Predicting aggregate morphology of sequence-defined macromolecules with recurrent neural networks

Abstract

Self-assembly of dilute sequence-defined macromolecules is a complex phenomenon in which the local arrangement of chemical moieties can lead to the formation of long-range structure. The dependence of this structure...

Data-Driven Design of Polymer-Based Biomaterials: High-throughput Simulation, Experimentation, and Machine Learning

“…Interestingly, long short-term memory networks (or similar recurrent neural network approaches) were not noted as remarkably deficient in our prior work 97 or that of others. 95,98 We suggest that the long short-term memory mechanism, which is useful for modeling sequence correlations, may struggle to internalize and maintain representations of the overall size of polymer sequences. By contrast, (graph) convolutional neural networks have pooling mechanisms that more readily retain such information.…”

“…In a similar vein, Bhattacharya et al have demonstrated the utility of gated recurrent units (another kind of recurrent neural network) in predicting the aggregate morphology of model sequencedefined polymers. 95 It is also natural to consider polymers as graphs wherein the graph vertices or nodes contain features of constitutional units, and edges indicate the connectivity among them. Mohapatra et al utilized such an approach with graph convolutional networks 119 to predict the immunogenicity of glycans and the target activity of antimicrobial peptides.…”

“…Having observed features relevant for achieving target values, the model could leverage this information, at least in some capacity, for designing new sequences with higher pattern complexity. Relatedly, Bhattacharya et al identified sequences with self-assembly behavior within the scope of that observed in their training data, 95 while Patra showed that sequence-based surrogate models struggle to predict property labels outside of the range observed in the training data. 134 Second, we note that the set of sequences constituting the training data were carefully constructed to cover all sequence variations up to four constitutional units.…”

“…For example, to study the properties of intrinsically disordered proteins, we used the one-bead-per-resiude model developed by Regy et al, which provides sufficient computational efficiency and accuracy to characterize the conformational behavior and relaxation dynamics of 2,585 intrinsically disordered proteins . Presently, most machine learning studies involving polymers predominantly use phenomenological models, − which enables data generation without wrestling with nuanced parametrization over a large chemical space. Using this approach, we created a simple model for single-chain nanoparticles and simulated nearly 2,000 distinct sequences altogether .…”

Polymer sequence design via molecular simulation-based active learning

“…We note that machine learning models are established to predict sequence-defined properties of polymers. 26,[38][39][40] These models are very powerful and can be used for high-throughput screening and the design of polymers. However, these models are built on a larger amount of pre-existing sequence-property data that are calculated using molecular simulations a priori.…”