AI designs protein with applications for heart disease

• For the first time, AI has taken general instructions and generated a functional protein.
• Unlike AlphaFold or ESM-based methods, our AI designed a novel protein with just a programmable description.
• The AI-designed protein experimentally reacts to ATP, offering potential applications in treating heart disease, stroke, and more.

Instructing for ATP Reactivity

A key contributor to heart failure is cellular energy depletion, which impairs the heart’s ability to contract effectively. Cardiomyocytes are among the most metabolically active cells in the body, requiring a constant supply of ATP—the cell’s primary energy currency—to sustain function. During heart failure, ATP production is compromised, leading to rapid energy loss, reduced contractile force, arrhythmias, and ultimately, cardiac death.

To address this problem, we tasked our proprietary AI model with a focused objective: design a molecule capable of efficiently catalyzing ATP generation. The resulting protein bore substantial similarity (~70%) to naturally occurring adenylate kinases—enzymes known to be involved in ATP metabolism—but featured several key modifications to enhance stability, affinity, and performance. These differences represented a clear departure from native proteins, demonstrating how AI can explore a broader molecular space to engineer novel and optimized biological functions.

We verified these findings in the lab, demonstrating that this AI-designed protein could be experimentally generated and purified in multiple expression systems. Furthermore, preliminary results suggest an interaction with ATP, a crucial first step towards functionality. More experiments are underway to confirm that AI can create proteins that are both novel and functional, going beyond what evolution has produced.

Insights from this first-generation design have informed the next phase of our AI model development. By integrating structural and functional data from experimental testing, we are enhancing the model’s ability to design molecules for context-specific applications. This iterative feedback loop enables us to train the model on desired outcomes—ultimately paving the way for more sophisticated, task-specific molecules tailored to targeted disease treatment.

MP4 and the Future

MP4 (Molecule Programming V4), developed by 310 AI, is the first end-to-end generative protein model, creating protein sequences directly from text prompts to solve the "molecule programming" challenge—designing proteins for specific functions. MP4 is fully generative, trained on 138K tokens and 3.2B datapoints with 3,800 AMD-Instinct GPU-days.

MP5 will expand to complex non-protein molecules, emphasizing de novo designs, pharma-aware features, and multi-modal molecules.

Computational benchmarks and lab results are cited in white papers here and the GEM workshop paper. Check out all 107 AI designs in this repo.

Acknowledgements

This work was made possible by Trome, 310 AI’s proprietary platform for drug development.

We would also like to thank the AMD Instinct team and Arctoris Ltd for their generous time and dedication, which made the training and evaluation of MP4 possible.