Official Review · ICLR 2026

This paper introduces GENOME-FACTORY, a Python library designed to unify the workflow for genomic foundation models (GFMs). The framework integrates six components: (1) Genome Collector for data acquisition and preprocessing, (2) Model Loader supporting diverse genomic models (HyenaDNA, DNABERT-2, Caduceus, Nucleotide Transformer, EVO, GenomeOcean), (3) Model Trainer enabling full-parameter, LoRA, and adapter-based fine-tuning, (4) Inference Engine for embedding extraction and sequence generation, (5) Benchmarker providing evaluation on GUE and Genomic Benchmarks, and (6) Biological Interpreter using sparse autoencoders (SAE) for interpretability. The authors validate their framework across three dimensions: compatibility with diverse models, benchmarking performance, and biological interpretation using DNABERT-2. * Comprehensive Integration: The paper addresses a practical need by consolidating various components of the genomic model workflow into a single framework. The integration of data preprocessing, model training, inference, and evaluation is useful for practitioners. * Broad Model Support: Supporting six different genomic foundation models with varying architectures (attention-based, state-space models, generative models) demonstrates reasonable engineering effort and provides users with flexibility. * Multiple Fine-tuning Strategies: The inclusion of full-parameter, LoRA, and adapter-based tuning with detailed efficiency analysis (Table 2, Table 3) provides actionable insights about computational trade-offs. * User-Friendly Interfaces: Offering both CLI and WebUI lowers the barrier to entry for non-expert users, which could facilitate broader adoption in the biology community. * Systematic Benchmarking: The evaluation across multiple models and fine-tuning methods on two benchmark suites (GUE and Genomic Benchmarks) provides comprehensive performance comparisons. * Limited Research Novelty: The paper primarily presents an engineering integration effort rather than novel research contributions suitable for ICLR. While the authors claim to introduce "the first integrated Python library" for genomic models, this represents an engineering milestone rather than a fundamental research advance. Each component relies on existing techniques without significant innovation: data preprocessing employs standard bioinformatics operations, fine-tuning methods are direct applications of LoRA and adapters from NLP literature, and the interpretability module uses standard sparse autoencoders without genomics-specific innovations. The primary contribution is integrating these existing components into a unified interface, which, while potentially useful for practitioners, does not constitute the level of methodological or scientific novelty expected at ICLR. * Weak Biological Interpretation Validation: The biological interpretation component (Section 4.3), positioned as a key contribution and claimed to provide interpretability for genomic foundation mod