Simulation surrogate in structural biology and protein science

How it is used

Structural biology works out the shapes of proteins and how those shapes do chemical work. A large part of the field relies on molecular dynamics: software that computes the forces between atoms and steps them forward in tiny increments to watch a molecule move. The examples here train neural networks on the output of such runs so the network can supply the answer the simulation would have given. In one, a model learns to generate protein backbone shapes from a long all-atom trajectory. In others, simulations supply labelled training data, and the network learns to predict a property straight from the amino acid sequence, either through engineered descriptors of that sequence or through a language model pretrained on protein sequences.

Why researchers use it

The appeal is arithmetic. A single simulation can run for a long stretch of simulated time to sample one molecule properly, and the quantity a researcher wants, an average shape, a free energy, a measure of how a chain swells or sticks, only emerges once enough of that sampling has accumulated. Designing peptides or screening thousands of sequences means paying that cost again for each candidate. A trained stand-in answers in a fraction of the time, so sequences can be ranked in bulk and only the interesting ones simulated properly. The examples concentrate on intrinsically disordered proteins, which have no single fixed shape and so demand especially heavy sampling.

In the record

structural-biologyA compact neural network learns to generate shape-shifting protein structuresA compact model trained on a long all-atom trajectory generated backbone conformations, using graphs of nitrogen, alpha-carbon and carbon atoms with orientation-aware edge features.structural-biologyPeptides designed by simulation and machine learning sit at condensate surfacesA small two-layer network trained on coarse-grained simulation outputs predicted interface-partitioning free energies and a virial coefficient from engineered sequence descriptors.structural-biologyLanguage model trained on protein sequences predicts shape and stability of disordered proteinsA protein language model, pretrained on sequences and fitted with a regression head, predicted properties labelled by coarse-grained simulations of disordered protein sequences.

Limits

A stand-in can only be as good as the simulations that taught it, and those rest on approximate descriptions of atomic forces, coarser still when groups of atoms are lumped together. Errors in the underlying physics pass straight through, invisibly. The models are also narrow: trained on particular proteins or a particular pool of sequences, they may fare poorly on sequences unlike those, and they rarely signal when they are out of their depth. Averaged properties can be reproduced while rare but important shapes are missed.

Written by AIxSci: the background from general knowledge, the examples from checked records.