Sparks: Multi-Agent Artificial Intelligence Model Discovers Protein Design Principles
This addresses the challenge of enabling AI to independently conduct rigorous scientific inquiry and identify new principles, which is significant for researchers in protein science and AI-driven discovery, though it appears incremental as it builds on existing multi-agent and generative methods.
The paper tackles the problem of autonomous scientific discovery by presenting Sparks, a multi-agent AI model that executes the entire discovery cycle without human intervention, and applied to protein science, it uncovered two previously unknown phenomena, including a length-dependent mechanical crossover and a chain-length/secondary-structure stability map.
Advances in artificial intelligence (AI) promise autonomous discovery, yet most systems still resurface knowledge latent in their training data. We present Sparks, a multi-modal multi-agent AI model that executes the entire discovery cycle that includes hypothesis generation, experiment design and iterative refinement to develop generalizable principles and a report without human intervention. Applied to protein science, Sparks uncovered two previously unknown phenomena: (i) a length-dependent mechanical crossover whereby beta-sheet-biased peptides surpass alpha-helical ones in unfolding force beyond ~80 residues, establishing a new design principle for peptide mechanics; and (ii) a chain-length/secondary-structure stability map revealing unexpectedly robust beta-sheet-rich architectures and a "frustration zone" of high variance in mixed alpha/beta folds. These findings emerged from fully self-directed reasoning cycles that combined generative sequence design, high-accuracy structure prediction and physics-aware property models, with paired generation-and-reflection agents enforcing self-correction and reproducibility. The key result is that Sparks can independently conduct rigorous scientific inquiry and identify previously unknown scientific principles.