Experimentally Calibrated Computational Prediction Enables Accurate Fine-Tuning of Near-Infrared Rhodamines for Multiplexing

Abstract

A significant barrier inhibiting multiplexed imaging in the near-infrared (NIR) is the extensive trial and error associated with fine-tuning NIR dyes. In particular, the need to synthesize and experimentally evaluate dye derivatives in order to empirically identify those that can be used in multiplexing applications, requires a large investment of time. While coarse-tuning efforts benefit from computational prediction that can be used to identify target dye structures for synthetic campaigns, errors in computational prediction remain too large to accurately parse modifications aimed at fine-tuning changes in dye absorbance and emission. To address this issue, we screened different levels of theory and identified a time-dependent density functional theory (TD-DFT) approach that can rapidly, as opposed to synthesis and experimental evaluation, estimate absorbance and emission. By calibrating these computational estimations of absorbance and emission to experimentally determined parameters for a panel of existing NIR dyes, we obtain calibration curves that can be used to accurately predict the effect of fine-tuning modifications in new dyes. We demonstrate the predictive power of this calibrated dataset using seven previously unreported dyes, obtaining mean percent errors in absorbance and emission of 2.2 and 2.8%, respectively. This approach provides a significant timesavings, relative to synthesis and evaluation of dye derivatives, and can be used to focus synthetic campaigns on the most promising dye structures. The new dyes described herein can be utilized for multiplexed imaging, and the experimentally calibrated dataset will provide the dye chemistry community with a means to rapidly identify fine-tuned NIR dyes in silico to guide subsequent synthetic campaigns.

Graphical Abstract

Calibration to experimental data produces accurate computational predictions of NIR dye absorbance and emission for unknown derivatives, allowing for identification of fine-tuning modifications.

Introduction

Fluorescence imaging has revolutionized our understanding of biological systems.^{[1, 2]} Nonetheless, this optically-based imaging technique has traditionally been limited to cell culture or transparent organisms due to the absorbance and scattering properties of tissues. Diminished absorbance of light in the far-red to NIR (650 – 900 nm) has re-invigorated chemistry-focused efforts to identify robust fluorophores capable of NIR absorbance and emission,^{[3, 4]} due to the potential for deep tissue imaging in the NIR. For example, heteroatom substitutions of the bridging oxygen atom (10-position) in the tretramethylrhodamine scaffold with Si,^[5] P,^{[6, 7]} S,^{[8, 9]} Ge,^[5] Se,^{[8, 9]} and SO₂^[10] as well as other modifications^{[11, 12, 13]} can lead to significant red-shifts in absorbance and emission, in some cases ≥100 nm. These coarse-tuning efforts have yielded new rhodamine-based dyes with NIR absorbance and emission characteristics necessary for imaging in multicellular organisms. With these new parent rhodamine-based NIR scaffolds in-hand, a remaining challenge for the field is further fine-tuning of spectral properties to identify NIR dyes for multiplexed imaging applications. Such dyes would further complement the growing palette of multiplexable cyanine dyes.^{[14, 15, 16, 17]} Here, as compared to coarse-tuning efforts with rhodamines, subtle changes in absorbance and emission ranging from ~10 – 30 nm are desirable to maximize the number of observable fluorophores using spectral unmixing algorithms.

Traditionally, efforts to fine-tune the absorbance and emission properties of dyes have relied on intuition and iterative rounds of synthesis and testing of derivatives (Figure 1). While useful, this approach is time consuming and requires human intuition to select targets for synthesis. Alternatively, TD-DFT provides a rapid means to estimate relative absorbance and emission transitions for coarse-tuning applications.^{[7, 11, 13, 18, 19, 20, 21]} Unfortunately, the error in predicted absorbance and emission wavelengths of TD-DFT calculations is often too large to reliably parse modifications aimed at fine-tuning absorbance and emission wavelengths for multiplexing applications.

Figure 1.

Approaches for fine-tuning absorbance and emission of xanthene-based dyes. The generic NR dye scaffold is shown with numbering. The calibrated computational approach described herein enables rapid identification of fine-tuned dyes.

Our lab has demonstrated the ability to dramatically red-shift the absorbance and emission of xanthene dyes by ~100 nm via incorporation of phosphinate substituents at the bridging position, yielding the Nebraska Red (NR) series.^{[7, 22]} During the process of derivatizing NR dyes for various applications,^{[23, 24]} we have obtained a catalog of experimentally characterized dyes with absorbance and emission in the far-red to NIR. This catalog of existing dyes provides us with standards that can be used to calibrate computational methods. Herein, we screen levels of theory in Gaussian for their ability to provide rapid estimations of absorbance and emission wavelengths. We then calibrated computational estimates to experimentally determined values from our dye catalog. Using this approach, we obtained a trendline capable of accurately predicting the effect of fine-tuning modifications on absorbance and emission of seven previously unreported dyes with a mean percent error (MPE) of 2.2 and 2.8%, respectively. Lastly, we show that four new dyes within this set can be resolved by spectral unmixing.

Results and Discussion

As a first step, we sought a level of computational theory that would provide a rapid estimation of absorbance and emission maxima of NR dyes. Using the published NR₆₆₆ phosphinate-rhodamine as a representative test case,^[7] we optimized ground state geometries and estimated absorbance wavelengths in Gaussian 16^[25] using a panel of functionals including PBE,^[26] PBE0,^[27] M05,^[28] B3LYP,^[29] M06,^[30] CAM-B3LYP,^[31] HSE06,^{[32, 33, 34, 35, 36, 37]} ωPBE,^{[38, 39, 40]} and ωB97X-D.^[41] The polarizable continuum model (PCM)^[42] for water and the 6-311+G(2d,p) basis set^{[43, 44]} were used throughout. While several studies have investigated xanthene excited states using B3LYP,^{[7, 11, 13, 18, 19, 20, 21]} we observed a 2-fold decrease in computational time for the PBE exchange-correlation functional relative to B3YLP (Figure S1). Additionally, while all methods underestimated the absorbance wavelength for NR₆₆₆ (resulting in a 15 – 26% error in estimated absorbance maxima), PBE was the most accurate (Figure S2). Since our goal was to use computational guidance to rapidly predict the effect of fine-tuning modifications in NIR dyes, as compared to relatively time-consuming synthesis and experimental characterization, we chose to proceed with PBE due to its ability to produce the most rapid estimations of dye absorbance within our panel of functionals (Figure S1). To further test the predictive potential of PBE, we estimated absorbance and emission wavelengths for a series of previously published heteroatom-containing rhodamines (Table S1). Comparison to published data yielded strong correlations between computationally predicted and reported absorbance and emission maxima (R² = 0.86 and 0.91, respectively, Figure S3). This initial work indicated that the relatively fast PBE level of theory could be used to estimate the effect of coarse-tuning modifications on xanthenes.

Next, we used PBE to estimate absorbance and emission wavelengths for a panel of previously published NR dyes containing an array of auxochrome, phosphinate ester, and pendant phenyl ring modifications (Figure 2a, 2b, and Table S2).^{[7, 22]} These modifications include coarse-tuning modifications (e.g. NR₆₀₀ versus NR₆₆₆) as well as structural changes that more finely tune spectroscopic properties (e.g. NR₆₆₆ versus NR₇₀₀ and NR₆₆₆ versus NR₆₉₈). Using this published NR dyes series, we obtained computationally predicted absorbance and emission wavelengths that, while underestimated, correlated extremely well with experimentally observed values (Figure 2c and d; R² = 0.99 and 0.98, respectively). Comparison of oscillator strengths for absorbance to experimentally determined molar extinction coefficients (also known as molar absorption coefficients or molar absorption factors) showed little correlation between these values (Figure S4). Thus, oscillator strengths obtained using this method should not be used to predict the relative efficiency of dye absorbance.

Figure 2.

Experimental calibration of a computational dataset using previously described NR dyes. a) Core structure of xanthene-based NR dyes with variable sites R₁, R₂, X, Y, and Z indicated. b) Previously published NR dyes with computationally predicted as well as experimentally observed absorbance and emission maxima. Correlations between computationally determined and experimentally observed absorbance (c) and emission (d) maxima.

We next asked whether the trendlines in Figure 2c and d could be used to rapidly predict the absorbance and emission of new NR dyes. Accordingly, we envisioned seven new NR dyes bearing possible fine-tuning modifications (Figure 3a). This panel of new dyes leveraged cyclization of the auxochrome to the 2,7- or 4,5-positions of the xanthene ring. In particular, cyclization to the 4,5-positions of heteroatom xanthenes is uncommon and, to the best of our knowledge, benzomorpholine xanthenes have not been reported (e.g. NR₆₇₇ and NR₇₂₉). Thus, given the limited information concerning dyes within this panel, using human intuition to make predictions would be challenging. To address this issue, we employed PBE to predict absorbance and emission maxima for each new dye and then used the trendlines from Figure 2c and d to convert computational values into predicted absorbance or emission wavelengths (Figure 3b). Using established procedures,^[7] we next synthesized and purified each dye in Figure 3a and experimentally determined their absorbance and emission maxima in DPBS (Figure S5) as well as photophysical properties (Table 1). Interestingly, we observed a general trend of blue-shifted absorbance and emission for cyclization at the 4,5-position relative to isomeric dyes with cyclization at the 2,7-position (e.g. NR₆₆₇ versus NR₆₉₉, NR₆₇₂ versus NR₇₂₁, and NR₆₇₇ versus NR₇₂₉). These observations reinforce the need for computational approaches to parse fine-tuning modifications aimed at identifying dyes for multiplexing applications, as these trends would be difficult to predict using human intuition. Comparison of predicted to experimentally obtained values showed that trendlines from Figure 2c and d produced reliable predictions of both absorbance and emission wavelengths with MPEs of 2.2 and 2.8%, respectively. It should be noted that the % error for NR₇₂₉ was appreciably higher than for other dyes in the panel. Since other NR dyes in this spectral range (e.g. NR₇₂₁ and NR₇₄₄) show good correlations to the computational approach used here, we attribute the relatively larger error in predictions for NR₇₂₉ to unique electronic features of this dye that are not approximated well by the level of theory employed. In addition to the trends in absorbance and emission wavelength of isomeric dyes, we also observed relatively higher molar extinction coefficients and quantum yields for dyes cyclized at the 4,5-postion compared to their isomeric derivatives (e.g. NR₆₆₇ versus NR₆₉₉, NR₆₇₂ versus NR₇₂₁, and NR₆₇₇ versus NR₇₂₉). Indeed, NR₆₆₇, NR₆₇₂, and NR₆₇₇ are all brighter than NR₆₆₆ (Table S2) the brightest NR dye previously described in this wavelength range.^[7] This enhancement in fluorescence brightness is primarily due to an increase in quantum yield of all three previously unreported dyes compared to NR₆₆₆. Although structural modifications that increase quantum yields of xanthenes within the visible range have been described,^{[45, 46, 47, 48, 49]} general approaches for increasing quantum yields of NIR xanthenes have remained elusive.^[50] We are currently assessing whether structural modifications at the 4,5 position can be used as a general approach to increase quantum yields of NIR xanthenes. Overall, the level of error observed herein is well within the range required to identify dyes with fine-tuned absorbance and emission for multiplexing applications, providing a relatively rapid approach for screening new dye structures as compared to time-consuming synthesis and experimental evaluation.

Figure 3.

Accurate prediction of absorbance and emission maxima of unknown dyes. a) Structures of previously unreported NR dyes. b) Predicted and experimental absorbance and emission maxima for unknown dyes along with percent errors in each prediction. Abs. is Absorbance, Ex. is Experimental, and Em. is Emission.

Table 1.

Photophysical properties of previously unreported NR dyes.

Dye	Ex (nm)	Em (nm)	ε (M−1·cm−1)	Φ	ε × Φ
NR 667	667	681	146,660	0.52	76,263
NR 672	672	689	187,240	0.48	89,875
NR 677	677	689	184,810	0.51	94,253
NR 693	693	720	67,700	0.25	16,925
NR 699	699	725	120,000	0.13	15,600
NR 721	721	758	82,540	0.06	4,952
NR 729	729	781	60,100	0.03	1,803

All parameters were measured in DPBS (pH = 7.2, with 0.5 % DMSO).

The ability to accurately predict the influence of fine-tuning modifications allows for the selection of target dyes with desirable absorbance and emission maxima for multiplexing. As a proof-of-principle for multiplexed imaging applications using the new dyes identified herein, we investigated the ability of spectral unmixing to resolve separate samples of NR₆₆₇, NR₆₇₇, NR₆₉₃, and NR₇₂₁ imaged on the same 96-well plate (Figure 4). Pure samples of each dye were prepared in DPBS and aliquoted into separate wells of a 96-well plate. The plate was then imaged using 605 or 745 nm excitation to resolve the fluorescence signal of NR₆₆₇, NR₆₇₇ and NR₆₉₃ or NR₇₂₁, respectively, using spectral unmixing algorithms (Figure S6). This experiment clearly demonstrates the ability to resolve fluorescence emission from each of these four dyes (Figure 4), indicating their potential as multiplexable probes for imaging. Furthermore, these experiments demonstrate the ability to utilize experimentally calibrated computational prediction to identify fine-tuned NIR rhodamines with desired absorbance and emission maxima for synthetic campaigns.

Figure 4.

Proof-of-principle for multiplexing fine-tuned NR dyes. Structures of NR dyes used for spectral unmixing, along with color coding, are shown. Spectrally unmixed fluorescence emission channels from wells containing 50 μM of the indicated dye in the same 96-well plate imaged using either 605 (NR₆₆₇, NR₆₇₇, and NR₆₉₃) or 745 (NR₇₂₁) nm excitation. The inset shows composite images for each dye in the respective channel. Error bars represent the standard deviation of 10 replicates.

Conclusions

Coarse-tuning efforts are tolerant to relatively large errors (~15 – 26%) in computational estimation of absorbance and emission since the goal of these studies is to qualitatively rank structural modifications such that new dyes with significant changes in absorbance and emission can be identified. However, fine-tuning efforts for identification of multiplexable dyes require higher accuracy in prediction of absorbance and emission since structural modifications that shift these parameters by ~10 – 30 nm are desirable to maximize the number of observable dyes. Consequently, errors in computational prediction of ≤ 4% are desirable for dyes in the NIR. To address this issue, we have screened levels of theory to identify a TD-DFT approach that can rapidly, as opposed to synthesis and experimental evaluation, estimate dye absorbance and emission. Calibrating estimated computational values to experimentally determined parameters for a panel of NIR dyes produces calibration curves that can accurately predict absorbance and emission of new dyes with MPEs of 2.2 and 2.8%, respectively. We show that this calibrated computational dataset can be used to accurately predict the influence of structural modifications on NR dye absorbance and emission using seven previously unreported dyes as a test case. The ability to rapidly identify fine-tuned NIR dyes for synthetic campaigns is an important step towards expanding multiplexing capabilities within this wavelength region. Indeed, we demonstrate a proof-of-principle for multiplexing four new dyes described herein, providing a basis for further development of multiplexed imaging probes. More broadly, this computational approach can be applied to other heteroatom xanthenes (Figure S3). Thus, we envision that this work will enable rapid fine-tuning of NIR dyes to illuminate biology.