AI for Science

Large language models (LLMs) have shown growing promise in biomedical research, particularly for knowledge-driven interpretation tasks. However, their ability to reliably reason from gene-level knowledge to functional understanding, a core requirement for knowledge-enhanced cell atlas interpretation, remains largely underexplored. To address this gap, we introduce SciHorizon-GENE, a large-scale gene-centric benchmark constructed from authoritative biological databases. The benchmark integrates curated knowledge for over 190K human genes and comprises more than 540K questions covering diverse gene-to-function reasoning scenarios relevant to cell type annotation, functional interpretation, and mechanism-oriented analysis. Motivated by behavioral patterns observed in preliminary examinations, SciHorizon-GENE evaluates LLMs along four biologically critical perspectives: research attention sensitivity, hallucination tendency, answer completeness, and literature influence, explicitly targeting failure modes that limit the safe adoption of LLMs in biological interpretation pipelines. We systematically evaluate a wide range of state-of-the-art general-purpose and biomedical LLMs, revealing substantial heterogeneity in gene-level reasoning capabilities and persistent challenges in generating faithful, complete, and literature-grounded functional interpretations. Our benchmark establishes a systematic foundation for analyzing LLM behavior at the gene scale and offers insights for model selection and development, with direct relevance to knowledge-enhanced biological interpretation.

Foundation models trained on DNA sequences have achieved strong performance across biological tasks including variant effect prediction and genome design. These models rely on massive public genomic datasets comprising trillions of nucleotide tokens. Unlike natural language, DNA sequences lack semantic transparency, making corrupted or adversarially crafted entries difficult to detect during data curation. We present the first systematic study of training data poisoning in genomic language models, targeting both pre-training and fine-tuning stages. At pre-training, using Evo 2 and GENERator architectures, we show that less than 1% adversarially crafted sequences in the training corpus can selectively degrade generative performance on targeted genomic contexts while leaving unrelated sequences unaffected. We evaluate three scenarios: corruption of TATA-box promoter motifs, disruption of CTCF binding sites, and insertion of synthetic sequences absent from all training genomes. At fine-tuning, we demonstrate two additional attacks. First, poisoning a subset of CTCF sites in a ClinVar-derived corpus installs a conditional backdoor in a LoRA-adapted model that activates almost exclusively when the trigger sequence is present. Second, using frozen Evo 2 7B embeddings, targeted label corruption of downstream training data selectively compromises a clinically relevant variant classification task, demonstrated on BRCA1 variant effect prediction. These results show genomic foundation models are susceptible to targeted data poisoning with minimal footprint. We urge the field to adopt data provenance tracking, integrity verification, and adversarial robustness evaluation as standard components of the genomic model development pipeline.