Machine learning for molecule design is a fast-growing field with massive literature, to the best of our knowledge, we are the first to **comprehensively** review this field, the preprint is now available at Arxiv .
🧵1/6 Introducing Diffusion Constrained Samplers 🥳🥳🥳
Interested in optimization problems where (partial) constraints are unknown (protein/molecule design)? We show diffusion models learn it implicitly and optimize feasible solutions thru sampling!
If you are working on or interested in graph generation, you may want to check out our recent survey about deep graph generation with methods and applications, available in Arxiv now .
(1/3) I am thrilled to announce that
@AI_for_Science
workshop is back with
#NeurIPS2023
! This year we put together several new programs with a new theme "from theory to practice", including a panel discussion to align the expectation between academia and funding agencies.
Happy to announce our new initiative AI4Science101. We wrote a series of documents to encourage knowledge sharing and collection in AI for Science from both the view of AI and Science researchers to motivate them to learn, join and work on AI for Science.
(1/n) After a month of "on hold" on arXiv, I am excited to share our latest work on unlocking the potential of ML for materials discovery! ML has been successfully applied to modeling molecular structures, esp. biomolecules.
Last year, we started AI4Science101 () which aims to bridge AI and Science community with a series of blogs designed for beginners. After a year, we have 12 blogs and many more are coming soon! We hope you enjoy reading the blog and consider joining us!
2022 was a year of new and exciting ventures for me:
1. I stepped out of the ML industry and joined a startup led by a group of physicists, chemists, etc.
2. I launched AI4Science101, a blog series aimed at building a new knowledge system for AI for Science.
(1/n)
Finally accepted
@NeurIPSConf
! Cannot wait to discuss molecular dynamics, transition path sampling and stochastic optimal control in the coming December!
#NeurIPS2023
Excited to share our latest work at the intersection of machine learning and computational chemistry; Path Integral Stochastic Optimal Control for Sampling Transition Paths between molecular conformations. With
@YuanqiD
*,
@priyankjaini
, Ferry Hooft,
@BerndEnsing
and
@wellingmax
One of the coolest topics to listen at Hawaii (ICML 2023) this summer: structured probabilistic inference & generative modeling! This field has been rapidly growing and can’t wait to follow the new frontier!
Our
#ICML2023
workshop proposal "Structured Probabilistic Inference & Generative Modeling" has been accepted 🎉. We can't wait to engage in insightful discussions with experts in probabilistic ML and other areas at the beautiful Hawaii 🌴🏖️.
Check🔍:
Revisited Prof. Weinan E’s opinion paper “The Dawning of a New Era in Applied Mathematics”! The history of applied math is very inspiring and “generalize to/align with” the development of comp. tool (numerical, CS, AI) in Science. Highly recommended!
Don’t miss the deadline for the Structured Probabilistic Inference and Generative Modeling workshop at ICML 2023 (May 26th)!
We are calling submissions covering all aspects of probabilistic machine learning!
Website:
Submission:
I will be traveling to
#ICML2023
in Hawaii next week! I will present two papers, main conf track (Flexible Diffusion) and oral presentation
@TAGinDS
workshop (LEFTNet). Happy to chat about Geometric DL, Generative Models and AI for Science! DM me if you'd like to chat!
AI for Science is a fast-growing and promising field for both AI and Science community. We started building this community several years ago and are super excited to share the first paper by the community effort, lead by
@marinkazitnik
, co-lead by
@hcwww_
*,
@TianfanFu
* and me!
✈️✈️✈️I am visiting Georgia Tech tmr and next Monday! I am flattered to give a talk at the Applied and Comp Math seminar. I will be sharing some of my work and thoughts on "Accelerating Molecular Discovery with ML: A Geometric, Sampling and Optimization Perspective"!
Happy holidays, graph and geometric deep learning community! Do not forget that if you like to become one of the
@LogConference
organizers in 2024, you still have time to sign up for the application form . We will review them on January!
We tried this idea about two years ago. The interesting observation is that our architecture and modality are sensitive to different part of information thus make them “complement” each other. Glad this idea has been picked up. Joint work with
@Zhu_Yanqiao
Multi-Modal Representation Learning for Molecular Property Prediction: Sequence, Graph, Geometry
Three chemical modalities are contrasted against each other and used for property prediction.
Unfortunately, only evaluated on the MoleculeNet benchmarks
Our new work on generative models for chemical reactions: much faster inference with flow matching (OT path) training scheme, better leveraging our knowledge about the problem is a key to solve science problems! Check out the paper if you are interested!
New paper alert: React-OT: Optimal Transport for Generating Transition State in Chemical Reactions (). React-OT formulates TS search as a transport problem, approaching chemical accuracy while taking only 0.5 seconds in inference on a single GPU.
#compchem
🥳🥳🥳 The AI for Science slack channel has almost reached 1,000 users, super excited about this growth! Welcome anyone interested in AI for Science to join, chat and post any related events, resources, or hiring info!
@AI_for_Science
The third year of
@AI_for_Science
workshop
@NeurIPSConf
, covering more diverse areas, featuring more speakers from science communities,
@OpenCatalyst
competitions, and a panel from funding agencies about the future of AI for Science! Join us on December 16th!
🎉 Join us at NeurIPS 2023 AI for Science Workshop on 12/16:
7 speakers on cutting-edge AI research across fields🧠
Future-focused panel with funding agencies 💼
Open Catalyst Challenge announcement 🏆
📢Our December issue is now live! Highlights include an approach to identify transition state structures in chemical reactions, a denoising method for fluorescence images, and an approach to identify stable surface reconstructions of complex materials.
👉
I find this encouraging in my life, I was not trained enough in mathematics and physics in early years so I have a hard time catching up (slowly). I find it so relieving to admit things I don’t know and I’m learning. It doesn’t frighten me to ask question and understand why.
Don't get frightened by not knowing things. I have approximate answers, and possible beliefs, and different degrees of certainty about different things, but I'm not absolutely sure of anything. There are many things I don't know anything about. It doesn't frighten me.
I will be in NYC Dec 3-6th, traveling to NeurIPS Dec 10-17th, and staying around Seattle till January. Let me know if anyone would like to chat at any point if we come across! 🌟🌟🌟
#NeurIPS2023
I am happy to chat anything about AI for Science and Science for/of AI/Science.
Here it is: the first Learning on Graphs Conference! 🎊
We think this new venue will be valuable for the Graph/Geometric Machine Learning community.
What makes it so important+unique? See our blog post!
1/6
Many people asked me why AI for Science and I answered with several arguments and beliefs in this blog. Some interesting observations and trends in 2023 and hope the momentum continues for 2024!
🚀 Exciting News! Our blog “AI for Science in 2023: A Community Primer” is now live! In this blog, we delve into how AI intersects with various scientific fields - from Chemistry, Biology, Physics, Computer/Math. Science, Neuroscience to Earth Science.
One more week to submit your work to
@AI_for_Science
workshop at
@NeurIPSConf
!
Do not miss this opportunity to attend one of the best annual events about AI for Science!
We are also soliciting education-related papers to lower the barrier to entering the field!
#NeurIPS2023
Thanks for all the speakers, organizers, authors, reviewers, area chairs to make the AI for Science workshop a great success! Looking forward to meeting you in the coming year!
Check out our new paper using diffusion model for structure-based drug design! Diffusion models are particularly powerful with inpainting and could be suitable for many drug design scenarios. Play with our demo if you are interested!
Excited to share our latest work at the intersection of machine learning and computational chemistry; Path Integral Stochastic Optimal Control for Sampling Transition Paths between molecular conformations. With
@YuanqiD
*,
@priyankjaini
, Ferry Hooft,
@BerndEnsing
and
@wellingmax
Super excited about this topic and see how scaling plays a role in AI for Science (alone from scaling that was long studied in science)! We also put together a highly diverse program to share ideas/lessons across different fields!
🥳🥳🥳 We are excited to share that AI for Science workshop will be held again with
@icmlconf
2024, Vienna! This time, we focus on scaling in AI for Science (as a new dimension to theory, methodology and discovery)! Tentative schedules can be found:
Now accepted at
@NeurIPSConf
! Feel free to play with our code to adapt state-of-the-art ML models to your materials problems or develop your new models with diverse materials datasets!
#NeurIPS2023
(1/n) After a month of "on hold" on arXiv, I am excited to share our latest work on unlocking the potential of ML for materials discovery! ML has been successfully applied to modeling molecular structures, esp. biomolecules.
I really enjoyed reading this discussion from prestigious ML researchers about **now** and **future** of ML research, particularly
@andrewgwils
's view on discovering scientific theory, highly recommended, a wonderful holiday read!
@AI_for_Science
Happy new year, my friends on twitter! 2023 has been a challenging yet awarding year for me: both in work and life: (1) going through early PhD crisis - find out what interests me the most and what are the essentials to support them, (2) putting consistent effort in education
A nice introduction for molecular dynamics (I found it very helpful for myself when working with Yanze)! This is also under the AI4Science101 initiative, more blogs are coming out soon!
In our latest community blog, Yanze Wang and
@YuanqiD
provide an introductory overview of molecular dynamics simulations.
If you're interested in learning more, you can read the full blog here:
Interested in reviewing and learning the frontiers of probabilistic inference and generative models? Sign up this form for ICML 2023 workshop Structured Probabilistic Inference and Generative Modeling!
#ICML2023
This Monday, we discuss a new 3D GNN framework with the authors Weitao Du,
@YuanqiD
, and
@limei69990587
:
Come discuss with the authors what is new about this one more 3D GNN
Join us at 11am EDT / 5pm CEST on Zoom:
❓New to Geometric GNNs, GDL, PyTorch Geometric, etc.? Want to understand how theory/equations connect to real code?
Try this practical notebook before diving into this exciting area!
**Geometric GNNs 101**
Super excited to kick off this seminar by
@StefanoErmon
on diffusion models and applications in science! This seminar series will be open to all, live-streamed and recorded! Zoom link:
We are excited to announce the AI for Science seminar series! The seminar will feature both pioneers and Schmidt Futures AI for Science postdocs on advances and challenges at the frontier of AI for scientific discovery.
We hope you’ll join us!
Sign up for the second annual Learning on Graph conference! Follow the most recent progress in the field and enjoy the big party of graph machine learning community! 🔥🔥🔥
@LogConference
I will be traveling to
#ICML2023
in Hawaii next week! I will present two papers, main conf track (Flexible Diffusion) and oral presentation
@TAGinDS
workshop (LEFTNet). Happy to chat about Geometric DL, Generative Models and AI for Science! DM me if you'd like to chat!
We are excited to announce that this Friday, March 22nd Dr. Michael Bronstein will be joining us for a Physical Perspective on Graph Neural Networks. Hope you'll join us!
Zoom:
New Preprint! We believe the “optimal” forward process of diffusion models should be learned. Inspired by symplectic structure and Riemannian metric, we build the **first** learnable forward process of diffusion models with theoretical guarantees!
I knew
@ZimingLiu11
more than three years ago when I just entered this field. He has been a great researcher, mentor and friend for me. He re-ignited my curiosity in physics and discovery. If you are looking for someone who works on AI + physics (science), reach out to him!
This fall, I’ll be on job market looking for postdoc and faculty positions in US! My research interests span in AI + physics (science). If there’re opportunities to present in your school, institute, group, seminar, workshop etc., I really appreciate it! 🥹
We use the advances in geometric deep learning and generative models to accelerate the “searching” of transition state in chemical reactions with high efficiency and accuracy! Great work with
@chenru_duan
@KulikGroup
!
📢
@chenru_duan
,
@KulikGroup
,
@YuanqiD
and colleagues introduce a diffusion model that generates chemical reactions in 3D with all desired symmetries preserved.
👉
Join the meeting group if you are interested! Previous machine learning potential work often considers predicting force/energy to drive the simulation, while we tackle another challenging problem to sample transition paths between two states (e.g. modeling protein folding)!
Tomorrow
@HoldijkLars
presents his paper "Path Integral Stochastic Optimal Control for Sampling Transition Paths" ()
I think the SOC ideas might have even more applicability in this field!
Join on Zoom at 11am EDT / 5pm CET:
Education is the foundation of AI for Science. We are calling people who are interested in joining us and discussing the future this December at NeurIPS! If you have any questions or suggestions regarding the submission format for this track, please let us know!
We are opening a new track for Education
@AI_for_Science
workshop this year
#NeurIPS2023
. Education has been and will continue to be one of the largest gaps in AI for Science. We are calling for contributions of educational resources in flexible formats!
A very nice and long review (w/ perspective) about AI for (Physical) Science! Education and community building are indispensable yet challenging parts in
@AI_for_Science
. This paper could serve as a great educational resource for people who are interested in this promising area!
Will attend
#NeurIPS2022
in-person next week! Can’t wait to meet new and old friends (many we haven’t met due to COVID)! If you would like to chat anything with me (research, community, etc.), feel free to message! I will also be around
@AI_for_Science
workshop on December 2nd.
It was a wonderful chat with Jaanak brothers about my experience and perspective about AI for Science research. If you are interested in how I got into the field, found my direction and my thoughts about the present and future of the field, check out this podcast!
We greatly enjoyed speaking with
@YuanqiD
about research at the intersection of biology and computer science, the importance of multidisciplinary research, and the PhD journey. Please feel free to listen below! :)
We designed an object-aware equivariant diffusion model tailored for transition state generation in chemical reaction! Check more details in the preprint if you are interested!
We received around 200 submissions last year, responding to the high demand, we are openly recruiting new reviewers () and (new this year!) area chairs ()! Please sign up the form if you are interested!
(3/3) We are also extremely honored to partner with
@OpenCatalyst
and have the competition announcement as part of our workshop schedule!
For more details, please check our website:
We are looking forward to seeing you all back in New Orleans this year!
A great memory visiting friends and chatting about everything related to AI, Physics, Mechanistic Interpretability, Science of Science, etc. and (more interestingly) the smiling pattern is similar to the "God fathers of Deep Learning"!
Yes, these have also been shown in the benchmark PMO developed by
@WenhaoGao1
and
@TianfanFu
! Part of my reasoning about this is that I felt deep generative model-based methods have not yet fully used all available data. (So I'm not surprised CO methods work much better).
Very thought provoking:
"...we have shown that genetic algorithms are very strong baselines for molecular generation tasks, performing at least as well as many more complicated methods..."
@austinjtripp
@jmhernandez233
As many conf. submission/rebuttal deadlines are close, we just decided to extend our ddl for extra 7 days. Take this opportunity to prepare your submission to the best conf. in graph machine learning/geometric deep learning with high-quality reviews and engaged discussions!
The abstract submission deadline of the
#LoG
conference is extended to August 18th AoE!
Known for high-quality reviews :)
There is a nonarchival 5-page track as well.
Thanks for the feedback about the gap between abstract and paper submission deadlines!
Molecules inherently have multiple modalities and it’s interesting to note despite they may have almost exact same information in the eyes of chemists, neural architectures treat them differently. It could be a great testbed to understand inductive biases.
I will be at NeurIPS the whole week, let me know if you like to discuss anything related to
#AI4Science
, geometric deep learning, deep generative models! Don’t miss our workshop on December 2nd!
#NeurIPS2022
is taking off today! Looking forward to meeting you all on Friday, Room 388-390 for the
#AI4Science
workshop. We have a stellar lineup of speakers covering a wide range of AI for science topics, along with 5 contributed talks and amazing posters!
LoG 2022 has been a success with recognitions from our community.
In LoG 2023, we will (1) continue the reviewer award mechanism to explore the best way for our community to grow, (2) have decentralized local meetups to encourage and form local communities.
Welcome back! After a very successful first edition in 2022, we are thrilled to announce the second venue of the Learning on Graphs Conference!
📆27 - 30 November 2023
💻Virtual & free-to-attend
🤗Stronger emphasis on local meetups around the world
LoG 2023 is less than one month away, and we're super excited to hear from our keynote speakers,
@loukasa_tweet
@jure
@KyleCranmer
and Stefanie Jagelka!
🥳🥳🥳 We are recruiting new reviewers for the Structured Prob. Inference and Generative Modeling workshop
@icmlconf
2024! Sign up to follow the frontiers of prob. inference, sampling, decision-making, uncertainty, optimization and beyond with structures!
We are committed to building a home for graph machine learning and putting incentive into our peer review process! If you would like to support us, please consider sponsor us!
We are actively looking for sponsors for LoG'23
🤗95% of the sponsorship goes back to the community towards paying our top reviewers -- we think this initiative helped elevate the review quality in graph ML last year based on positive community feedback!
The current review process is totally biased by perspective of the **assigned** reviewer. Some like SOTA performance, some like cool application, some like lengthy theory and proof, some like simple/some like complicated method, etc.
As an ML community, if we want a healthy review process, we need to educate the reviewers. As area chairs, now is the time to do so. There are many papers with sound ideas that got low scores based on wrong reasons. See below what I say to some of the reviewers in those cases:
@miniapeur
It feels quite true. Do you suspect any reason behind that? Is it because I’m early days there was no clear boundary between fields as today and people do science more driven by curiosity? Interesting to retrospect this and inspire our next-generation education.
(2/3) As always, we aim to improve diversity and bring more "non-NeurIPS regulars" to attend NeurIPS! We open a new Education track to solicit collected education resources (of flexible format) to improve the knowledge collection in AI for Science.
The new LoG conference is looking for more reviewers! We have a special emphasis on review quality via high monetary rewards, a more focused conference topic, and low reviewer load (max 3 papers). But for this we need your help! Sign up here:
Final proceedings from the inaugural edition of the Learning on Graphs Conference are now available on PMLR!
Our sincere gratitude to the program committee, authors, supporters, and research community for making LoG possible 🤗
📖:
Very excited for this! After a year, we would like to look back the success and failure in AI for Science and pave the way for the future of AI for Science research! Looking forward to receiving your amazing submission and meeting you
@NeurIPSConf
!
The third AI for Science workshop is coming in-person/hybrid again with
@NeurIPSConf
, New Orleans Dec 2022. This time, we focus on the progress and promises of AI for Science and aim to discuss what has been the key to the success of AI for Science and what’s next?
With the
@LogConference
abstract deadline approaching, there are still a few days left to help out and get a stab at our best reviewer rewards! (money)
You can select how many papers to review and bid on papers at our conference!
Sign up here:
Graphein is an amazing library for network analysis on biomolecular structure and interaction network. I’m so happy to be part of it and working with
@arian_jamasb
(and the team) is so joyful! Highly recommended if you are working on this area!
Thrilled our paper “Graphein - a Python Library for Geometric Deep Learning and Network Analysis on Biomolecular Structures and Interaction Networks” has been accepted at
#NeurIPS2022
Check out the code (and give us a 🌟) here:
There are 12 days left if you want to submit your paper to the
@LogConference
!
August 11th is the deadline.
If you have exciting work related to learning on Geometries, Graphs, 3D objects ... submit it here:
Please share with friends and family :)
There are many good interpretations and discussions already. My personal take on this is that AI not only provides tools for one particular science field but the problems across fields can be understood together and cross-field discussion can be brought up again!
This paper on generalization in diffusion models is very nice
@ZKadkhodaie
gave a talk about it, and the whole audience, including me, loved it. So tomorrow we'll discuss it with her in the reading group!
On Zoom 11am ET:
Super excited to see frame-based approaches are finally getting so much attention they deserve! We had two works along this line (ClofNet 2021) and (LeftNet 2023). Local/global frames are simple, efficient, expressive and scalable! This is at least my personal de facto choice!
LoGG is one of the most open and diverse community and support the growth of the
@LogConference
community! I watched a lot of videos and they are really helpful in understanding the viewpoints of the authors!
@klindt_david
@fredericpoitev1
@ninamiolane
Nice work! We also did some early attempt in learning disentangled representation in molecules but they seem to fail because of challenges in modeling discrete structures. But we found you can actually identify semantics in the latent space .
We are hosting the first Comp Sust workshop at NeurIPS 2023! Sustainability has so many challenging problems for and needs machine learning to solve. As this field is even more heterogenous than others, we setup goals to highlight not only promises but pitfalls!
We are excited to announce the first
#NeurIPS2023
workshop on Computational Sustainability with the theme Promises and Pitfalls from Theory to Development!
📃Papers due Oct 3
✅Notification date Oct 21
🌿Workshop date Dec 15
Learn more:
There has been a surge of interest in developing foundation models in molecular ML. In practice, uncertainty quantification is key for real-world discovery workflow. We implement and benchmark many commonly used UQ methods and pre-trained models (all codes are public! ). 🔥🔥🔥
Our paper reviews recent advances in graph structure learning which have many potential applications in scientific discovery where graph structures are unknown or mis-represented. In particular, it could lead to interpretation of data such as long-range contact.
Checkout our thoroughly updated survey on graph structure learning for more details in this fast growing field!
Joint work w/ Weizhi, Jinghao,
@YuanqiD
, Jieyu,
@yangji9181
, Qiang, and Shu. 2/2
A nice concept “structure regression”, this reminds me when I started working on interpretability-related research, I was trying to find a trade-off between fully symbolic Eq and black-box NN. My answer was to enable the interaction between human and AI (interact with NNs).
Many scientific problems can be formulated as regression. In this blogpost, I argue structure regression is probably a better goal than symbolic regression. If you are interested in applying structure regression to your scientific field, please DM me!😀
🚀 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