Advanced Circuits, Architecture, and Computing

Nihar R. Mahapatra | | www.egr.msu.ed/~nrm


Advanced Circuits and Architectures

In the past, semiconductor technology scaling and circuit and architecture innovations have delivered exponential performance growth within roughly fixed power and cost budgets. This has yielded wide ranging benefits across all sectors of the economy by making software increasingly powerful (to solve compute-intensive problems), easier to develop (through high-level languages and tools that enhance software productivity), and more cost effective (e.g., by enabling distributed cloud computing applications supported by advertisement revenue). However, extremely scaled technologies pose a number of challenges to this trend: the fraction of the transistor budget on a chip that can be actively engaged at any given time within fixed power and thermal constraints is decreasing in current architectures, devices are becoming more unreliable (because of process, voltage, and temperature variations and susceptibility to soft and hard errors), and communication and memory access are becoming increasing bottlenecks. Further, emerging applications demand greater security and are data intensive.

To address the above challenges, we have developed a number of advanced circuit- and architecture-level techniques for modeling and design optimization that exploit the fact that the behavior of computation, communication, and storage components on a chip are dependent on the data value and timing of signals they process. Examples of these include spatio-temporal energy and thermal modeling of on-chip interconnects, operand encoding and operation bypass techniques for computation components, static, dynamic, and adaptive encoding techniques for interconnects, and interconnect and memory system compression. Our current research is focused on cross-layer design techniques suitable for extremely scaled technologies, for modern computing infrastructure (sensors, portable edge devices, and cloud servers), and for meeting the demands of emerging big data and security-sensitive applications.

Advanced Computing

We have designed efficient sequential and scalable parallel branch-and-bound algorithms for solving hard discrete optimization problems. Our current efforts are centered on machine learning and optimization algorithms for big data and bioinformatics problems. We are also applying computational principles to natural and engineered systems (e.g., cyber-physical systems) to solve problems, develop design methods, facilitate automation, and advance understanding of their behavior.

Structure Based Drug Design

Drug design involves two main steps: first, the enzyme, receptor, or other protein responsible for a disease of interest is identified; second, a small molecule or ligand is found or designed that will bind to the target protein, modulate its behavior, and provide therapeutic benefit to the patient. The process of finding ligands that bind strongly to a macromolecular target involves screening millions of them. Determining the binding affinity (BA) of each ligand against a target protein in vitro and/or in vivo is a prohibitively expensive process and impractical for large databases even with the use of high-throughput screening approaches. As a result, in silico virtual-screening techniques are employed to filter large compound databases to manageable sets of most promising ligands. Molecular docking (see below) is a popular computational approach that “docks” a ligand into the 3D structure of macromolecular target and scores its potential complementarity to the binding site by predicting its BA.

Accurately predicting the BAs of large sets of protein-ligand complexes efficiently is a key challenge in computational biomolecular science, with applications in drug discovery, chemical biology, and structural biology. Since a scoring function (SF) is used to score, rank, and identify potential drug leads, the fidelity with which it predicts the affinity of a ligand candidate for a protein's binding site has a significant bearing on the accuracy of virtual screening. Despite intense efforts in developing conventional SFs, their limited accuracy has been a major roadblock toward cost-effective drug discovery. To address this problem, our research focuses on the development of highly accurate machine learning SFs in conjunction with a variety of physicochemical and geometrical features characterizing protein-ligand complexes for the ligand docking, screening, ranking, and scoring problems. Our efforts to date have yielded a machine learning SF with a predictive accuracy of 0.822 in terms of Pearson correlation coefficient between predicted and measured BA compared to 0.644 achieved by a state-of-the-art conventional SF on the core test set of PDBbind benchmark.