RUBICON: A Framework for Designing Efficient Deep Learning-Based Genomic Basecallers

TL;DR

RUBICON addresses the bottlenecks of deep learning genomic basecalling by introducing a hardware-aware framework that couples quantization-aware neural architecture search (QABAS) with skip-connection removal (SkipClip). The combined approach yields RUBICALL, a hardware-optimized basecaller that delivers substantial throughput gains with maintained or improved accuracy, and superior downstream assembly and mapping performance. Empirical results across multiple datasets and platforms demonstrate up to multi-fold speedups and improved assembly quality, underscoring the importance of hardware-conscious design in genomics workloads. The framework is modular and extensible, offering a path to broader adoption of efficient basecallers across diverse hardware, species, and sequencing contexts.

Abstract

Nanopore sequencing generates noisy electrical signals that need to be converted into a standard string of DNA nucleotide bases using a computational step called basecalling. The accuracy and speed of basecalling have critical implications for all later steps in genome analysis. Many researchers adopt complex deep learning-based models to perform basecalling without considering the compute demands of such models, which leads to slow, inefficient, and memory-hungry basecallers. Therefore, there is a need to reduce the computation and memory cost of basecalling while maintaining accuracy. Our goal is to develop a comprehensive framework for creating deep learning-based basecallers that provide high efficiency and performance. We introduce RUBICON, a framework to develop hardware-optimized basecallers. RUBICON consists of two novel machine-learning techniques that are specifically designed for basecalling. First, we introduce the first quantization-aware basecalling neural architecture search (QABAS) framework to specialize the basecalling neural network architecture for a given hardware acceleration platform while jointly exploring and finding the best bit-width precision for each neural network layer. Second, we develop SkipClip, the first technique to remove the skip connections present in modern basecallers to greatly reduce resource and storage requirements without any loss in basecalling accuracy. We demonstrate the benefits of RUBICON by developing RUBICALL, the first hardware-optimized basecaller that performs fast and accurate basecalling. Compared to the fastest state-of-the-art basecaller, RUBICALL provides a 3.96x speedup with 2.97% higher accuracy. We show that RUBICON helps researchers develop hardware-optimized basecallers that are superior to expert-designed models.

Figures (20)

• Figure 1: Overview of RUBICON framework.
• Figure 2: Effect of pruning the elements and channels of Bonito_CTC using unstructured and structured pruning, respectively, on: (a) validation accuracy and (b) model size.
• Figure 3: Basecalling using quantized models.
• Figure 4: Effect of quantizing weight and activation of Bonito_CTC on model size. We quantize both the weight and activation with static precision. Since weights are the trainable parameters in a neural network, only weights contribute to the final model size.
• Figure 5: Comparison of average basecalling throughput for RUBICALL-MP with state-of-the-art basecallers in terms of: (a) average basecalling accuracy, (b) model parameters, and (c) model size. RUBICALL-MP provides higher compute performance with lower model size when compared to RUBICALL-FP because of the mixed-precision computation.