Large language model (LLM)-based agents can increasingly support and empower human scientists in their work by reading literature, generating hypotheses, designing experiments, analyzing data, interpreting results, and drafting manuscripts. In parallel, robotics and programmable, cloud-connected laboratories have the potential to transform wet-lab operations into software-addressable infrastructure. Yet biological experiments face a persistent reality gap in which silent failures, temporal drift, contamination, sample mix-ups, and ambiguous protocol intent can invalidate an otherwise sound experimental plan. Closing this gap is essential for bringing the speed of modern computational discovery to experimental biology and requires audit-grade execution and verification. Here, we distinguish scripted automation from flexible, AI-enabled autonomy and outline a staged roadmap toward human-led and trustworthy “self-driving” labs for biology. We define laboratory autonomy in the context of a human-led scientific endeavor as a bounded control loop in which scientists set goals and constraints, agents help plan and coordinate experiments, instruments execute experiments and report machine-checkable evidence, and humans verify and interpret results and govern high-consequence decisions. We discuss where such systems are most appropriate and relevant today and where caution or deferral is needed. We map key lab-side potential failure modes and argue for verifiability-first autonomy. If developed responsibly, such systems could enable scientists to more efficiently decipher combinatorially large perturbation, design, and optimization spaces that are currently impractical to test experimentally, while empowering, rather than replacing, human creativity and judgment in experimental biological discovery.
A fundamental property of DNA is that the reverse complement (RC) of a sequence often carries identical biological meaning. However, state-of-the-art DNA language models frequently fail to capture this symmetry, producing inconsistent predictions for a sequence and its RC counterpart, which undermines their reliability. In this work, we introduce Reverse-Complement Consistency Regularization (RCCR), a simple and model-agnostic fine-tuning objective that directly penalizes the divergence between a model's prediction on a sequence and the aligned prediction on its reverse complement. We evaluate RCCR across three diverse backbones (Nucleotide Transformer, HyenaDNA, DNABERT-2) on a wide range of genomic tasks, including sequence classification, scalar regression, and profile prediction. Our experiments show that RCCR substantially improves RC robustness by dramatically reducing prediction flips and errors, all while maintaining or improving task accuracy compared to baselines such as RC data augmentation and test-time averaging. By integrating a key biological prior directly into the learning process, RCCR produces a single, intrinsically robust, and computationally efficient model fine-tuning recipe for diverse biology tasks.