Introduction

Reporter genes for Magnetic Resonance Imaging (MRI) would allow genes to be measured in a non-invasive and safe manner, paving the way for translational use in human subjects. One of the best mechanisms for generating such a reporter gene is Chemical Exchange Saturation Transfer (CEST), which generates contrast without addition of outside metals and allows for the contrast mechanism to be controlled. Previous CEST reporter genes1,2can be further optimized via protein evolution but have proven difficult to improve using typical protein optimization engineering methods. We hypothesized that these limitations could be overcome via application of machine learning. To achieve this goal, we utilized Genetic Programming (GP) which is a field in computer science studying algorithms inspired by biological evolution. GP is a powerful technique that evolves models to difficult tasks and has been used in the past to in array of biological studies. In this study we sought to develop contrast which was generated at two different saturation frequencies. The first experiments were used to optimize contrast at 3.6ppm downfield from water, since this is the saturation frequency currently used in research due to the plentiful amines present in amino acids. Later experiments were used to optimize contrast at 5.0 ppm downfield from water, since there is less background saturation at this frequency, and it demonstrates the ability for POET to develop a novel function less well understood in research.

Methods

We have thusfar performed 10 generations of development for peptides with a CEST peak at 3.6 ppm, and 6 generations of development for peptides with a CEST peak at 5.0 ppm. For the 3.6 ppm peptides we started with training data obtained from prior literature values of 12 amino acid long peptides. For the 5 ppm peptides we used the 5ppm data gathered from the first four generations of the 3.6 evolution. In each generation of development, POET was used to predict 10 peptides, which were then synthesized by a vendor. We dissolved the peptides to a concentration of 5 mg/mL in PBS and were imaged on 7T preclinical MRI at 37 C, TR 10 s, Tsat 4s, Bsat 4.7 μT with conditions as described before3. The experimental results would then be added into the training data and the genetic program would then develop predictive models and predict proteins based on the new information learnable from the increase in data. Exchange rate (Kex) was determined using QUEST4. Each sample was scanned alongside a sample of Poly-L-Lysine (PLL), which served as a control and standard. PLL was chosen due to its relation to the Lysine Rich Protein (LRP) which is the current state of the art CEST reporter gene1.

Results

The first peptides developed using POET were mostly insoluble, with only one peptide which was able to dissolve and produce contrast. The next two Later generations showed steadily increased contrast. Later generations level out in terms of median and maximum contrast, however, Across all generation, 42 peptides were discovered that produce greater contrast than PLL (figure 1). More importantly, several peptides that produce contrast that is substantially higher than PLL were identified. When the exchange rate (Kex) – the hallmark of a good CEST agent) was measure using the QUEST method, some selected peptides show Kex as high as 822 Hz which is ~2 fold of PLL (Kex=479).
Next, we tested if POET can be used to develop a new function which is generating contrast at 5 ppm. The advantage of contrast at 5 ppm is that there is a very low background contrast from physiological molecules in comparison to the conventional contrast at 1.8 or 3.6 ppm. The first peptides developed for 5ppm showed greater contrast on average than the training data. Later generations have shown less contrast than the first generation (figure 2). Remarkably, throughout 6 generations, several peptides were identified that can produce contrast that is up to 80% of the contrast that PLL produces at 3.6 ppm. while this evolution experiment is still in progress, our data point out that POET can be used to evolve new function.

Discussion

This mechanism has shown the ability to discover new peptides that have challenged the current understanding about what chemical properties allow MRI contrast to be generated. Although the training data, and prior research was focused on peptides that were both positively charged and basic, POET had led to the discovery of CEST agents that are neutral in charge and pI but produce more contrast than PLL.

Conclusion

The contrast of the peptides generated by POET produce substantially more contrast than PLL after a few generations of using the program. Meanwhile, selected peptides developed using POET, that showed superior contrast can be assembled into a new MRI reporter gene improving its sensitivity over previous generations of reporters.

Acknowledgements

This research was made possible by financial support from NIH/NIBIB: R01 EB031008; R01EB030565; R01EB031936; P41-EB024495