papers_r = FileAttachment("../data/all_papers.json").json()
n_columns = Object.entries(papers_r["doi"]).length
index_array = Array.from({length: n_columns}, (_, i) => i)
papers = index_array.map(
i => Object.fromEntries(
Object.entries(papers_r).map(([column, values]) => [column, values[i]])
)
)
function displayTable(papers_) {
return Inputs.table(papers_, {
columns: [
"doi",
"title",
"journal.name",
"journal.volume",
"year",
],
header: {
"journal.name": "journal",
"journal.volume": "volume",
},
format: {
doi: doi => htl.html`<a href=https://doi.org/${doi} target=_blank>${doi}</a>`,
title: title => htl.html`${title}`,
"journal.volume": volume => htl.html`<b>${volume}</b>`,
year: year => htl.html`${year}`,
},
width: {
"title": 10
},
layout: "auto",
})
}
// Every section below selects its papers this way. The `topics` field is
// attached at export time from `data/topics.yml`, which is also where a new
// paper gets assigned to a section -- see `scripts/topics.py`. Selecting by
// coauthor instead, as this page used to, silently adopts whatever that person
// publishes next; selecting by a DOI list written into the page means the page
// has to be remembered every time a paper comes out, and it was not.
function topic(key) {
return displayTable(papers.filter(p => p.topics.includes(key)))
}Research topics
Research topics of the group: free energies from generative models, measure transport across resolutions, high-throughput screening of thermodynamic properties, molecular discovery by active learning, and coarse-graining across chemical space.
Overview
Our research sits at the interface of statistical mechanics, molecular simulation, and machine learning. We develop multiscale methods for soft-condensed matter and biomolecules, and we apply them to problems in which two landscapes compound one another: the configurational landscape of a single molecular system at fixed composition, and the compositional landscape of chemical compound space, in which the identity of the molecule is itself the variable. Sampling either is expensive. Having to sample both at once is what makes molecular discovery hard.
Coarse-graining is our main instrument against both. A reduced representation smooths the free-energy landscape and accelerates sampling; and because a small set of bead types maps many molecules of similar shape and hydrophobicity onto the same parametrization, it also contracts chemical space by close to an order of magnitude. That many-to-one mapping is what makes high-throughput screening of thermodynamic properties tractable, and what turns structure–property relationships into design rules. It also costs chemical specificity, and a good deal of our methodological work is concerned with measuring that price and with recovering what the reduction discards.
Machine learning enters as a way of tailoring models to physics rather than substituting for it. The recurring move is to take a known analytical result—the Boltzmann distribution, a Fokker–Planck steady state, thermodynamic integration—and impose it as an inductive bias on a generative model, instead of letting the model infer that structure from data alone.
For an unsupervised view of the same corpus, clustered by abstract rather than by hand, see the research landscape. Much of what follows is released as open-source software, and all of it is the work of the group.
Current directions
Free energies from generative models
Free-energy calculations remain among the most expensive routine tasks in molecular simulation. We reformulate them as learning problems in which the physics supplies the structure of the estimator. Neural thermodynamic integration parametrizes the interpolating potential along the alchemical pathway with a neural network, so that a single learned model stands in for the ladder of intermediate states a conventional calculation has to simulate; it is demonstrated on the excess chemical potential of a Lennard-Jones fluid. The construction extends to solvation, where both end states are non-trivial, by geodesic interpolation between their samples. Fokker–Planck score learning arrives at the same goal from a different constraint: the analytical steady-state solution for a particle in a periodic potential under constant force is imposed on a neural score function, which recovers potentials of mean force efficiently and is most advantageous in the low-data regime.
Fokker-Planck score learning Neural TI
Measure transport across resolutions
Going from atomistic to coarse-grained resolution is straightforward; going back is not. We treat the inverse—backmapping—as a probabilistic transport problem rather than a regression. Conditional generative adversarial networks reconstruct equilibrated atomistic structures autoregressively, bond by bond, sampling the atomistic Boltzmann distribution without pre-equilibration, with remarkable transferability from melt to crystalline phases and, in later work, across chemistries. Split-Flows recasts the problem as continuous-time measure transport: augmenting a coarse-grained configuration with auxiliary noise dimensions renders the map to the fine-grained space one-to-one, and a flow matched between the two resolutions supplies both a generative backmapping and, for the first time tractably and for an arbitrary coarse-graining map, the mapping entropy that quantifies how much information the reduction discards.
High-throughput screening of thermodynamic properties
Rather than arbitrarily selecting compounds, we systematically explore chemical space according to the target property of interest. Coarse-graining is what makes this affordable: a linear relation between bulk partitioning and the potential of mean force turns an expensive insertion free energy into a surrogate that can be evaluated across an entire library. We used it to screen more than 500,000 small molecules for drug–membrane permeability, yielding a permeability map across a sizeable part of organic chemical space, benchmarked against experiment and released as a public dataset of coarse-grained trajectories, insertion PMFs, and permeability coefficients. Symbolic regression over that dataset recovers compact algebraic expressions in a handful of physical descriptors—size, polarity, hydrogen-bonding capacity—and the century-old Meyer–Overton correlation emerges from hydrophobic transfer and bilayer deformation alone, without invoking receptor-specific binding.
Molecular discovery by active learning and Bayesian optimization
Screening is only half of discovery; the other half is deciding what to compute next. We close that loop with active learning and Bayesian optimization on top of rigorous coarse-grained free-energy calculations. Applied to cardiolipin, iterative cycles inside a low-dimensional chemical-space embedding identified small molecules predicted to bind the mitochondrial lipid selectively over structurally similar phospholipids, and the resulting physicochemical design rules were then used to filter a chemical-vendor database: of twenty compounds tested, three proved selective in vitro and one in vivo. Multilevel Bayesian optimization organizes the same idea as a funnel across resolutions, exploring broadly where the model is cheap and exploiting where it is accurate. The approach reaches beyond small molecules—to gas-separation polymers that exceed the CO2/CH4 Robeson upper bound and were subsequently synthesized and measured, and to peptide sequences whose learned embeddings support de novo design.
Coarse-graining as a reduction of chemical space
Transferable coarse-grained models are useful precisely because they are degenerate, and we study the consequences of that degeneracy quantitatively. Varying the number of Martini bead types from 5 to 16 across thousands of molecules establishes the resolution limit of data-driven models spanning chemical space, and marks where transferability begins to cost chemical specificity. Treating exploration as importance sampling over coarse-grained representations improves coverage of the property landscape by factors of 2–10 over uniform sampling, at the scale of some \(10^6\) compounds. Adjacent questions concern how condensed-phase structure ought to be represented at all—adapting the SLATM expansion in one-, two-, and three-body local environments to thermodynamic observables of molecular liquids—and how far kernel-based models can be trusted, where an audit of machine-learned hydration free energies exposes the penalty incurred when the training database is chemically narrow.
Automated force-field parametrization
High-throughput screening is only possible if parametrization requires no human in the loop. Auto-Martini generates Martini topologies for small organic molecules directly from SMILES, assigning the mapping, bead types, and bonded parameters automatically. The pipeline has been rebuilt for Martini 3, validated against expert-designed models, and exercised on a screen of 100,000 compounds from the Enamine library.
Reproducible and FAIR simulation workflows
Screening campaigns generate large numbers of simulations whose provenance is easy to lose and whose results are easy to recompute by accident. Martignac represents Martini simulations as directed acyclic graphs with full provenance tracking, connected to the NOMAD database, so that a prior calculation is retrieved rather than repeated. The framework is the practical end of a longer commitment to FAIR data in materials science—community metadata schemas spanning electronic structure, sampling, and workflows, and the case that the transition to data-centric materials science requires shared infrastructure and stewardship rather than algorithms alone.
Continuing threads
Membrane biophysics
Lipid membranes are the physical system our methods return to most often, and the question we most often ask of them is which of their features are thermodynamically dictated and which demand active cellular maintenance. Semi-grand canonical simulations, with chemical potentials imposed on individual lipid species, reframe membrane composition as a constrained-equilibrium problem: acyl-tail saturation homeostasis in ectotherms and saturation asymmetry between leaflets both emerge from thermodynamics alone, narrowing the list of properties a cell must actively regulate. In parallel we study how small molecules partition into membranes and reshape them—phase separation in binary mixtures, nanoparticle contact mediated by copolymer coronas, the bacteriostatic action of 2-phenylethanol derivatives, the reorganization produced by nitrated fatty acids—together with the folding and insertion thermodynamics of transmembrane peptides.
Kinetics of coarse-grained models
Coarse-graining accelerates dynamics, and it does so unevenly: a model parametrized against structure alone will generally misrepresent barrier-crossing kinetics. We parametrize concurrently against static and kinetic information, using Markov state models biased with external information to keep the interpretation of simulation kinetics consistent. Conformational surface hopping couples distinct force fields to distinct conformational basins, in analogy with nonadiabatic transitions between electronic surfaces; incorporating more surfaces improves cross-correlations systematically and recovers barrier-crossing dynamics that a single surface cannot. Related work identifies metastable states without hand-chosen collective variables, through hidden Markov models of solvation substates and Gaussian-mixture variational autoencoders whose mixture components correspond to conformational basins.
Non-equilibrium dynamical reweighting
We extend the concept of statistical reweighting to non-equilibrium steady states. Stochastic thermodynamics analytically relates the forward and backward probabilities of any pathway through the external nonconservative force, which supplies the constraint; a maximum-caliber—maximum path entropy—formalism then reweights microtrajectories from one driven steady state onto another. The framework extends to collective variables, so that free-energy landscapes along a reaction coordinate can be recovered from short driven runs.
Polymer and organic-materials morphology
Multiscale simulation of polymer morphology, where the observable of interest sits beyond the reach of either resolution on its own. Metadynamics in a phase space of collective variables, referenced against DFT, resolves the free-energy landscape between polymorphs of syndiotactic polystyrene and yields \(\alpha/\beta\) free-energy differences that rationalize the experimentally observed phase populations.
Foundations
Earlier lines of work, largely concluded, that the current programme grew out of.
Static atomic multipole electrostatics
Point charges are a crude description of the molecular electrostatic potential. Static atomic multipoles, derived from the electrostatic potential through a spherical-harmonic expansion, provide a systematically better one, and exploiting their symmetries makes force propagation affordable enough for condensed-phase molecular dynamics. Applications include free energies of hydration, two-dimensional infrared spectroscopy of solvation dynamics, and the \(\sigma\)-hole of halogenated compounds, where quadrupolar terms on the adjacent atoms prove necessary.
Machine learning of non-covalent interactions
Physics-based potentials with machine-learned coefficients, combining atomic multipoles and many-body dispersion for accurate and transferable non-covalent interactions. The coefficients are learned by kernel ridge regression across both conformations and composition of small organic molecules; a Voronoi-tessellation route to van der Waals coefficients dispenses with explicit electron densities altogether.
Peptide coarse-graining and folding cooperativity
A top-down, implicit-solvent coarse-grained peptide model, used to study secondary and tertiary structure formation across environments and scenarios: \(\alpha\)-helix against \(\beta\)-sheet folding, the microcanonical analysis that locates the origin of folding cooperativity in the interplay between the two, the structural alignment of capsid interfaces, and peptide–membrane interactions from a simple cross-parametrization.
Early methodological work
Assorted early work on estimators and integration schemes: optimized convergence for multiple-histogram analysis, a multi-timestep integrator for the modified Andersen barostat, and a simulated molecular walker.