Machine Learning for Molecules

• Course overview and introductions. Course goals, structure, expectations, and logistics.

• Molecular and material representations. SMILES and InChI, graphs, fingerprints, 3D coordinates, chirality, crystals, and symmetry. Key papers: Duvenaud et al., Neural Molecular Fingerprints (2015)
• Quantum foundations. The Schrödinger equation, wavefunctions, orbitals, density functional theory, molecular orbitals, and band structure.
• Chemical and statistical foundations. Chemical forces, statistical mechanics, free energy, molecular dynamics, potential-energy surfaces, and reaction kinetics.
• Drug discovery. Small molecules and biologics, pharmacokinetics and pharmacodynamics, target identification, hit-to-lead optimization, clinical trials, and experimental assays. Key papers: Wallach et al., AtomNet (2015)
• Materials discovery. Properties, stability, synthesis, characterization, batteries, semiconductors, catalysis, and direct air capture. Key papers: Xie and Grossman, CGCNN (2018)

• Message-passing graph neural networks and graph transformers. Message passing, attention, positional encodings, and global token interactions; GCN, GraphSAGE, GIN, GAT, and Graphormer. Key papers: Gilmer et al., MPNNs (2017) · Ying et al., Graphormer (2021)
• Molecular property prediction and benchmarks. Supervised prediction from fingerprints, graphs, and three-dimensional representations, with molecular benchmarks and foundation models. Key papers: Wu et al., MoleculeNet (2018) · Zhou et al., Uni-Mol (2023)
• Geometric deep learning foundations. Geometric graphs, Euclidean groups, group actions and representations, invariance, equivariance, and spherical harmonics. Key papers: Bronstein et al., Geometric Deep Learning (2021)
• Invariant models. SchNet, DimeNet, and GemNet. Key papers: Schütt et al., SchNet (2018) · Gasteiger et al., DimeNet (2020) · Gasteiger et al., GemNet (2021)
• Vector-equivariant models. EGNN and PaiNN. Key papers: Satorras et al., EGNN (2021) · Schütt et al., PaiNN (2021)
• Spherical tensor models. Irreducible representations, Clebsch-Gordan products, Tensor Field Networks, and NequIP. Key papers: Thomas et al., Tensor Field Networks (2018) · Batzner et al., NequIP (2022)
• Modern interatomic-potential architectures. Modern machine-learning interatomic potentials for energy and force prediction; MACE, eSCN, eSEN, Orb-v3, and UMA. Key papers: Batatia et al., MACE (2022) · Rhodes et al., Orb-v3 (2025) · Wood et al., UMA (2025)

• Midterm examination period. No class.

• Probabilistic modeling and continuous normalizing flows. Probability distributions, change of variables, likelihoods, transport maps, neural ODEs, and continuous normalizing flows. Key papers: Chen et al., Neural ODEs (2018) · Grathwohl et al., FFJORD (2019)
• Stochastic dynamics and probability evolution. ODEs and SDEs, Brownian motion, Itô calculus, Fokker-Planck equations, Langevin dynamics, continuity equations, and numerical simulation.
• Score-based diffusion models. Forward and reverse diffusion, score functions, denoising score matching, reverse-time SDEs, and probability-flow ODEs. Key papers: Ho et al., Denoising Diffusion Probabilistic Models (2020) · Song et al., Score-Based Generative Modeling through SDEs (2021)
• Flow matching and stochastic interpolants. Conditional probability paths, marginal velocity fields, simulation-free training objectives, and stochastic interpolants. Key papers: Lipman et al., Flow Matching for Generative Modeling (2023) · Albergo et al., Stochastic Interpolants (2023)
• Geometric generative processes. Generative modeling on manifolds, tangent vector fields, rotations, SO(3), periodic coordinates, and equivariant probability paths. Key papers: Chen and Lipman, Flow Matching on General Geometries (2024)
• Discrete generative processes. Discrete diffusion and flow matching, continuous-time Markov chains, molecular graphs, sequences, and hybrid discrete-continuous states. Key papers: Campbell et al., Continuous-Time Discrete Denoising (2022) · Gat et al., Discrete Flow Matching (2024)
• Conditional and controlled generation. Conditioning, guidance, energy-based steering, composition, constrained generation, and inverse problems; how these methods modify a base distribution for molecular design. Key papers: Dhariwal and Nichol, Classifier Guidance (2021) · Ho and Salimans, Classifier-Free Guidance (2022)
• Modern molecular generative systems. Diffusion and flow models for molecular graphs, three-dimensional molecules, conformers, protein-ligand structures, protein structures, and crystals across continuous, discrete, geometric, and hybrid state spaces. Key papers: Hoogeboom et al., Equivariant Diffusion for Molecule Generation in 3D (2022) · Corso et al., DiffDock (2023) · Hassan et al., ET-Flow (2024)

• Molecular modeling and simulation. Molecular dynamics, learned collective variables, free-energy estimation, enhanced sampling, time-coarsened dynamics, and learned atomic transport. Key papers: Noé et al., Boltzmann Generators (2019) · Klein et al., Timewarp (2023) · Nam et al., LiFlow (2025)
• Quantum chemistry and electronic structure. Neural wavefunctions, learned density functionals, machine-learning Hamiltonians, electron densities, orbitals, energies, and forces. Key papers: Pfau et al., FermiNet (2020) · Kirkpatrick et al., DM21 (2021) · Kim et al., QHFlow2 (2026)
• Crystal and materials discovery. Crystal representations, structure and property prediction, stability, generative discovery, and materials foundation models. Key papers: Merchant et al., GNoME (2023) · Zeni et al., MatterGen (2025)
• Goal-directed and synthesis-aware molecular design. Property guidance, reinforcement learning, Bayesian optimization, and GFlowNets for goal-directed and multi-objective design; reaction prediction, retrosynthesis, route search, and reaction-based generation for synthetic feasibility, with oracle reliability and experimental feasibility as design constraints. Key papers: Koziarski et al., RGFN (2024) · Seo et al., RxnFlow (2025) · Gao et al., SynFormer (2025)
• Protein structure prediction. Sequence-to-structure models combine evolutionary information, protein language models, geometric reasoning, and confidence estimation. Key papers: Jumper et al., AlphaFold 2 (2021) · Abramson et al., AlphaFold 3 (2024)
• Protein representation and design. Sequence-based representation learning and single-sequence folding with ESM-2 and ESMFold; structure-conditioned inverse folding with ProteinMPNN; generative backbone design with RFdiffusion. Key papers: Lin et al., ESM-2 and ESMFold (2023) · Dauparas et al., ProteinMPNN (2022) · Watson et al., RFdiffusion (2023)
• Genomics. DNA and RNA foundation models support long-range sequence modeling, functional annotation, regulatory and variant-effect prediction, and sequence design. Key papers: Brixi et al., Evo 2 (2026) · Avsec et al., AlphaGenome (2026)
• Virtual cells. Single-cell representations, multimodal foundation models, perturbation prediction, and models of cell state and response. Key papers: Cui et al., scGPT (2024) · Adduri et al., State (2025) · Noutahi et al., Virtual Cells (2025)

• Juno Nam, MIT. Machine learning for molecular simulation, with a focus on free-energy modeling and enhanced sampling. Key papers: Nam et al., LiFlow (2025) · Nam et al., Universal Thermodynamic Interatomic Potentials (2026)
• Minji Lee, Columbia University. Generative modeling for protein structure and design, including ConforNets and Genie 3. Key papers: Lee et al., ConforNets (2026) · Lin et al., Genie 3 (2026)

The Dec. 15 and 18 talks follow Korea University's class schedule.