Advanced Graduate Course · Computational Biophysics
Meta-Learning for Protein Simulation
MAML, memory kernels, equivariant neural potentials, and quantum-corrected dynamics — a unified treatment connecting machine learning methodology to the deep physics of non-Markovian protein dynamics.
About This Course
Classical molecular dynamics relies on force-field evaluations that are either physically approximate (empirical) or prohibitively expensive (ab initio). This course develops a unified framework in which meta-learning — learning to learn — serves simultaneously as a computational accelerator for force-field adaptation and as a physically interpretable model for non-Markovian protein dynamics.
The central theoretical result of the course is the formal correspondence between the Model-Agnostic Meta-Learning (MAML) inner-loop adaptation and the Nakajima–Zwanzig memory integral — identifying the meta-learning algorithm as a discrete approximation to a continuous physical operator. The cusp catastrophe provides the universality class for bifurcations in both the loss landscape and the protein conformational free-energy surface; ring-polymer MD on the MAML-adapted potential gives quantum-corrected rates; and the suppression of tunneling by geometric confinement becomes the experimental anchor connecting all four threads.
Key Numbers
12
Modules
~40
Equations
6–8 h
Study time
PhD
Level
~50
DFT configs to adapt (MAML-NequIP)
~10–100×
Less data than vanilla NequIP
Twelve Modules
M0
Introduction & Motivation
Three hard problems of protein simulation; how meta-learning + equivariant NNPs + RPMD address them. Learning objectives and prerequisites.
M1
MAML Mathematical Formulation
Bi-level optimisation, the second-order meta-gradient, the energy/force loss, and FOMAML / Reptile / ANIL variants.
M2
Equivariant NNPs — NequIP
E(3)-equivariant message passing on geometric tensors. Clebsch–Gordan products, Wigner D-matrices, and the data-efficiency advantage.
M3
MAML + NequIP Pipeline
Phase 0 meta-dataset construction, Phase 1 meta-training, Phase 2 task adaptation, Phase 3 QM/MM integration. Full Python code.
M4
Nakajima–Zwanzig Memory
Zwanzig projection operators, the generalised master equation, and the memory kernel \(\mathcal{K}(t-t')\). Markovian and non-Markovian limits.
M5
MAML as a Discrete Memory Kernel
The central theoretical result: MAML inner-loop adaptation as a Volterra approximation to the NZ memory integral. Term-by-term correspondence.
M6
Cusp Catastrophe Theory
Thom’s cusp normal form, fold and bifurcation sets, Hessian spectroscopy as bifurcation diagnostic on the MAML loss landscape.
M7
Quantum Tunneling & KIE
Bell and Wigner tunneling corrections, the kinetic isotope effect, WKB integrals, and the geometric-confinement suppression mechanism.
M8
Ring Polymer Molecular Dynamics
Path-integral ring polymers, the Bennett–Chandler procedure, i-PI input, bead-count temperature scaling.
M9
Geometric Confinement & Tunneling
The unified chain: rigidity → short NZ memory → shallow MAML → KIE → 1. Quantitative scaling and Hessian-spectroscopy diagnostic.
M10
Research Program & Open Problems
Five open problems from non-Gaussian MAML to NMQJ couplings. Connection to the JCP “Memory in Biomolecular Dynamics” special topic.
M11
References & Further Reading
Foundational papers (Finn 2017, Batzner 2022, Nakajima 1958, Zwanzig 1960, Thom 1972, Craig & Manolopoulos 2004) and advanced reading.
Cross-Links
Molecular Biology,Biochemistry,Quantum Mechanics,Statistical Mechanics,Machine Learning for Science,Numerical Methods,Probability & Statistics,Differential Geometry.