Leash Bio, a startup founded in 2021 by Ian Quigley and Andrew Blevins, is attempting to navigate the complex world of machine-learning-assisted drug discovery with a focus on ethical modeling. The article presents a metaphorical scenario where predicting which of 116 billion balls will adhere to a velcro strip mirrors challenges in small-molecule design. The point is clear: researchers often unconsciously introduce biases into their models, leading to results that may not be truly useful. This situation is exacerbated by long feedback loops in the industry, where flaws may not be discovered until it's too late, creating a tendency to overlook potential issues. Leash Bio aims to combat this trend by ensuring their models genuinely learn useful patterns rather than simply reflecting the biases of their training data. Their approach, while grounded in established scientific principles—utilizing large datasets of small-molecule and protein interactions—distinguishes them through their commitment to accuracy. They recently partnered with Monte Rosa Therapeutics to develop binders for novel targets, highlighting external validation of their efforts. The article also touches on a notable incident in the machine-learning community involving DiffDock, a small-molecule docking model. Criticism arose regarding its training and testing protocols, which could lead to overfitting. A subsequent update to DiffDock addressed these concerns, but it raised questions about chemical diversity in training datasets. The ongoing discussions in this space underscore the importance of rigorous methodologies in machine learning for drug discovery, a challenge that Leash Bio seems keenly aware of as they pursue their research.