(L-469) Facilitating clinical implementation of intratumoral injections by developing a panel of imageable drug surrogates

Cancer drugs are typically administered systemically, but delivery is hindered by biophysical obstacles such as elevated interstitial fluid pressure, dense tumor extracellular matrix, and hypoperfusion. Further, systemic exposure can cause dose-limiting side effects. Intratumoral injection (i.t.) is an appealing alternative. Image guidance can facilitate delivery of concentrated drug wherever a needle can be placed and pressurized fluid flowing from the needle tip can disperse drug over a greater volume than is possible with systemic delivery. However, clinical implementation is hindered by an incomplete understanding of the injected drug distribution and difficulty selecting injection protocols. One source of ambiguity is the unknown relationship between drug molecular weight and the injected distribution volume. Our hypothesis is that larger drugs will distribute over a lower volume but the distribution volumes are correlated.

Our goals are: 1) to demonstrate the correlation between differently-sized radiopaque and fluorescent surrogates and 2) to determine the dependence of distribution volume on drug size. Demonstrating the correlation between differently-sized radiopaque and fluorescent surrogates will facilitate inclusion of CT-imageable contrast agents with drug for i.t. to estimate the drug distribution in real-time. To determine the dependence of distribution volume on drug size we have developed a panel of fluorescent imageable surrogates to allow for high-resolution imaging and simultaneous discrimination of multiple surrogates.  While CT imaging entails a sacrifice in spatial resolution, it allows for rapid three-dimensional morphological analysis and is poised for clinical use.  Confocal microscopy and a linear regression model are used to unmix colocalized fluorophores.

Assessing correlation between radiodense and fluorescent drug surrogates

Ex vivo bovine liver was injected with a mixture of fluorescent albumin and iodine-containing contrast agent (iodixanol) at volumes of 1, 2, and 4 mL (n=3). Samples were frozen, imaged with microCT, sectioned at 20 microns, and imaged with a fluorescent slide scanner.

Training model to unmix colocalized fluorescent drug surrogates

The panel of imageable drug surrogates is composed of either dextran particles of 70, 150, or 500 kDa conjugated with the fluorophores FITC, Antonia Red, or TRITC or albumin conjugated with AlexaFluor680. The size of each surrogate was determined with dynamic light scattering. To train the unmixing model individual surrogates and all possible combinations were pipetted onto a 20-micron section of ex vivo bovine liver, allowed to dry, and then imaged with a confocal microscope. Samples were imaged with excitation lasers of 488, 561, 594, and 633nm and emission was collected in 32 9-nm wide bins ranging from 414 to 690 nm. To reduce the parameter size 8 combinations of excitation lasers and emission bins were selected. Each image was subsampled 100 times and fed into a linear regression model with a training-test split of 80-20.

To demonstrate the utility of the trained model the panel of imageable drug surrogates was mixed with a CT contrast agent, injected into ex vivo bovine liver, frozen, imaged with a microCT, and then cut into 20-micron thick sections.

Fluorescent albumin and iodixanol have non-identical but correlated distribution volumes. Figure 1A illustrates that the smaller iodixanol (MW = ~1 kDa) distributes over a greater volume that that of fluorescent albumin (MW = ~70kDa, Figure 1B). Figure 1C depicts that the distribution volume of iodixanol is approximately 7x greater than that of fluorescent albumin (MW = ~70kDa) and the correlation coefficient between the distribution volumes of the two contrast agents is 0.89.  Despite the large difference in distribution volumes (7x) between our fluorescent and radiopaque imageable surrogates, the high correlation coefficient (R² = 0.89) demonstrates the predictive value of CT imaging to visualize the distribution of a differently-sized drug.

Fluorescent dextrans of nominal molecular weight 70, 150, and 500 kDa were measured with dynamic light scattering and had diameters of 11.6 ± 0.2 nm, 16.5 ± 1.7 nm, and 22.6 ± 0.9 nm. Fluorescent albumin had a diameter of 12.5 ± 0.4 nm.  Each fluorescent drug surrogate was imaged with all four excitation lasers across all emission bins (Figure 1D-G). The spectra of individual surrogates as well as each possible combination was input into a linear regression model. The model was accurate to within 10% in quantifying the concentration of individual solutions. The accuracy decreased to 25% when applied to a mixture of all four fluorophores and to combinations three fluorophores. Combinations of two fluorophores yielded the lowest accuracy with the separation of 500kDa-TRITC and 150kDa-Antonia Rd causing the largest discrepancies.  Application of the linear regression model to an injection sample yielded a mixture of the 70 and 150kDa dextrans and albumin, but no 500kDa dextran.

These results will be extended by conducting a spatial analysis of the injection distribution of the panel of fluorescent surrogates. We hypothesis that larger surrogates will be most concentrated in the center of the distribution while smaller surrogates will distribute over a larger volume and be more dilute. More sophisticated supervised learning models such as a convolutional neural network will be evaluated. Future characterization will include measure the zeta potential and the lipid/water partition coefficient to approximate the binding affinity of surrogates to tissue.