Speaker
Description
Our objective is to develop an flexible in silico pipeline for designing small protein binders targeting an immune checkpoint receptor, aiming to overcome immune suppression and enhance cancer cell killing. Such protein design pipelines are complex, integrating multiple AI-based methods, including diffusion models, deep neural networks, and protein structure prediction tools.
We discuss the evaluation of existing protein design pipelines and the development of a customized workflow, supported by both human experts and the AI coding agent Codex. Particular attention is given to scientific, logical, and numerical errors identified in individual components of these pipelines. Some errors produce clearly incorrect results that compromise downstream predictions, while others introduce numerical inaccuracies without necessarily affecting the biological interpretation of the results. Distinguishing between these categories is essential for assessing the reliability of computational predictions.
We argue that scientific software, particularly AI-based tools, should undergo standardized quality assessments that extend beyond conventional software testing to include scientific validity, algorithmic correctness, and numerical reliability. AI coding agents are increasingly capable of identifying scientific and logical flaws in complex computational workflows and could play an important role in establishing reproducible, systematic quality control in computational biology.