Session C04: Machine Learning for Biomolecular Design and Simulation

Live

Mitochondria are organelles of eucaryiotic cells involved in a number of physiological pathways. Cardiolipin (CL) is a phospholipid unique to the inner mitochondrial membrane. It plays a central role in mitochondrial functions and dynamics, and CL abnormalities are implicated in diseases. Our goal is to find compounds with high selectivity that can act as CL probes.
We explore the capabilities of using a coarse-grained (CG) model to find structures with certain properties. The 5-bead-type reduced Martini force field (T5) is a physics-based model that incorporates both the essential chemical features with a robust treatment of statistical mechanics. It simplifies the molecular representation through a small set of bead types that encode a variety of functional groups. This offers two advantages: first, many molecules map to the same CG representation and second, screening boils down to systematically varying among the set of CG bead types available. We have combined coarse-grained free energy calculations with deep representational learning and Bayesian optimization to efficiently screen the chemical space represented by all T5 compounds up to 400 Da molecular weight. The chemical-space exploration provides general design rules to further optimize selectivity over known CL probes.   

Live

The last years have seen an immense increase in high-throughput and high-resolution technologies for experimental observation as well as high-performance techniques to simulate molecular systems at a microscopic level, resulting in vast and ever-increasing amounts of high dimensional data. However, experiments provide only a partial view of macromolecular processes and are limited in their temporal and spatial resolution. On the other hand, atomistic simulations are still not able to sample the conformation space of large complexes, thus leaving significant gaps in our ability to study molecular processes at a biologically relevant scale. We present our efforts to bridge these gaps, by combining statistical physics with state-of-the-art machine-learning methods to design optimal coarse models for complex macromolecular systems. We derive simplified molecular models to reproduce the essential information contained both in microscopic simulation and experimental measurements.   

Live

In the sciences, computational chemists and physicists have been using ML for the prediction of physical phenomena, such as atomistic potential energy surfaces and reaction pathways. Transferable ML potentials, such as ANI-1x, have been developed with the goal of accurately simulating organic molecules containing the chemical elements H, C, N, and O. Here we provide an extension. The new model, dubbed ANI-2x, is trained to sulfur and halogens. These new features open a wide range of new applications within organic chemistry and drug development. To show that these additions do not sacrifice accuracy, we have tested this model across a range of organic molecules and applications, including dihedral rotations, conformer scoring, and non-bonded interactions. ANI-2x is shown to accurately predict molecular energies compared to DFT with a ~10⁶ factor speedup. A resulting model is a valuable tool for drug development that can potentially replace both quantum calculations and classical force fields for myriad applications.   

Live

Despite the importance of the molecular dipole polarizability in governing key intra- and inter-molecular interactions (such as induction and dispersion), determining the spectroscopic signatures of molecules, and being an essential ingredient in polarizable force fields, an accurate and computationally efficient prediction of this fundamental quantum mechanical response property still remains a challenge to date. In this work, we present a benchmark database [1] of highly accurate static dipole polarizability tensors of 7,211 small organic molecules computed using linear response coupled cluster singles and doubles theory (LR-CCSD). Using a symmetry-adapted machine-learning approach [2], we also demonstrate that it is possible to predict these LR-CCSD polarizabilities with an error that is an order of magnitude smaller than that of hybrid density functional theory (DFT). The resulting AlphaML model is robust and transferable, and able to yield molecular polarizabilities for a diverse set of 52 larger molecules (including challenging conjugated systems, carbohydrates, small drugs, amino acids, nucleobases, and hydrocarbon isomers) with a similar level of accuracy and at a negligible computational cost.
[1] Sci Data 6, 152 (2019).
[2] Proc Natl Acad Sci USA 116, 3401 (2019).   

Live

The free energy to pull apart stacked DNA bases is found to be much lower than classical simulations to date have suggested. Thermodynamic calculations are made in explicit water using a machine learning molecular dynamics method, trained on a novel dataset of quantum calculations making an advanced treatment of dispersion interactions. While the novel results contrast with previous classical simulations, they are consistent with values extrapolated down to the nanoscale from single molecule pulling experiments on large DNA double helices. The presented machine learned Hamiltonian for DNA is generally applicable to nucleic acids and is efficient to apply, therefore suggesting wide application in the future.   

Live

We study the spontaneous self-assembly of two single-stranded DNA (ssDNA) fragments using the coarse-grained oxDNA2 implementation [1]. Successful assembly is a rare event that requires crossing free energy barriers of several k_BT. To accurately determine different states and transition rates, we use trajectories from molecular dynamics simulations to construct a Markov state model. To this end, one needs one or more order parameters (OP) that faithfully describe the transition towards an assembled state. We formulate these OP based on structural information, which we map onto structural descriptors. Specifically, we investigate the latent space of EncoderMap [2] and how it changes with the amount of information contained in the descriptor.
With a proper OP, we investigate the stochastics of the self-assembly of two ssDNA molecules in detail.

[1]-Snodin et al., J. Chem. Phys.(2015), 142, 234901
[2]-T. Lemke and C. Peter,J.Chem.TheoryComput.(2019), 15, 1209-1215   

Live

Engineered proteins have emerged as novel diagnostics, therapeutics, and catalysts. Often, poor protein developability - quantified by expression, solubility, and stability - hinders commercialization. The ability to predict protein developability from amino acid sequence would reduce the experimental burden when selecting candidates. Recent advances in screening technologies enabled a high-throughput (HT) developability dataset for 10⁵ of 10²⁰ possible variants of protein scaffold Gp2. In this work, we evaluate the ability of neural networks to learn a developability representation from the HT dataset and transfer the knowledge to predict recombinant expression beyond the observed sequences. Mimicking protein theory, our model convolves learned amino acid properties to predict expression levels 42% closer to the experimental variance compared to a non-embedded control. Analysis of learned amino acid embeddings highlights the uniqueness of cysteine and the importance of hydrophobicity and charge, and unimportance of aromaticity, when aiming to improve developability. We identify clusters of similar sequences with increased developability through nonlinear dimensionality reduction (UMAP) and explore the inferred developability landscape via nested sampling.   

Live

A goal of physical modeling is to relate system-level phenomena to physical mechanisms. The complexity of biological systems makes it difficult to build models that balance parsimony and physical realism. The Manifold Boundary Approximation Method can derive simple models with limited scope from biological first-principles. However, these models may not reliably transfer since they abstract away the mechanisms irrelevant for their target context. I describe an approach to improve the transferablility of these reduced models. I consider the space of all possible reduced models. Given two minimal models, I construct the simplest model that can be reduced to them. This model is the least upper bound in complexity, and so we refer to it as the "supremum." By unifying the mechanistic explanations for different phenomena, the supremum model is predictive under diverse conditions. I illustrate for the Wnt signaling pathway. I build minimal models that describe Wnt signaling for two different developmental stages and construct their supremum. The supremum model predicts a new phenomenon: controlled pulsing, analogous to expression patterns of anterior-posterior axial development. The supremum principle is broadly applicable to create parsimonious, transferable models.   

Live

Prediction of protein functional properties from sequence is a central challenge that would allow us to discover new proteins with specific functionality. Experimental breakthroughs allow data on the relationship between sequence and function to be rapidly acquired that can be used to train and validate machine learning models that predict protein function directly from sequence. However, the cost and latency of wet-lab experiments require methods that find good sequences in few experimental rounds, where each round contains large batches of sequence designs. In this setting, I will discuss model-based optimization approaches that allow us to take advantage of sample inefficient methods and find diverse optimal sequence candidates for experimental evaluation. The potential of this approach is illustrated through the design and experimental validation of viable AAV capsid protein variants for gene therapy applications in addition to the design and validation of peptides as potential therapeutics.   

Live

A novel version of the Recurrent Geometrical Network (RGN¹) algorithm, which geometrically reasons over protein conformations, is used to predict protein structures. We use a transfer matrix formalism, which enables reasoning over protein backbones using a discrete version of the Frenet-Serret equations (dFSE) that leverages the fact that protein backbones are intrinsically discrete one-dimensional curves. dFSE-based RGNs are used with a context-based encoding of amino acid residues – AminoBERT – derived strictly from raw amino acid sequences without making explicit use of any evolutionary information. For building AminoBERT a reformulated version of the BERT language model is used to train a transformer over protein sequences to predict missing amino acids conditioned on the flanking sequence. Amino acid residues are thus mapped onto a higher-dimensional representation.

1. AlQuraishi, M. End-to-End Differentiable Learning of Protein Structure. Cell Syst. (2019) doi:10.1016/j.cels.2019.03.006.   

Live

In this work we employ multi-fidelity Bayesian optimization to fuse experimental and computational datastreams for the design of self-assembling π-conjugated peptides with emergent optoelectronic properties. We consider a family of peptides composed of a central π-core flanked by oligopeptide wings that have been demonstrated to self-assemble into supramolecular pseudo-1D nanoaggregates with emergent optoelectronic properties. Exhaustive traversal of the molecular design space of π-cores and peptide wings by either simulation or experiment is prohibitively expensive. This motivated the construction of a multi-fidelity Bayesian optimization platform to fuse cheap, high-volume, and approximate simulation data with expensive, low-volume, and accurate experimental data to rationally traverse the design space and efficiently identify molecules with engineered optoelectronic properties. New molecules identified by this active learning platform for experimental synthesis and testing yield superior optoelectronic properties compared to the best performing previous candidates.