ML Researcher

• Research models that combine molecular information, formulation and dose, assay metadata, video-derived behavior, and laboratory context
• Develop active-learning and sequential experiment-selection methods that balance predicted efficacy, uncertainty, novelty, and information value
• Define retrospective and prospective evaluations, including holdouts by chemical scaffold, laboratory, colony, and time
• Investigate which behavioral signals generalize across experiments and which reflect confounding, measurement noise, or laboratory-specific effects
• Translate model failures into new labels, assay variants, controls, or experiments that improve the next training cycle
• Communicate results with enough precision that experimental scientists can understand why a recommendation should or should not be trusted

• Ph.D. or equivalent research record in machine learning, statistics, computational science, or a closely related field
• Demonstrated ability to formulate open-ended research questions, build strong baselines, and design evaluations that survive distribution shift
• Strong software skills in Python and a modern machine-learning framework
• Experience working with noisy, limited, multimodal, or experimentally generated datasets
• Ability to move between theory, implementation, and scientific interpretation

Desired Attributes

• Experience with active learning, Bayesian optimization, reinforcement learning, causal inference, or scientific foundation models
• Experience in molecular discovery, biology, animal behavior, robotics, or another domain where models learn from physical experiments
• Track record of prospective validation rather than benchmark-only research
• Strong research taste and comfort abandoning an attractive idea when the evidence does not support it

We offer opportunities to do your life’s work while helping solve one of the most important technical and moral challenges of our time.

Full-time, in-office in Emeryville, California. Compensation includes equity.

Develop the learning methods that turn repeated assays into better scientific decisions. The central question is prospective: can a model use prior compound, assay, and behavior data to recommend an experiment that is more informative than the one scientists would otherwise run?