VN-EGNN: E(3)-Equivariant GNNs with Virtual Nodes Enhance Protein Binding Site Identification
New method to find binding pockets of proteins. Virtual nodes allow to employ distance losses directly.
P:
C:
🤗:
The ELLIS ML4Molecules workshop will also happen this year on November 28 in VIRTUAL format!
Please find the announcement and the call for papers here:
Looking forward to your contributions!
🎾🥳DEEP LEARNING GRAND SLAM🥳🎾
(get a paper accepted at ICLR, ICML, NeurIPS within one year 😜)
ICLR2023: [MHNfs]()
ICML2023: [CLAMP]()
NeurIPS2023: Initialization for Input-Convex Nets
Credits should go to the PhD students 🎓👏
Finally, someone made the decision-tree learning differentiable. Reformulation of the classification function to dense representations; approximation of step function with sigmoids, entmax function etc.
Good results on a large number of datasets.
Generating molecule with a given 3D shape.
The method ShapeMol uses a conditional diffusion probabilistic model and equivariant layers. Benchmarked on experiments suggested by SQUID.
Self-supervised pre-trained networks are usually more robust to distribution shifts than networks pre-trained in supervised or un-supervised fashion. Thorough analysis of this behaviour:
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
Bayesian Deep Learning (BDL) Library released:
Use BDL easily: e.g. MC-Dropout with a vision transformer can readily be coded in few lines
Paper:
Repo:
New physics-inspired descriptors for non-bonded interactions. Overcoming limitations of local geometry-based ML approaches, this method incorporates long-range effects with a focus on diverse non-bonded potentials. 🚀📊
Honored to be invited to talk about ArtificiaI Intelligence in the Austrian parliament!
Will provide a re-recording of my talk here soon.
Should it be English or German?
PREFER, a Python-based framework powered by AutoSklearn, assists molecular property prediction. Effortlessly compare diverse representations and ML models for accelerated discovery. Open-source on GitHub. 🧪🔬
#Cheminformatics
#ML
Details:
At NeurIPS,
@HochreiterSepp
critized LLMs for using the parameters for storing phrases; there would be better ways, such as a modern Hopfield network, to store this. Well, here is a retrieval system with 500x less parameters that follows this approach:
What makes CLIP work?
The contrast with negatives via softmax?
The more negatives, the better -> large batch-size?
We'll answer "no" to both in our ICCV oral🤓
By introducing SigLIP, a simpler CLIP that also works better and is more scalable, we can study the extremes.
Hop in🧶
Geometric Learning concepts (Equivariance, Invariances) carried over to interpretability-methods: if a neural net is invariant to a certain transformation of the input, also the feature-importance should be invariant.
Overly complicated notation, imho
AlphaFold is also vulnerable to adversarial attacks: small changes in the protein sequence can lead to drastic changes in folding structure. Shown on data with controlled changes to the protein sequence and on COVID-19 protein variants:
One GNN encodes the graph in the usual way; a second GNN gets the input graph with an adjacency matrix, in which edges are determined by node similarity (kNN). Contrastive learning forces the GNN to preserve the structure:
This work reports strong improvements at "protein design":
This task is defined as predicting amino-acid sequence from 3D structure (coordinates) -- that means it's a kind of *inverse-AlphaFold*.
The 3rd edition of the ELLIS ML4Molecules Workshop has been announced!
⏰ Virtual event on December 8 ⏰
Call for contributions open!
Participation is free -- join us!!!
WOW! Diffusion models..
"We redesign the network layers
to preserve activation, weight, and update magnitudes on expectation. [..] systematic application of this philosophy [..] results in considerably better networks at equal computational complexity"
Machine learning and deep learning methods compared on a large number of tasks on molecules (ADME-prediction, retro-synthesis, ...) with high imbalancedness:
Large language model pre-trained on masked SMILES. Then applied to MoleculeNet.
Obviously, we can now predict clinical tox with AUC 99.7. Basically all wet-lab tox and clinical trials unnecessary now - LOL.
The community must stop to deceive itself..
Prediction of drug synergies with Deep Learning: now there seem to be sufficient data to predict the synergistic effects for new compounds . Advances the ideas of DeepSynergy ( ).
🗜️CLAMP (Contrastive Language-Assay-Molecule Pre-training), a new 🚀 method for *zero-shot drug discovery* that utilizes textual assay descriptions for molecular property prediction.
Shows that scientific language models are bad at activity prediction.
New self-supervised learning: instead of predicting the masked part of an image, the method tries to match the embedding of the masked part of the region with the embedding from the context (for the masked region). Sry, difficult without formulas.. ;)
A message-passing network (MPNN) with one virtual node, connected to all nodes, can approximate a Transformer layer. Many groups have already used such virtual nodes in the past...
We've used CLOOB to develop a search engine that unlocks querying a bioimaging database with chemical structures. The CLOOME encoder produces bioimage embeddings that can clearly distinguish new cell phenotypes:
Search engine:
We are excited to announce that our latest work, CLOOB, was accepted at this year's NeurIPS 🎉
CLOOB consistently outperforms CLIP at zero-shot transfer learning on a large variety of datasets.
🧵 1/4
Can a molecular graph be reconstructed from chemical fingerprints using machine learning? In this study ( ) a Transformer reconstructs molecules (SMILES) from ECFP4. Only 1-4% of SMILES could be correctly reconstructed -- this should work much better...
Great news! We are excited to announce that our
#ELLISforEurope
research activities are expanding further! The proposals ‘Machine Learning for Molecule Discovery’ and ‘Learning for Graphics and Vision’ have been accepted as new
#ELLISPrograms
!
#AI
#ML
🔬A “Google” for microscopy images and molecules🔬
Given an image of cells treated with a molecule, CLOOME can correctly identify this molecule – a task considered impossible even for human experts!
Work by
@ana_sanchezf
of
@AiddOne
.
Paper:
Oh, finally someone else also calling for more scientific rigor in the GNN-community :)
Methods with and without hyperparameter selection compared on longe-range graph benchmark. In 2018, we also described this "hyperparameter selection bias"..
In-Context Learning for Drug Discovery:
Embedding-based few-shot learning methods are equivalent to "in-context learning" of LLMs.
Here this concept is used again (but introduced before in MHNfs by Schimunek et al, 2023):
Excited to share our paper “A community effort in SARS-CoV-2 drug discovery" (after 3 years in work)!
We report the results of an open science community effort to identify small-molecule inhibitors against SARS-CoV-2.
Paper:
LAST REMINDER
Tomorrow (Dec 8), 9am CET, this year's
@ELLISforEurope
Machine Learning for Molecules Workshop starts!
Open for everyone to join for free!
Schedule and registration:
Predicting molecular activity with XGBoost. 📊🧪
Study on feature importance, highlighting the need for expert interpretation. Hyperparameter optimization is crucial. Valuable guidelines for
#cheminformatics
practitioners.
#ML
Paper:
Beam enumeration: an approach for goal-directed molecular generation.
Substructures with high likelihood are kept during learning and others are discarded. Language model as basis for the generation process.
Paper:
HyperPCM for predicting drug target interactions:
Join us today at the poster session of the
@AI_for_Science
workshop at
#NeurIPS2022
. 12:05pm-1pm and 5:10pm-6pm, room388!
@AiddOne
Key elements underlying molecular property prediction?
Here self-supervised learning methods for small molecules are compared against descriptor-based methods.
Another call for scientific rigor and to not over-hype GNNs for small mols..
Pretraining a GNN encoder for molecular structures on a conformation sampling task.
Builds on the smart idea (that has been around) to use a diffusion model on the atom coordinates.
NVIDIA is also generating small molecules now: ;).
VAE-like approach to generate molecules and optimize their properties by sampling from latent space. They are aware of the problems of these optimization cycles ( ).
Happy Easter! Luckily, it's Monday and arxiv does not take vacation:
A graph neural network approach to predict chemical properties using an initial (bad) 3D conformation. Then gradually improves this conformation with de-noising. Good results on QM9:
Study of drug-ranking with anti-cancer properties for particular cell lines characterized by their gene expression profile.
Interesting fact 1: bilinear regression is used to combine modalities
Fact 2: Morgan FPs plus MLP worked better than GNN (sry folks)
Neural architecture searching (NAS) for new graph neural network architectures. The found architectures (Table 8 in Appendix) are relatively shallow and often LSTMs are selected to aggregate layers. Tests on non-molecule tasks (CiteSeer, PubMed).
New way of goal-directed optimization of molecular structure: GNNs reformulated as mixed integer linear programming (MILP) problems. Solvers for those problems then provide chemical structures.
Paper:
OPEN POSITIONS!
We have open positions for PhD students and PostDocs at my lab both in core machine learning as well as in "Machine Learning in Life Sciences":
🔥🔥🔥 Contrastive learning unleashed 🔥🔥🔥
A powerful, transferable MICROSCOPY IMAGE and MOLECULE ENCODER has been trained on CellPainting data through self-supervised learning (SSL).
Paper:
Code:
App:
HyperPCM: Robust Task-Conditioned Modeling of Drug–Target Interactions
Hypernetworks can carry over information from one protein activity prediction task to another.
work by
@EmmaJMSvensson
in
@AiddOne
Large dataset for ACTIVITY CLIFF modeling consists of 400K matched molecular pairs. Deep Learning models perform ok, again ECFP plus deep multi-layer perceptrons perform best -- in accordance with works by other groups...
Exploring Causality in Single-Cell Genomics 🧬🤖 ML adapts causal techniques to high-dimensional single-cell data, addressing challenges and paving the way for informed experimental design. 📊🔬
Paper:
Multi-modal anything to anything :)
NextGPT can take text, image, audio or video as input and generate output in these modalities. Quite efficient because only input projectors and output are trained.
Self-supervised representation learning for time series:
a sequence is split into past and future, and the learned representations have to be similar. Does not need negatives -- I don't see why mode collapse should not happen..
It's a matter of the initial molecular representations. That's why I always for using ECFP/Morgan plus MLPs as baseline in all studies. Would have saved us from the pile of the zillion molecule encoders that we are now facing...
Activity cliff: two molecules have similar structure, but a big difference in bioactivity.
ML approaches don't do well with activity cliffs, but 'old-school' ML models tend to work better than deep learning.
@DerekvTilborg
@korney34
@fra_grisoni
🚀 Transforming molecules discovery with few-shot learning 🚀! Our new method **MHNfs** enriches molecule representations with CONTEXT. MHNfs sets a new state-of-the-art on the FS-Mol benchmark dataset.
#drugdiscovery
#fewshotlearning
#AI
Paper:
Context dependent molecular representations are enabled by LLM-like architectures and self-supervised learning strategies.
Performance and benchmarking are still limited.
Paper:
I really appreciate this result about the expressiveness of the molecular fingerprint representations.
The authors used transformers to translate from chemical fingerprints to SMILES; showed that conversion re-construct the connectivity of the molecule
Please consider submitting to your work to our special issue "AI meets Toxicology"!
We are happy to get submissions until November 30 (will remind you again ;)) !
Researchers have been striving to bridge BIOLOGY 🧫 and CHEMISTRY ⚗️ via transcriptomics, metabolomics, and whatever....
... funny that the bridge turned out to be the oldest biological technique: CELL MICROSCOPY.
Generative model for molecules in 3D space: using latent diffusions on point-structured latent spaces.
As for all generative models,metrics is difficult: stability & molecule stability, validity, and validity & uniqueness used.
How good are large language models at answering questions about PATHWAYS, MOLECULAR INTERACTIONS, and MECHANISMS?
Answer: Not too good, but the LLM trained on bio-literature (BioGPT) are better than the usual GPTs, etc..
Study:
Offline black-box optimization with diffusion models. The denoising process is conditioned on the label (output of black-box function). Also used for optimizing molecules (ChEMBL), but unclear how this was exactly done...
Based on seminal work of using Normalizing Flows to learn Boltzmann distributions, here is an approach that tries to alleviate the problem of the costly simulation of training data:
Combining Bayesian Neural Networks and Contrastive Learning: Unlabelled samples are used to learn a prior distribution of weights in a style similar to SimCLR.
From
@patwalters
at ELLIS ML4Molecules:
**STOP using MoleculeNet and TDC!**
I fully support this statement (and have criticized those datasets myself for a long time)
Modeling long range intramolecular forces by extending message-passing networks: on addition to the local messages, nodes are mapped to non-local Fourier space and updated there (Ewald message passing). Improves MAE on energy in OC20:
Almost identical work to CLAMP "Contrastive Language-Assay-Molecule-Pretraining" (ICML2023): common space for natural language and mols.
CLAMP allows to steer activity prediction with natural language models for zero-shot activity/property prediction.
Introducing our new work:
GIMLET: A Unified Graph-Text Model for Instruction-Based Molecule Zero-Shot Learning
Say goodbye to the supervised molecule property prediction and embrace the instruction-based zero-shot paradigm with GIMLET!
Since the LLM community has an evolutionary tree, I thought we should have one, too.. :)
Sharing my personal perspective on the evolution of Deep Activitiy prediction Models (DAMs).
PDF w links
Bib
A survey of LLMs with a practical guide and evolutionary tree.
Number of LLMs from Meta = 7
Number of open source LLMs from Meta = 7
The architecture nomenclature for LLMs is somewhat confusing and unfortunate.
What's called "encoder only" actually has an encoder and a decoder…
Self-supervised learning for molecular graphs is still in its infancy according to this paper: . For example, on the MUV dataset, the methods don't even beat a random method. Downstream: molecular property prediction. Suggest a new suite of probe tasks.
Transformer architecture learns chemical substructures together with natural language and can answer to questions about chemistry:
(similar to our BioassayCLR method, but less focused on activity/property prediction).
Wow, this is really cool:
Contrastive pre-training on DNA sequences an the human genome to build a retrieval system. Then making a DNA vector database -- allows to align reads via maximum-inner-product-search!