Attribution assignment for deep-generative sequence models enables interpretability analysis using positive-only data.

GAMA enables interpretability of deep-generative models trained only on positive biological sequences, revealing key features in antibody-antigen interactions.

• Why it matters: Understanding how generative models make predictions is crucial for biological insights, especially when negative data are unavailable or unreliable, limiting traditional analysis methods.
• What they did: The authors developed GAMA, an attribution method based on Integrated Gradients, and validated it with synthetic data and real antibody-antigen binding datasets, focusing on positive-only training.
• The result: GAMA successfully uncovers biologically relevant features and supports sequence design strategies, advancing interpretability in biological modeling without requiring negative data.