Lecture 8

Generative Models

Diffusion, flow matching and the dynamical-systems view of generative modelling — how to fine-tune and steer a pretrained generator toward a reward, and what it takes to make any of this work for proteins, enzymes and molecular crystals.

Chenghao Liu
90 min

Slides developed in part with Alex Tong.

Back to schedule
Slides (PDF) Recording (YouTube) Companion Tutorial



Chen et al. — “Neural Ordinary Differential Equations” Paper
NeurIPS, 2018 · Continuous-depth networks trained via the adjoint method — the origin of continuous normalizing flows
Grathwohl et al. — “FFJORD: Free-Form Continuous Dynamics for Scalable Reversible Generative Models” Paper
ICLR, 2019 · Hutchinson's trace estimator makes CNFs scale to high dimensions with unrestricted architectures
Song et al. — “Score-Based Generative Modeling through Stochastic Differential Equations” Paper
ICLR, 2021 (Oral) · Unifies score matching and diffusion under a single forward/reverse SDE framework
Sahoo et al. — “Simple and Effective Masked Diffusion Language Models” Paper
NeurIPS, 2024 · A principled variational bound for masked discrete diffusion, closing much of the gap to autoregressive LMs
Rojas et al. — “Diffuse Everything: Multimodal Diffusion Models on Arbitrary State Spaces” Paper
ICML, 2025 · A general denoising Markov model for multimodal diffusion across heterogeneous continuous/discrete spaces
Ouyang et al. — “Training Language Models to Follow Instructions with Human Feedback” Paper
NeurIPS, 2022 · The InstructGPT paper — SFT plus RLHF (PPO on a learned reward model) as the canonical alignment recipe
Wallace et al. — “Diffusion Model Alignment Using Direct Preference Optimization” Paper
CVPR, 2024 · Adapts DPO to diffusion models, aligning text-to-image generators to preference pairs with no reward model
Korbak, Perez & Buckley — “RL with KL Penalties Is Better Viewed as Bayesian Inference” Paper
Findings of EMNLP, 2022 · KL-regularized RL fine-tuning is variational inference over a Bayesian posterior — why it converges to the reward-tilted target
Dhariwal & Nichol — “Diffusion Models Beat GANs on Image Synthesis” Paper
NeurIPS, 2021 · Introduces classifier guidance and pushes diffusion sample quality past GANs on ImageNet
Ho & Salimans — “Classifier-Free Diffusion Guidance” Paper
arXiv, 2022 · Guidance without a separate classifier, by jointly training and interpolating conditional/unconditional scores
Skreta et al. — “Feynman–Kac Correctors in Diffusion: Annealing, Guidance, and Product of Experts” Paper
ICML, 2025 (Spotlight) · Weighted-SMC correctors that provably steer a pretrained diffusion model to the reward-tilted target
Del Moral, Doucet & Jasra — “Sequential Monte Carlo Samplers” Paper
J. R. Stat. Soc. B, 2006, 68, 411 · The SMC machinery behind Feynman–Kac steering: weighted particle propagation between distributions
Dryden, Thomson & White — “How Much of Protein Sequence Space Has Been Explored by Life on Earth?” Paper
J. R. Soc. Interface, 2008, 5, 953 · Bounds on the sliver of protein sequence space life has actually sampled
Jumper et al. — “Highly Accurate Protein Structure Prediction with AlphaFold” Paper
Nature, 2021, 596, 583 · Near-experimental-accuracy structure prediction from sequence alone
Watson et al. — “De Novo Design of Protein Structure and Function with RFdiffusion” Paper
Nature, 2023, 620, 1089 · Fine-tunes RoseTTAFold as a diffusion model over backbones — de novo binders, oligomers and motif scaffolds
Peng, Bezemek et al. — “Path Planning for Masked Diffusion Model Sampling” Paper
arXiv, 2025 · Splits masked-diffusion sampling into a planning stage (what to unmask) and a denoising stage — the self-correction idea
Peng, Bezemek et al. — “Planner Aware Path Learning in Diffusion Language Models Training” Paper
ICLR, 2026 (Oral) · A “Planned ELBO” that trains against planner-guided denoising paths instead of random ones
Krishna et al. — “Generalized Biomolecular Modeling and Design with RoseTTAFold All-Atom” Paper
Science, 2024, 384, eadl2528 · Extends RoseTTAFold to arbitrary biomolecular assemblies — proteins, nucleic acids, small molecules, metals
Stark et al. — “BoltzGen: Toward Universal Binder Design” Paper
bioRxiv, 2025 · An all-atom generative model unifying binder design and structure prediction, validated in the wet lab
Butcher et al. — “De Novo Design of All-Atom Biomolecular Interactions with RFdiffusion3” Paper
bioRxiv, 2025 · An all-atom diffusion model generating protein structure in context with ligands and nucleic acids, ~10× cheaper
Arnold et al. — “Genetically Programmed Chiral Organoborane Synthesis” Paper
Nature, 2017, 552, 132 · Directed evolution of a carbene-transferase cytochrome — the first enzymatic C–B bond-forming catalyst
Zhang et al. — “Enzymatic Assembly of Carbon–Carbon Bonds via Iron-Catalysed sp³ C–H Functionalization” Paper
Nature, 2019, 565, 67 · Evolved cytochrome P411 enzymes form new C–C bonds by intermolecular carbene C–H insertion
Kennemur, Ko & Arnold — “Enzymatic Stereodivergent Synthesis of Azaspiro[2.y]alkanes” Paper
J. Am. Chem. Soc., 2025, 147, 27165 · Engineered carbene transferases give scalable, stereodivergent access to azaspiroalkanes
Coelho et al. — “Olefin Cyclopropanation via Carbene Transfer Catalyzed by Engineered Cytochrome P450 Enzymes” Paper
Science, 2013, 339, 307 · The founding demonstration that engineered heme proteins catalyze abiological carbene transfer
Stenner et al. — “A De Novo Peroxidase Is Also a Promiscuous yet Stereoselective Carbene Transferase” Paper
PNAS, 2020, 117, 1419 · A fully de novo (not evolved-from-natural) heme protein catalyzing stereoselective carbene transfer
DeGrado et al. — “De Novo Design of Porphyrin-Containing Proteins as Efficient and Stereoselective Catalysts” Paper
Science, 2025, 388, 665 · Fully computationally designed porphyrin proteins catalyze carbene cyclopropanation and Si–H insertion