Protein-protein interactions are among today’s most exciting and promising targets for therapeutic intervention. To date, identifying small-molecules that selectively disrupt these interactions has proven particularly challenging for virtual screening tools, since these have typically been optimized to perform well on more “traditional” drug discovery targets. Here, we test the performance of the Rosetta energy function for identifying compounds that inhibit protein interactions, when these active compounds have been hidden amongst pools of “decoys.” Through this virtual screening benchmark, we gauge the effect of two recent enhancements to the functional form of the Rosetta energy function: the new “Talaris” update and the “pwSHO” solvation model. Finally, we conclude by developing and validating a new weight set that maximizes Rosetta’s ability to pick out the active compounds in this test set. Looking collectively over the course of these enhancements, we find a marked improvement in Rosetta’s ability to identify small-molecule inhibitors of protein-protein interactions.

No abstract

Protein design is challenging because it requires searching through a vast combinatorial space that is only sparsely functional. Self-supervised learning approaches offer the potential to navigate through this space more effectively and thereby accelerate protein engineering. We introduce a sequence denoising autoencoder (DAE) that learns the manifold of protein sequences from a large amount of potentially unlabelled proteins. This DAE is combined with a function predictor that guides sampling towards sequences with higher levels of desired functions. We train the sequence DAE on more than 20M unlabeled protein sequences spanning many evolutionarily diverse protein families and train the function predictor on approximately 0.5M sequences with known function labels. At test time, we sample from the model by iteratively denoising a sequence while exploiting the gradients from the function predictor. We present a few preliminary case studies of protein design that demonstrate the effectiveness of this proposed approach, which we refer to as "deep manifold sampling", including metal binding site addition, function-preserving diversification, and global fold change.

Antibody complementarity determining regions (CDRs) are loops within antibodies responsible for engaging antigens during the immune response and in antibody therapeutics and laboratory reagents. Since the 1980s, the conformations of the hypervariable CDRs have been structurally classified into a number of canonical conformations by Chothia, Lesk, Thornton, and others. In 2011 (North et al, J Mol Biol. 2011), we produced a quantitative clustering of approximately 300 structures of each CDR based on their length, a dihedral angle metric, and an affinity propagation algorithm. The data have been made available on our PyIgClassify website since 2015 and have been widely used in assigning conformational labels to antibodies in new structures and in molecular dynamics simulations. In the years since, it is has become apparent that many of the clusters are not canonical since they have not grown in size and still contain few sequences. Some clusters represent multiple conformations, given the assignment method we have used since 2015. Electron density calculations indicate that some clusters are due to misfitting of coordinates to electron density. In this work, we have performed a new statistical clustering of antibody CDR conformations. We used Electron Density in Atoms (EDIA, Meyder et al., 2017) to produce data sets with different levels of electron density validation. Clusters were chosen by their presence in high electron density cutoff data sets and with sufficient sequences (at least 10) across the entire PDB (no EDIA cutoff). About half of the North et al. clusters have been retired and 13 new clusters have been identified. We also include clustering of the H4 and L4 CDRs, otherwise known as the DE loop which connects strands D and E of the variable domain. The DE loop sometimes contacts antigens and affects the structure of neighboring CDR1 and CDR2 loops. The current database contains 6,486 PDB antibody entries. The new clustering will be useful in the analysis and development of new antibody structure prediction and design algorithms based on rapidly emerging techniques in deep learning. The new clustering data are available at http://dunbrack2.fccc.edu/PyIgClassify2.