Next-Generation Adaptable Computing for Omics

Abstract: Genome analysis is the foundation of many scientific and medical discoveries as well as a key pillar of personalized medicine. After an individual's genome is sequenced, many computational steps are taken to reconstruct and analyze the genome. Unfortunately, these computational tasks are often very slow and energy hungry, in many cases requiring very expensive computational resources. As a result, even though sequencing technology improvements have enabled high-throughput and portable sequencing devices, like nanopore sequencers, interesting and potentially critical analyses still take days or even weeks and also cannot be performed in portable devices. This talk describes our ongoing journey in greatly improving the performance and energy efficiency of genome analysis (with a focus on at least two major issues in genome analysis, i.e., read mapping and metagenomics analysis). We show that significant improvements are possible with both algorithmic and hardware-based approaches and their combination. We conclude with a foreshadowing of future challenges brought about by very low-cost new sequencing technologies and their potential use cases in public health, science, and medicine. A short accompanying paper, which appeared at DAC 2023, can be found here and serves as recommended reading: "Accelerating Genome Analysis via Algorithm-Architecture Co-Design" https://arxiv.org/abs/2305.00492 A longer overview & survey of modern genome analysis and how to make it intelligent and efficient can be found here and also serves as recommended reading: "Going From Molecules to Genomic Variations to Scientific Discovery: Intelligent Algorithms and Architectures for Intelligent Genome Analysis" https://arxiv.org/abs/2205.07957

Abstract: Regex Matching (REM) represents a crucial computational kernel in many scenarios, from networking to proteomics, from databases to natural language processing. However, it carries two intrinsic disadvantages: it is embarrassingly sequential and there is no real consensus on which is the optimal engine for regex execution. This talk will explore the research direction of co-designing a unified architectural solution to the longstanding performance–flexibility trade-off.

Abstract:Analyzing genomic data provides critical insights for many important applications, including understanding and treating diseases, personalized medicine, and outbreak tracing. Modern genome sequencing technologies, such as nanopore sequencing, can rapidly generate large volumes of genomic data at low cost. A standard first step in genomic data analysis is to translate raw sequencing data (e.g., electrical signals) into DNA letters using complex AI models. However, this translation step introduces significant computational overhead that limits the performance, scalability, and portability of genome analysis pipelines, especially on resource-constrained devices. This talk focuses on designing new techniques to address these computational limitations. We discuss how we can take a fundamentally different approach that directly analyzes electrical signals generated by nanopore sequencing devices, without translating them into DNA characters. By efficiently processing these signals, for the first time, we enable real-time analysis of human genomes as sequencing data are generated. Building on our prior work, we explore new genomic applications enabled by the direct analysis of nanopore electrical signals. We show that genomes can be constructed directly from electrical signals instead of DNA letters. We conclude by discussing how such approaches open the door to genomic data analysis that can be performed anywhere and any time, enabling new opportunities in medicine and genomics.

Abstract:Scaling AI for biology requires efficiency not just in hardware, but in the underlying algorithms. In this talk, I present a unified perspective on improving efficiency across the lifecycle of omics AI, spanning data acquisition, modeling, and interpretation. I discuss approaches that leverage structure and amortization to reduce data requirements, accelerate learning, and enable scalable interpretability. This perspective reveals common principles that cut across traditionally separate problems and suggests new directions for co-design between machine learning algorithms and adaptable computing systems. The result is a pathway toward more efficient, scalable, and broadly deployable AI for genomics and proteomics.