Freecurve Labs
Publications
The key paper · October 2023
Illarionov et al. - Journal of the American Chemical Society - 2023
In referee terms, "the right answers for the right reasons."
Adding a neural-network interaction term finally makes agreement with QM accurate enough for predictive calculations across all systems, including ions.
We have an annotated list of more publications below.
In chronological order
Annotated List of Publications
These manuscripts are not just theoretical proofs. All these results were produced by parameterization stack(s), molecular dynamics stacks, and free energy stack(s) that we wrote. (In retrospect we would not have written the full MD stack. However, we have the expertise and capability to do so; and in current times such capability can only be matched by a handful of teams 10X the size).
Much of the work deals with solvation and hydration. These processes are much more than that; they demonstrate (up to sampling) a complete and faithful representation of both the energy (all relevant sub-components) and also the free energy (e.g. the entropy at a given temperature) of the molecular ensemble being modeled.
The 'benchmarks' of current molecular AI models are simply energy comparisons to quantum mechanics for sets of molecules. Their free energy results - when that capability is enabled - will be seriously off. In contrast, having training sets built from first principles has been our guiding principle for quite some time.
Comparisons with structural data-mining generative AI methods (i.e. Alpha-Fold) would take another white-paper and we are happy to discuss during further due diligence.
Whether it’s worthwhile to embed physics structure into an AI model is a long conversation involving Sutton’s bitter lesson. Briefly, we strongly believe that physics is still very necessary and will continue to be so for a decade. Why not use it if we have it?
Donchev et al. - Journal of Computational Chemistry - 2007
First analytical models (Levitt group); Quantum Mechanics - fitted, polarization, but non-transferable
AUTHOR'S NOTE: Relevant origin paper; not predictive
This is a predecessor paper (our analytical model director is a co-author). Michael Levitt supervised this group and he is a co-author on some of the publications. It utilizes physics-inspired (analytical) expressions to fit a (too economical) training set of quantum mechanics data. Polarization was used for transferability of models to bulk, but also to claim that interactions A <->B & A<->C would correctly determine A <->C (not true). While it did not achieve predictability nor transferability, it was a great start.
Pereyaslavets et al. - Proceedings of the National Academy of Sciences - 2018
Accurate alkane/water models; stack(s) foundation
AUTHOR'S NOTE: First accurate thermodynamic modeling
Formally this paper shows that a QM-faithful classical Hamiltonian produces significant errors unless light atom motion is treated quasi-classically. This is not as important for drug design, as the error partially cancels out.
Informally, and more importantly, this paper announces two things.
1. We have achieved (really in 2016) an accurate and descriptive energetic and thermodynamic representation of alkanes and water.
2. We have built much of our molecular dynamics (MD) and analytical model stacks
Accurate determination of solvation free energies of neutral organic compounds from first principles
Pereyaslavets et al. - Nature Communications - 2022
Solvation Free Energy (FE) prediction from dimers’ QM only
AUTHOR'S NOTE: First accurate first-principles modeling of (neutral) molecular ensembles
Aptly named, first and still only: Accurate determination of solvation free energies of neutral organic compounds from first principles. An analytical model, trained on dimer calculations only, is an accurate and descriptive thermodynamic representation of any ensemble of neutral molecules.
This paper deals with neutral molecules only because charged interactions did not work (and prompted the inclusion of the NN terms, next paper).
Chemistry magazine also published a less academic summary of this : https://communities.springernature.com/posts/molecular-dynamics-from-first-principles
This figure shows (blue) excellent agreement with experiment (also much better than (non-transferrable) models specifically fit to reproduce such results); along with the yellow dots that illustrate a point from the previous publication.
This figure shows that, indeed, dimer calculations can be used as training sets for ensembles of arbitrary size (as do the final results of course).
Nawrocki et al. - Journal of Chemical Theory and Computation - 2022
Analytical models insufficient for strong binding interactions
AUTHOR'S NOTE: Isolates model errors, prompts pivot toward NN correction
This paper definitely separates sampling issues from model issues in protein-ligand systems, and its real point is to conclude that strong intermolecular interactions cannot be sufficiently modeled by tractable analytical expressions.
Illarionov et al. - Journal of the American Chemical Society - 2023
NN-augmented models achieve predictive power (ions, ligands, etc.)
The key paper where everything comes together!
In referee terms, "the right answers for the right reasons."
Adding a neural-network interaction term finally makes agreement with QM accurate enough for predictive calculations across all systems, including ions.
AUTHOR'S NOTE: Predictive power across all systems
The NN correction targets the two-body interaction — the largest share of both total energy and modelling error (many-body terms are already sufficient). There is no practical way to reach this accuracy with analytical expressions: attempts stalled a decade ago at a water model, and any tractable analytics produce very large errors.
The diagram of the NN term (a minimal GNN, geometrically symmetrized, specific to a pair of atom types (i.e. aliphatic C <-> aromatic N; determined by local intermolecular environment, 'book-kept' by a database (DB). (Bottom) An illustration of the generation of the fingerprints and the NN structure for interaction of two water molecules.
The pair-specific (i.e. property of the pair) NN term is not decomposable into single atom properties (the analytical term is - essentially the analytical term(s) extract all the 'nice' 'understandable' [e.g. we even give them names: Van der Waals / dispersion] properties and short, (intermediate) and long-range behavior). Another key technique here is the explicit decoupling of the non-bonded (intermolecular) and (much stronger) bonded (intramolecular) interactions.
The final results: electrolyte / ionic systems, neutral molecular ensembles (ofc) and protein-ligand interactions (literally very complex) are modelled and predicted accurately, with the models created from dimer QM calculations only.
Accuracy is essentially identical to the QM calculations (here, 'gold-standard' coupled-cluster). A possible exception is very large aromatic systems, where QM itself sometimes struggles. The models run at ~0.5× the speed of a normal polarizable analytical model — whereas atomistic AI models are ~1000× slower, digesting a large neighbourhood for every atom at every step.
Butin et al. - Journal of Chemical Theory and Computation - 2024
Predicts pH of water, settles speciation debate
AUTHOR'S NOTE: Demo of full free energy modeling for catalysis and ionic species. Somewhat of a victory lap from the 2023 paper.
This paper uses the models to predict the PH of water from first principles and without any approximations, and it does so correctly. It also settles the debate on which ionic species of water is prevalent at temperature. This is a warm-up for simulating and analyzing catalytic and enzymatic reactions with full explicit solvent.
Current R&D is subordinate to generating and testing models (protein-protein interactions next), refactoring, streamlining the code stack(s), templatizing common in-silico tasks, and servicing customers, and producing POC studies for prospective clients (e.g. antibody maturation benchmarks), and fundraising.
While energetic accuracy has largely been solved, the remaining barrier to predictive modeling is ensuring that the system explores relevant conformational states — i.e., sampling; and that is the aim of our current R&D efforts. We are working to achieve routine 0.5 µs / day simulation times with explicit solvent. We employ equivariant graph networks and flow-based NN's. Attempts to do this (Nature 2019) have stalled but we have understood the main obstacles and have devised several multi-scale techniques to bypass them; this project is currently going well.