Introduces SelectBench, a comprehensive benchmark assessing the selectivity of generative structure-based drug design (SBDD) models. Evaluates nine representative models across literature case studies, safety panels, and large-scale target pairs to quantify off-target discrimination.

Introduces TasteBench, a multimodal benchmark and competition for sensory prediction across 21K+ human evaluations on 215 plant-based foods and 15K flavor molecules, providing computational proxies to accelerate sustainable food discovery.

Introduces IGLOO, a diagnostic framework that evaluates whether residue-level embedding gradients from protein language models faithfully capture functional importance against synthetic and experimental fitness landscapes.

Couples deep learning turnover number predictions with stochastic simulated annealing to parameterize enzyme-constrained genome-scale models across 93 organisms, resolving long-standing biocatalytic data sparsity.