MACE-CG Biomolecular Force Field

Simulating a membrane, a ribosome, or a lipid nanoparticle at atomic resolution is out of reach for all-atom molecular dynamics: the systems run to millions of atoms, and a trajectory needs millions of force evaluations. Coarse-grained force fields solve this by collapsing groups of atoms into beads, and machine-learned versions promise the accuracy of a quantum reference at a fraction of the cost. The catch has been transferability. Deep-learning coarse-grained potentials have almost always been fit to a single molecule or a single narrowly defined system class, because the coarse-graining map itself is usually hand-written — Cα-only mappings for proteins, tabulated bead types per amino acid — and does not extend to nucleic acids or lipids.

This model, developed at InstaDeep with BioNTech, applies one trained checkpoint across biosystem classes. Demonstrating transferability over proteins, RNA, lipids, and small molecules is the paper's stated core contribution, and it is what separates the model from prior coarse-grained potentials. It is also the first coarse-grained variant of the MACE architecture, the equivariant message-passing network behind all-atom potentials such as MACE-POLAR-1.

The coarse-graining step is deliberately not learned: atoms take priorities from the spectral decomposition of weighted molecular-graph Laplacians and are aggregated into beads by a rule-based algorithm, leaving the mapping a pure preprocessing step at both training and simulation time.

#Key Features

• One potential across biosystem classes: A single trained model runs molecular dynamics on protein, RNA, and lipid systems it has never seen, and on structures larger than any in its training set.
• Tunable graph-based coarse-graining: Four Laplacians — a binary bond graph plus edge weightings by inverse distance, nuclear charge, and bond polarity — are combined with four coefficients, tuned by gradient-free search to minimize the average magnitude of the reference coarse-grained forces.
• Continuous bead descriptors: Every bead carries a seven-dimensional vector of its mass and its counts of H, C, N, O, S, and P, replacing the discrete bead-type tables of classical force fields and letting one model span arbitrary organic chemistry.
• Fragmentation-based training data: A pipeline cuts large proteins, RNA, and lipid nanostructures into fragments, samples their conformations, and computes reference forces, producing training data at a scale no existing dataset offers.

#Technical Details

The force field is a MACE network modified to accept continuous multi-dimensional node features: bead descriptors pass through a learned MLP embedding rather than the discrete element lookup used for atoms. It uses hidden irreps of 128x0e + 128x1o, a maximum spherical-harmonic order of 3, two message-passing layers, and 8 Bessel radial basis functions, trained with AMSGrad at a batch size of 500.

Training data comprises 4,910,710 structures drawn from 7,170 unique fragments and molecules, split 80:10:10 with all conformers of a fragment kept in the same partition. Fragments were cut from large proteins, RNA, and lipid nanostructures with an initial 7 Å radius, up to a largest fragment of 297 atoms, and supplemented with PubChem molecules; conformations were sampled by molecular dynamics, by Gaussian perturbation of MD snapshots, and by stochastic conformer generation, with reference forces from the semi-empirical GFN1-xTB method and roughly 16% of structures filtered as physically unreasonable. Two variants were trained: a standard model using the plain binary bond graph, and a tuned model using the optimized Laplacian combination, selected at 212 and 101 epochs respectively by validation 95th-percentile force error. In 100 ps NVT Langevin simulations at 300 K with a 5 fs timestep, run through JAX-MD, both models produce stable trajectories for most protein, RNA, and lipid test systems taken from the PDB, though some break apart and phosphorus-oxygen bonds in RNA phosphates stretch visibly. On a 10 ns trajectory of the IPP lactotripeptide, TICA analysis recovers three conformational basins and an energy range comparable to the Martini reference.

#Applications

The target use is molecular dynamics on biological assemblies too large for all-atom simulation — membranes and lipid nanoparticles, nucleic acid structures, and protein complexes — where a single transferable potential removes the per-system parameterization that both classical and earlier machine-learned coarse-grained force fields demand. Including lipid nanostructure fragments in training points at formulation chemistry alongside structural biology.

#Impact

The work establishes that a transferable coarse-grained machine-learning force field is achievable despite the information loss the mapping introduces, and that a scheme with only four tunable parameters measurably improves downstream training. The authors are candid that the picture is incomplete: the reference method is semi-empirical rather than DFT, no solvated structures are included, the tuned model's training advantage does not carry over to simulation stability, and the comparison against Martini and an all-atom force field is qualitative. Neither the trained potential, the training code, nor the biosystems dataset has been publicly released, so the contribution stands as a demonstrated approach rather than a tool other groups can pick up.