OwkinZero

OwkinZero is a family of biological-reasoning large language models from Owkin, the French-American AI biotech, aimed at drug-discovery questions that general chatbots answer poorly. Rather than pretrain a model from scratch, the authors post-train open-source Qwen3 instruction models so that a fixed checkpoint can reason across a broad set of biomedical tasks, from judging whether a gene is a tractable drug target to predicting how a compound perturbs gene expression.

The central idea is Reinforcement Learning from Verifiable Rewards (RLVR): every training example is framed as a question with a checkable answer, so the model is rewarded only when its final answer is objectively correct. To supply that signal at scale, the team curated eight benchmark datasets totalling more than 300,000 verifiable question-answer pairs spanning spatial and bulk transcriptomics, pathway activity, drug perturbation, and target druggability. Training uses Group Relative Policy Optimization (GRPO), the same reward-optimization approach behind recent reasoning models, adapted here to biology.

Described in "OwkinZero: Accelerating Biological Discovery with AI" (Bigaud, Cabeli, Gürel, Pignet, Klein, Wainrib, Durand; arXiv:2508.16315, August 2025), the work sits alongside other domain-specialized biological reasoning models and argues that small, carefully post-trained open models can beat much larger commercial systems on well-scoped scientific tasks. OwkinZero is productized inside Owkin's K Pro co-pilot and was developed in collaboration with NVIDIA using the NeMo framework.

#Key Features

• Verifiable-reward post-training: OwkinZero applies RLVR via GRPO on tasks with checkable answers, rewarding the model for correct final predictions rather than imitating reference text, which sharpens biological reasoning.
• Eight curated benchmark tasks: Training and evaluation span spatial differential expression, tumour-versus-healthy expression, gene-indication features, TCGA signature activity, DrugSeq differential expression, drug-pathway perturbation, therapeutic target profiling, and structural druggability.
• Small models beating large ones: The 8B and 32B checkpoints outperform larger commercial and open baselines including GPT-4o, o3-Mini, DeepSeek-R1, and MedGemma-27B on the held-out biological benchmarks.
• Cross-task generalization: Specialist models trained on a single task improve performance on unseen tasks, while mixture-trained models generalize more broadly across the full benchmark suite.

#Technical Details

OwkinZero post-trains Qwen3-8B-Instruct and its 32B counterpart. The GRPO objective is modified with a per-batch, token-level normalization scheme (termed BNPO) to counter length bias, and reward is decomposed into formatting compliance, tag correctness, valid multiple-choice selection, and answer accuracy, trained for a single epoch over the curated pairs. On the eight-task suite, OwkinZero-8B reaches roughly 76.8% average accuracy versus baselines near the 55% range, and approaches 99.7% accuracy on the strongest individual tasks. The evaluation compares against o3-Mini, GPT-4o, DeepSeek-R1, MedGemma-27B, Qwen2.5-7B, Qwen3-8B, and Qwen3-32B, with the specialized models leading despite their relatively small size.

#Applications

OwkinZero targets early-stage drug discovery and translational research, where teams must repeatedly assess target druggability, choose a therapeutic modality, and anticipate how perturbations reshape gene expression and pathway activity. Delivered through Owkin's K Pro co-pilot, it lets scientists pose these questions in natural language and receive answers grounded in verifiable biological tasks, complementing transcriptomic and structural analyses that would otherwise require bespoke pipelines.

#Impact

OwkinZero contributes to a fast-growing line of domain-specialized biological reasoning models and provides concrete evidence that RLVR on curated, checkable biomedical data can lift compact open models above far larger general-purpose LLMs. Its finding that single-task training transfers to unseen tasks is a useful signal for how far verifiable-reward post-training generalizes in science. The work is a preprint awaiting peer review, and OwkinZero is offered as a component of a commercial product: model weights, training code, and a license have not been released publicly, and the benchmark results are in-silico evaluations rather than prospective experimental validation.