Yuanqi Du · Adji Dieng · Yoon Kim · Rianne van den Berg · Yoshua Bengio ICLR 2022 - Workshop Deep Generative Models for Highly Structured Data

**Concept Graph & Resume using Claude 3 Opus | Chat GPT4 | Gemini Adv | Llama 3:**

graph LR
classDef deeplearning fill:#f9d4d4, font-weight:bold, font-size:14px;
classDef equivariant fill:#d4f9d4, font-weight:bold, font-size:14px;
classDef graphnets fill:#d4d4f9, font-weight:bold, font-size:14px;
classDef generation fill:#f9f9d4, font-weight:bold, font-size:14px;
classDef proteins fill:#f9d4f9, font-weight:bold, font-size:14px;
classDef bayes fill:#d4f9f9, font-weight:bold, font-size:14px;
classDef diffusion fill:#f9d4d4, font-weight:bold, font-size:14px;
classDef counterfactual fill:#d4f9d4, font-weight:bold, font-size:14px;
classDef dynamics fill:#d4d4f9, font-weight:bold, font-size:14px;
classDef inverse fill:#f9f9d4, font-weight:bold, font-size:14px;
classDef flows fill:#f9d4f9, font-weight:bold, font-size:14px;
A[Workshop Deep Generative Models

for Highly Structured Data

ICLR 2022] --> B[Deep learning for molecules, PDEs

using graph neural nets. 1] B --> C[Equivariant graph nets combine

GNNs, equivariance for 3D. 2] B --> D[GNNs for PDEs: learned stencils,

frequency marching. 3] A --> E[Conditional generation challenge: stable,

non-toxic, synthesizable molecules. 4] A --> F[Cryo-DRAGON: deep generative model

for 3D protein reconstruction. 5] F --> G[Cryo-DRAGON: coordinate nets, VAE,

pose inference for proteins. 6] F --> H[Cryo-DRAGON discovered structures,

visualized protein dynamics. 7] F --> I[Future: ab initio reconstruction,

data analysis, benchmarking, sequence info. 8] A --> J[DAG G-Flow Nets for Bayesian

structure learning. 9] J --> K[DAG G-Flow Nets approximate

posterior of DAGs. 10] J --> L[Detailed balance for G-flow

with terminating states. 11] J --> M[DAG G-Flow Nets outperformed

on synthetic, real data. 12] A --> N[Torsional diffusion: diffusion model

for molecular conformations. 13] N --> O[Torsional diffusion restricts to

torsions, reduces dimensionality. 14] N --> P[Torsional diffusion leverages Fourier

slice theorem, hypertorus functions. 15] N --> Q[Torsional diffusion outperformed

rule-based, ML methods. 16] A --> R[MACE: model-agnostic counterfactual

explanations for predictions. 17] R --> S[MACE generates local chemical

space, labels counterfactuals. 18] R --> T[Counterfactuals give intuitive insights

for drug-like molecules. 19] R --> U[XMol package: easy-to-use

implementation of MACE. 20] A --> V[DDPMs for generating molecular

conformations, trajectories. 21] V --> W[DDPMs learn by diffusing to

noise, learning to denoise. 22] V --> X[DDPMs capture Boltzmann distribution,

sample new energy regions. 23] V --> Y[Path-sampling LSTMs excel at

non-Markovian dynamics. 24] Y --> Z[Physics constraints improve

path-sampling LSTMs. 25] A --> AA[DDRM: unsupervised inverse

problem solver using diffusion. 26] AA --> AB[DDRM operates in spectral space

for general degradation. 27] AA --> AC[DDRM outperforms in PSNR,

perceptual quality. 28] A --> AD[Semi-discrete flows via Voronoi

for bounded supports. 29] AD --> AE[Voronoi enables flexible partitioning

for dequantization, mixtures. 30] class A,B,C,D deeplearning; class C,D,E equivariant; class B,C,D graphnets; class E,F,G,H,I generation; class F,G,H,I proteins; class J,K,L,M bayes; class N,O,P,Q,V,W,X,Y,Z diffusion; class R,S,T,U counterfactual; class V,W,X,Y,Z dynamics; class AA,AB,AC inverse; class AD,AE flows;

for Highly Structured Data

ICLR 2022] --> B[Deep learning for molecules, PDEs

using graph neural nets. 1] B --> C[Equivariant graph nets combine

GNNs, equivariance for 3D. 2] B --> D[GNNs for PDEs: learned stencils,

frequency marching. 3] A --> E[Conditional generation challenge: stable,

non-toxic, synthesizable molecules. 4] A --> F[Cryo-DRAGON: deep generative model

for 3D protein reconstruction. 5] F --> G[Cryo-DRAGON: coordinate nets, VAE,

pose inference for proteins. 6] F --> H[Cryo-DRAGON discovered structures,

visualized protein dynamics. 7] F --> I[Future: ab initio reconstruction,

data analysis, benchmarking, sequence info. 8] A --> J[DAG G-Flow Nets for Bayesian

structure learning. 9] J --> K[DAG G-Flow Nets approximate

posterior of DAGs. 10] J --> L[Detailed balance for G-flow

with terminating states. 11] J --> M[DAG G-Flow Nets outperformed

on synthetic, real data. 12] A --> N[Torsional diffusion: diffusion model

for molecular conformations. 13] N --> O[Torsional diffusion restricts to

torsions, reduces dimensionality. 14] N --> P[Torsional diffusion leverages Fourier

slice theorem, hypertorus functions. 15] N --> Q[Torsional diffusion outperformed

rule-based, ML methods. 16] A --> R[MACE: model-agnostic counterfactual

explanations for predictions. 17] R --> S[MACE generates local chemical

space, labels counterfactuals. 18] R --> T[Counterfactuals give intuitive insights

for drug-like molecules. 19] R --> U[XMol package: easy-to-use

implementation of MACE. 20] A --> V[DDPMs for generating molecular

conformations, trajectories. 21] V --> W[DDPMs learn by diffusing to

noise, learning to denoise. 22] V --> X[DDPMs capture Boltzmann distribution,

sample new energy regions. 23] V --> Y[Path-sampling LSTMs excel at

non-Markovian dynamics. 24] Y --> Z[Physics constraints improve

path-sampling LSTMs. 25] A --> AA[DDRM: unsupervised inverse

problem solver using diffusion. 26] AA --> AB[DDRM operates in spectral space

for general degradation. 27] AA --> AC[DDRM outperforms in PSNR,

perceptual quality. 28] A --> AD[Semi-discrete flows via Voronoi

for bounded supports. 29] AD --> AE[Voronoi enables flexible partitioning

for dequantization, mixtures. 30] class A,B,C,D deeplearning; class C,D,E equivariant; class B,C,D graphnets; class E,F,G,H,I generation; class F,G,H,I proteins; class J,K,L,M bayes; class N,O,P,Q,V,W,X,Y,Z diffusion; class R,S,T,U counterfactual; class V,W,X,Y,Z dynamics; class AA,AB,AC inverse; class AD,AE flows;

**Resume: **

**1.-**Max Welling discussed deep learning for molecules and PDEs using graph neural networks with equivariance properties.

**2.-**Equivariant graph neural networks combine graph neural networks with equivariance to model 3D molecular structures.

**3.-**Graph neural networks for PDEs enable solving many types of PDEs with learned stencils and frequency marching.

**4.-**Conditional generation remains a challenge - generated molecules must be chemically stable, non-toxic, and synthesizable.

**5.-**Ellen Zhong presented cryo-DRAGON, a deep generative model for reconstructing 3D protein structures from 2D cryo-EM images.

**6.-**Cryo-DRAGON uses coordinate-based neural networks, a VAE architecture, and exact pose inference to model heterogeneous protein structures.

**7.-**Cryo-DRAGON was used to discover new protein structures and visualize continuous protein dynamics from cryo-EM data.

**8.-**Future work includes ab initio reconstruction, exploratory data analysis, benchmarking, and incorporating protein sequence/structure information.

**9.-**Tristan Deleu introduced DAG G-Flow Nets for Bayesian structure learning of Bayesian networks.

**10.-**DAG G-Flow Nets provide an approximation of the posterior distribution of DAGs using generative flow networks.

**11.-**A new detailed balance condition for G-flow nets with only terminating states was introduced.

**12.-**DAG G-Flow Nets outperformed other Bayesian structure learning methods on both synthetic and real data.

**13.-**Bowen Jing and Gabriel Corso presented torsional diffusion, a diffusion model for molecular conformation generation.

**14.-**Torsional diffusion restricts diffusion to torsion angles, greatly reducing dimensionality compared to diffusing atomic coordinates.

**15.-**Torsional diffusion leverages the Fourier slice theorem and special functions on the hypertorus for equivariant generation.

**16.-**Torsional diffusion significantly outperformed existing rule-based and machine learning methods for conformation generation.

**17.-**Gimhani Eriyagama introduced MACE, a model-agnostic counterfactual explanation method for explaining predictions of arbitrary black-box models.

**18.-**MACE generates a local chemical space around an input molecule and labels counterfactuals using the black-box model.

**19.-**Counterfactual explanations provide intuitive, actionable insights into model predictions for drug-like molecules.

**20.-**The open-source XMol package provides an easy-to-use implementation of the MACE algorithm.

**21.-**Prateek Tiwari presented denoising diffusion probabilistic models (DDPMs) for generating sensible molecular conformations and trajectories.

**22.-**DDPMs learn distributions over molecules by diffusing to noise and then learning to denoise samples.

**23.-**DDPMs capture the Boltzmann distribution and generate samples from new regions of the energy landscape.

**24.-**For modeling non-Markovian dynamics from time series data, path-sampling LSTMs provide state-of-the-art results.

**25.-**Adding physics-based constraints to path-sampling LSTMs improves generation quality by reducing data noise.

**26.-**Bahijja Tolulope Raimi presented DDRM, an unsupervised method for solving inverse problems using pre-trained diffusion models.

**27.-**DDRM operates in spectral space, enabling denoising and inpainting for general degradation matrices.

**28.-**DDRM outperforms previous unsupervised inverse problem solvers in both PSNR and perceptual quality.

**29.-**Ricky T.Q. Chen introduced semi-discrete normalizing flows via differentiable Voronoi tessellation for modeling bounded supports.

**30.-**Voronoi tessellation enables flexible partitioning of a continuous space for dequantization and disjoint mixture modeling.

Knowledge Vault built byDavid Vivancos 2024