionals, glues and E3s
this week in bio:
- an evaluation of biosecurity risks of LLMs
- engineering linkers for PROTACs using reinforcement learning
- lessons in leadership+management in pharma R&D
- self-supervised transformer for predicting structure of unknown metabolites
Large multimodal models for accelerated science.
Is anyone with a strong engineering background interested in a **postdoc** at EPFL on scientific foundation/frontier model training and benchmarking?
Send me an email with your CV/motivation and join the vibrant
@SchwallerGroup
!
🚀 Excited for
#LoGNYC2024
? Dive deep into the world of machine learning geometries with us Tomorrow! 🌐 Learn how unlocking non-trivial geometries can revolutionize graph embeddings, dynamical systems, & GNN models. Don't miss out!
We will post the Zoom link later today
@thomasstruble
+ Experience in a fast moving field is at a steep discount rate. VantAI has no experience requirements - we only care about what you can build. If you learned that in short time even better! Hire for slope, not intercept
We screened ProtGPT2-generated sequences and surprisingly, even without any conditioning, for 1345/100k sequences Embeddings-based Annotation Transfer (EAT) predicted them to be biosynthesis genes of secondary metabolites (BGC), with 234 being similar to known BGC (3/5)
@alexeyguzey
+pytorch/tf/onnx, incredible productivity benefits. pretty unique in that in ML, working for a large for-profit company actually leads to higher chances of open sourcing game changing tools. also includes niche things like DL protein embeddings, like
@joshim5
's work
@SuperScienceGrl
We use
@DeepnoteHQ
- you can just fix anything they break in real time and commenting on individual code cells where they might have issues
@nic__carter
imo pretty useless if you already have an apple watch. i had one and returned it. here's a comparison both in my exp and according to publications/tests I found: Dreem 2 > Apple Watch >> Oura (check other vid of that channel)
@MoAlQuraishi
agree! also many people will be happy about torch. Could make sense to start with a fine-tuning use-case vs retraining given cost/benefit tradeoff?
@JocelynnPearl
Hottest take - just use a local first, open source, git controlled and markdown based solution like
@logseq
with first class pdf annotation support
Much more to come in this direction from Rebecca as she moves on to do her PhD with the great
@pschwllr
and
@befcorreia
. And we're always open for other research internships and full time positions!
@Atelfo
is one of the most knowledgeable people in our space, especially around where things are broken and need fixing. He knows all the knobs and dials where this tech will have impact. 10/10 would recommend for anybody looking for a new challenge!
Come work with me!
DM/email me if you're a software developer who wants to solve interesting problems in biotech using automation & LLMs. Open to informal/exporatory chats too, even if you're just curious
This shows the potential of using even unconditioned LLMs to, for example, design seed sequences in a directed evolution screen for molecular glues. Was great to work together with
@ferruz_noelia
,
@agoncear
,
@akdelmehmet
and
@sacdallago
, and much more to come! (5/5)
@andrewwhite01
that's what we do whenever we're speed bottlenecked. We often wrap RDKit in Rust then implement logic in Rust, which is less pain than dealing with c++/python imo. It's really seamless. Also check
@nu_shell
which provides dataframes backed by polars right in shell
@IanColdwater
VantAI is hiring for various open roles in the US to develop a novel class of medicines - rust,
python, k8s, GO and TS with lots exciting science!
@roydanroy
switched from i9 64gb mbp 16in to a 16gb mb air - it's 1/4 of the price and way faster in everything from compilation to training to UI rendering
@Niklas_TR
@ArtirKel
@LNuzhna
sounds like a job for logseq - has all these + plus fantastic paper workflow (native zotero + pdf viewer). My journey was Roam-> RemNote->Obsidian->Logseq
@sacdallago
Nature has long used molecular glue mechanisms to modulate not only their own proteome post-translationally, but also those of friends and foes. Given the difficulty of rational discovery of molecular glues, taking a page out of nature's book would be highly desirable! (2/5)
@Niklas_TR
@Dereklowe
useful for what it is - heavily biased towards certain target classes/chemotypes/binding modes. Important to explicitly measure what your model claims to be able to do (rarely done, esp. wrt applicability domain) and don't make claims beyond that
@MoritzW42
@JoinEdyfi
@southpkcommons
@launch_house
lived in a hackerhouse at university ave in palo alto during my time at Stanford. Best time ever, so many ppl I met that changed my trajectory - should be way more prevalent
@draparente
Selina is good option to stay w/ coworking space and fast wifi. they have monthly passes called coliv. Wework for cowork - brooklyn navy yard + union square best locations imo
One sequence showed strong similarity with a gene producing fusicoccin A, a known molecular glue phytotoxin further developed as an anti-cancer agent. Despite minimal sequence similarity, it was predicted to fold highly similar to know BGCs. (4/5)
@singareddynm
we've built so many tools that we'd love to open source, are game changing for our internal workflows, but just don't have bandwidth to do it + maintain, or no longer core focus. Clear, very fast process, funding mechanism to transfer in public domain would go a long way
@amos_schorr
@Halsted_19
I think both will happen - AI has ever expanding sphere of capabilities - if you are w/in capability frontier for the problem you're trying to solve, you'll need little specialization. As time goes on we will both have more and more problems within the sphere of capabilities, and
@LauraDeming
Alternative view to
@LNuzhna
's argument: external brains should maximize recall. If you agree with this premise, format is quite important. I use
@logseq
because 1) open source 2) local first 3) MD-based 4) built in Spaced Repetion 5) fastest PDF viewer
@draparente
Meroma is an absolutely amazing hidden gem, tell Rodney I said hi if you make it. Pujol is a nice experience but not a must - Quintonil much better I heard (only been to Pujol)
@KarlPfleger
@DIYPerks
My current vertical setup is 16in mbp + + put at the back vs bottom to stabilize hinge + - very happy, fast to set up since magnet snaps in place - had clips before but break easily when travel & slow setup
@andrewwhite01
- combines all w/ added benefit of allowing using imessage on android (except email, but email is mostly recruiter and sales spam anyways)
@TeejDeVries
@sourcegraph
global search (GH/BitBucket/GitLab) and then option to insert matching snipped and allow switching params - pretty please 😍 new paradigm of code completion