We present progress in developing stable, scalable and transferable generative models for visual data. We first learn expressive image priors using autoregressive models which generate high-quality and diverse images. We then explore transfer learning to generalize visual representations models to new data modalities with limited available data. We propose two methods to generate high quality 3D graphics from sparse input images or natural language descriptions by distilling knowledge from pretrained discriminative vision models. We briefly summarize our work on improving generation quality with a Denoising Diffusion Probabilistic Model, and demonstrate how to transfer it to new modalities including high-quality text-to-3D synthesis using Score Distillation Sampling. Finally, we generate 2D vector graphics from text by optimizing a vector graphics renderer with knowledge distilled from a pretrained text-to-image diffusion model, without vector graphics data. Our models enable high-quality generation across many modalities, and continue to be broadly applied in subsequent work.

In drug discovery, where a model of the protein structure is known, molecular docking is a well-established approach for predictive modeling. Docking algorithms utilize a search strategy for exploring ligand poses within an active site and a scoring function for evaluating the poses. This dissertation explores improvements to both aspects of docking, emphasizing the use of machine learning methods for improving scoring functions. The work is built upon an extensible software platform for modeling molecular interactions, called Surflex.

Performance evaluation has been carried out on benchmarks that have been made publicly available, some of which were constructed in the course of this work. The novel tool pdbgrind, developed as part of the infrastructure for this work, was used to generate the large amount of data necessary to create adequate training and test sets. While the dissertation focuses most strongly on the scoring function problem in docking, some effort was also spent on the tightly coupled problem of search, and modest improvements were shown by enhancing Surflex's representation of protein active sites.

The bulk of the work describes improvements to empirical scoring functions for protein-ligand interactions. This dissertation demonstrates a robust method for tuning scoring function parameters to improve modeling of known binding phenomena. Penalties for inter-atomic overlap and same-charge repulsion were learned using synthetic negative data. The new function was shown to be equivalent or better than the original function in terms of screening utility on a large and diverse benchmark. This approach was generalized for the entire scoring function to support the use of multiple constraints in refining scoring function parameters. Using the constraint-based optimization procedure, users can exploit multiple types of data to customize functions to suit a particular task or a particular protein target or family of targets. Significant improvement to screening utility was shown using data typical of applications in docking.

The main contributions of this dissertation are generalizable methods for generating and exploiting multiple types of data in refining scoring functions for docking. The approaches can be extended to other areas, including quantitative structure activity prediction or protein folding.

Computational approaches for binding affinity prediction are most frequently demonstrated through cross-validation within a series of molecules or through performance shown on a blinded test set. Here, we show how such a system performs in two realistic applications: 1. An iterative, temporal lead optimization exercise, and 2. A hybrid strategy that leverages diversified information as input. In the first evaluation, a series of gyrase inhibitors with known synthetic order formed the set of molecules that could be selected for "synthesis." Beginning with a small number of molecules, based only on structures and activities, a model was constructed using the newly developed Surflex-Quantitative Modeling (QMOD) approach. Compound selection was done computationally, each time making five selections based on confident predictions of high activity and five selections based on a quantitative measure of three-dimensional structural novelty. Compound selection was followed by model refinement using the new data. Iterative computational candidate selection produced rapid improvements in selected compound activity, and incorporation of explicitly novel compounds uncovered much more diverse active inhibitors than strategies lacking active novelty selection.

For the second evaluation we present a hybrid structure-guided strategy that combines molecular similarity, docking, and multiple-instance learning such that information from protein structures can be used to inform models of structure-activity relationships. The Surflex-QMOD approach has been shown to produce accurate predictions of binding affinity by constructing an interpretable physical model of a binding site with no experimental binding site structural information. Here we introduce a methodological enhancement to integrate protein structure information into the model induction process in order to construct more robust physical models. The structure-guided models accurately predict binding affinities over a broad range of compounds while producing more accurate representations of the protein pockets and ligand binding modes. Structure-guidance for the QMOD method yielded significant performance improvements, especially in cases where predictions were made on ligands very different from those used for model induction.