(394d) Active Learning to Understand and Predict the Electronic Behavior of Peptides

Proteins play a key role in electron transport processes in biology, but the structure-function relationships governing the electronic properties of peptides are not fully understood. Our work aims to understand the connection between conformational flexibility, hierarchical structures, and electron transport behavior in peptides, which can be used to design and predict the electronic behavior of protein assemblies. In recent work, we studied the electronic properties of single peptides, and our results revealed a bimodal distribution molecular conductance across a relatively small set of amino acid sequences. MD simulations, Gaussian mixture modeling (GMM), and principal component analysis (PCA) were used to show that this two-state conductance behavior arises due to the conformational flexibility of peptide backbones. The high-conductance state arises due to a more defined secondary structure (beta turn or 3₁₀ helices), and the low-conductance state is associated with the primary amino acid sequences. In this talk, I will discuss a large-scale collaborative approach that combines active learning with experiments, theory, and simulations to understand and predict the electronic properties of 324 tetrapeptides (M-X₁-X₂-M) and 5832 pentapeptides (M-X₁-X₂-X₃-M) over a large chemical sequence space. The N- and C-terminal residues are selected as methionine to facilitate binding to gold electrodes during single-molecule electronics experiments, and X₁, X₂ and X₃ represent all possible natural amino acids except methionine and cysteine. High-throughput MD simulations are performed using Folding@home with a cumulative sampling time of ~4 ms and ~17 ms for all tetra- and pentapeptides, respectively. Based on MD results, a subset of single-molecule experiments is performed on sequences that are predicted to show drastic differences in secondary structure formation. Experimental results are used in combination with MD to develop an active learning model utilizing classical and physics-inspired deep architectures that accurately predict molecular conductance for the entire tetra- and pentapeptide chemical space. Broadly, our work will demonstrate the ability to predict the electronic properties of folded molecular architectures based on the monomer sequence of amino acids. Overall, this work will provide new insights into understanding the role of higher order assembled structures on biological charge transport, in addition to showcasing the predictive power of a combined modeling and experimental approach to molecular electronics.