[Paper Review] GenBench: A Benchmarking Suite for Systematic Evaluation of Genomic Foundation Models

Zicheng Liu, Jiahui Li+6 more|arXiv (Cornell University)|Jun 1, 2024

Cancer Genomics and Diagnostics4 citations

TL;DR

GenBench is a comprehensive, modular benchmarking suite designed to systematically evaluate genomic foundation models (GFMs) across short- and long-range genomic tasks, including coding regions, non-coding regions, and genome architecture. It reveals that attention-based models outperform convolution-based models in short-range tasks, while performance gaps narrow in long-range tasks, suggesting hybrid architectures may optimize efficiency and representational capacity.

ABSTRACT

The Genomic Foundation Model (GFM) paradigm is expected to facilitate the extraction of generalizable representations from massive genomic data, thereby enabling their application across a spectrum of downstream applications. Despite advancements, a lack of evaluation framework makes it difficult to ensure equitable assessment due to experimental settings, model intricacy, benchmark datasets, and reproducibility challenges. In the absence of standardization, comparative analyses risk becoming biased and unreliable. To surmount this impasse, we introduce GenBench, a comprehensive benchmarking suite specifically tailored for evaluating the efficacy of Genomic Foundation Models. GenBench offers a modular and expandable framework that encapsulates a variety of state-of-the-art methodologies. Through systematic evaluations of datasets spanning diverse biological domains with a particular emphasis on both short-range and long-range genomic tasks, firstly including the three most important DNA tasks covering Coding Region, Non-Coding Region, Genome Structure, etc. Moreover, We provide a nuanced analysis of the interplay between model architecture and dataset characteristics on task-specific performance. Our findings reveal an interesting observation: independent of the number of parameters, the discernible difference in preference between the attention-based and convolution-based models on short- and long-range tasks may provide insights into the future design of GFM.

Motivation & Objective

• To address the lack of standardized evaluation frameworks for genomic foundation models (GFMs), which hinders fair comparison due to inconsistent datasets, training strategies, and model complexity.
• To provide a unified, modular, and extensible benchmarking platform that supports diverse GFM architectures and realistic genomic tasks across multiple biological scales.
• To investigate the impact of model architecture—particularly attention vs. convolution—on performance across short-range and long-range genomic tasks.
• To enable systematic evaluation of GFMs on real-world datasets spanning 43 diverse genomic tasks, including promoter prediction, gene clustering, and species classification.
• To offer insights into optimal model design and pre-training strategies by analyzing computational efficiency, representation quality, and task-specific performance.

Proposed method

• GenBench integrates 43 real-world genomic datasets spanning three core domains: coding regions, non-coding regions, and genome architecture, to ensure broad biological relevance.
• The benchmark evaluates ten state-of-the-art GFMs, including both attention-based (e.g., DNABERT2, Nucleotide Transformer) and convolution-based (e.g., HyenaDNA, Caduceus) models, across standardized pre-training and fine-tuning protocols.
• A modular codebase is implemented to standardize data preprocessing, tokenization, and evaluation, enabling fair and reproducible comparisons across models and tasks.
• Model performance is assessed using task-specific metrics such as AUC, accuracy, and F1-score, with t-SNE visualizations used to analyze learned embeddings for gene clustering and species differentiation.
• Computational efficiency is evaluated via FLOPS estimation across varying input lengths to compare the scalability of attention and convolution-based architectures.
• The framework supports systematic ablation studies on pre-training strategies, including k-mer and BERT-based tokenization, to assess their impact on downstream performance.

Experimental results

Research questions

• RQ1 How do attention-based and convolution-based genomic foundation models compare in performance across short-range and long-range genomic tasks?
• RQ2 To what extent does input sequence length influence model performance, particularly in tasks involving extended gene context?
• RQ3 How do different pre-training strategies—such as k-mer and BERT-based tokenization—affect the biological interpretability and downstream accuracy of genomic embeddings?
• RQ4 What is the relationship between model architecture, computational cost (FLOPS), and representation quality in genomic representation learning?
• RQ5 Can a hybrid architecture combining attention and convolutional components improve both performance and efficiency in genomic foundation models?

Key findings

• Attention-based models achieve superior performance on short-range tasks such as promoter prediction, where capturing local nucleotide dependencies is critical.
• Convolution-based models demonstrate comparable performance to attention-based models on long-range tasks, with performance gaps narrowing as input length increases.
• The Nucleotide Transformer (NT) achieved the highest species classification accuracy (0.761) among evaluated models, indicating strong representation learning for inter-species differentiation.
• HyenaDNA, despite being attention-free, showed lower accuracy (0.655) and less distinct t-SNE clustering, suggesting weaker discriminative capacity for species-level gene sequence separation.
• Computational cost analysis revealed that simple CNNs have the lowest FLOPS, while attention-based models require significantly more computation, especially at longer sequence lengths.
• The results suggest that hybrid architectures combining convolutional inductive biases with attention mechanisms may offer an optimal balance between performance and efficiency in genomic modeling.