Quantitative micro-Raman analysis of micro-particles in drug delivery

Polymeric micro and nanoconstructs are emerging as promising delivery systems for therapeutics and contrast agents in microcirculation. Excellent assets associated with polymeric particulates of tunable shape, size, mechanical and chemical properties may improve the efficiency of delivery and represent the basis of personalized medicine and treatment. Nevertheless, lack of effective techniques of analysis may limit their use in biomedicine and bioengineering. In this paper, we demonstrated Raman Spectroscopy for quantitative characterization of poly lactic-co-glycolic acid (PLGA) micro-plate drug delivery systems. To do so, we (i) acquired bi-dimensional Raman maps of PLGA micro-plates loaded with curcumin at various times of release over multiple particles. We (ii) realized an exploratory analysis of data using the principal component analysis (PCA) technique to find hidden patterns in the data and reduce the dimensionality of the system. Then, we (iii) used an innovative univariate method of analysis of the reduced system to derive quantitative drug release profiles. High performance liquid chromatography (HPLC), the consolidated method of analysis of macro-sized systems, was used for comparison. We found that our system is as efficient as HPLC but, differently from HPLC, it enables quantitative analysis of systems at the single particle level.


Supporting Information
.2. PC1 (a), PC2 (b) and PC3 (c) principal components extracted from Raman spectra of nine samples at the initial time of release . Displayed loading profiles = 0 indicate how much individual frequencies contribute to a specific principal component. Figure S1.2. PC1 (a) and PC2 (b) loading curves associated to a microplate at the initial time of release. In the images, colours indicate different spectral regions: Silicon peak (blue), principal curcumin resonances (orange), no signal region (green). The PC1 vs PC2 scatter plot in (c) indicates that the components associated to Silicon (blue) and curcumin (orange) are orthogonal (mutually independent).

SI-2. Details on the micro Raman post processing -1 st Normalization. In this Supplementary
section details on the first normalization (i.e. temporal normalization) on a reference Si sample and base-line subtraction are reported. In the Supporting Information Figure S2.1a the principal peak from Si at t=0 is shown and a Voigt fit is applied by Origin. Voigt function must be applied to RAMAN resonances due to its intrinsic mixed nature between a Gaussian and a Lorentzian shape.
In the insert of the Supporting Information Figure S2.1a the information from the fit are reported.
This fit is applied to Si reference for each time point. From this result, we acquired the x c central frequency and the related intensity is used to calculate the 1 st Normalization Constants as reported in the Supporting Information Figure S2.1b. How already explain in the main text of this paper, the x c frequencies, indicated as K Si , are not constant and an average value of 520.30.2 cm -1 is measured. Figure S2.1. Silicon spectra measured on a reference sample are used for the first normalization (a). Spectra are fitted using a Voigt function. Table in the inset (b) reports the amount of variation of the constants used for the first normalization at different time frames.

SI-3. Details on the micro RAMAN post processing -Baseline considerations. Definition and
subtraction of baseline from spectra is a common issue in spectroscopy. The presence of a baseline superposed to the spectra can be considered as a "reality effect". This is due to the variation of experimental condition during the acquisition of a measure as temperature variation, laser fluctuation or someone that open the lab door. Clearly all of them are not predictable a priori but only after the spectrum acquisition. Several techniques are used to define a polynomial function (one for each spectrum) that can be attributed to the baseline and subtracted to the spectrum. All of them, more or less accurate, have the limitation that if is unknown the nature of the baseline there is no way to verify if the selected function is really correct. The risk is to subtract a different function altering the spectrum. This is one of the main reasons because; also if theoretically Raman spectroscopy is a quantitative analysis, up to know is used only for qualitative evaluations. To overcome this problem the scientific community had start to work on different experimental configurations based on the Raman effect (i.e. CARS, SERS, TERS, etc…). Differently, the innovative analytical approach presented in this paper, evaluate the baseline on single frequencies over the entire map. This changing in the point of view is crucial to reach the Micro-Raman quantification, as stressed in the main text. In the Supporting Information Figure S3.1 the histograms of the Intensity distribution (after the 1 st normalization) over a map at five fixed frequencies for t=0 h is shown. The choice of these frequencies and the histogram interpretation is already described in the main text. What we notice is that the Gaussian distribution of the data referred to the map-points outside the particle (lower Intensity peak) has different central position

SI-4. Evaluation of PL bunch purity by Micro-Raman quantification.
As already explain in the paper, one of the main advantages of a Micro-Raman analysis compared to the HPLC is the possibility to study individual particles instead of a bunch. Studies on the entire population at once will considers also whatever is present in the solution where particles are dispersed in and, depending on the process used to synthetize the particles, it can alter their real pharmacological characteristics, for example for the presence of debris. In the Supporting Information Figure S4.1 we compared the CURC release evaluated by RAMAN quantification with the HPLC performed on a "purified" (as shown in Fig. 5) and "not purified" bunch of particles. How expected, RAMAN analysis can be considered as a reference for the purity of the sample from debris just because it is measured from the particles and no fragments are taken in account. From the Supporting Information Figure S4.1 it is possible to understand how critical is this consideration on the release behaviour. In effect, to have a controlled release only by the engineering of the particles, the definition of a purity reference becomes a need.
Supporting Information Figure S4.1. Comparison between curcumin release curves derived using (i) our Raman based method of analysis (blue); (ii) HPLC performed on a highly-purified sample (red); (iii) HPLC performed on an untreated sample (black).

SI-5. An alternative method base on clustering to separate the signal from the background in
the micro plates. We performed unsupervised clustering of Raman spectra acquired on the entire region of interest setting the maximum number of clusters as . Then, we compared the positions in 2 the plane of the spectra partitioned into two clusters to the position of the micro-platelet placed on the stage the micro Raman set up for analysis. From the comparison, it results that the clustering algorithm discriminated between in-silicon and out-of-silicon spectra with a precision as high as , with the highest assignment errors at the border of the particle. While this level of accuracy ~80% is relevant, clustering alone does not achieve perfect matching between the spectra and their position on the sample surface. On the contrary, histogram description of spectral intensities associated to the Raman band enables direct graphical representation of data, reduces 1630 -1 uncertainty, and achieves an efficiency of selectivity near unity. Figure S5.1 Supporting Information Figure S5.2 SI-6. An alternative method of analysis of Raman spectra, derivation of the drug release profile and comparison with the main findings of the paper. For each time step, for each particle per time step, we separate all spectra of a particle from the background using unsupervised clustering of Raman spectra setting the maximum number of clusters as . The clustering algorithm 2 discriminates between in-silicon and out-of-silicon spectra. Then, we baseline correct all spectra of a particle using an automatic procedure. After correction, we find the average of all the spectra associated to a particle. Then, we calculate the integral of the absolute values of extended over the entire spectral range:

Supporting Information
We then further select from the pool of spectra all those spectra with a characteristic integral that departs less than from , i.e. it is within the band . This serves as a procedure to 0.5 = ± 0.5 eliminate from the sample all anomalous spectra. Then, we take the average of the selection: the mean spectrum of a restricted group of spectra represents the characteristic spectrum of a particle, that will be used for deducing the release profile (Supporting Information Figure S6.1). Then, we report all means measured at different time frames in the same diagram (Supporting Information Figure S6.2). Of the means, we select a frequency band centered around (Supporting 1630 -1 Figure S6.2), and calculate the release profile from the particle as , where is the ( ) = 1 -/ Raman spectrum intensity measured at at a specific time , and is the intensity at 1630 -1 initial time. The fraction of drug released over time departs from that measured in the main article using the second normalization and HPLC methods to a great extent (Supporting Information Figure S6.3). Results reinforce the main findings of the work, that a second normalization of data