Polarity‐triggered anti‐Kasha system for high‐contrast cell imaging and classification

Kasha's rule, which states that all exciton emissions occur from the lowest excited state and are independent of excitation energy, makes high‐energy excitons difficult to use and severely hinders the widespread applications of organic photoluminescent materials in the real world. For decades, scientists have tried to break this rule to unleash the power of high‐energy excitons, but only minimal progress has been achieved, with no rational guiding principles provided, and few applications developed. So far, breaking Kasha's rule has remained a purely academic concept. In this paper, we introduce a design principle for a purely organic anti‐Kasha system and synthesise a series of compounds based on the design rule. As predicted, these compounds all display evident S2 emissions in dilute solutions. In addition, we introduce a highly accurate (over 90%) convolutional neural network as an assistant for the classification of cells using anti‐Kasha luminogens, thereby providing a new application direction for anti‐Kasha systems.


INTRODUCTION
The introduction of the Jablonski diagram, which has remained relevant for almost a century, finally provided in-depth insight into the physical process associated with photoluminescence. [1][2][3] The term 'photoluminescence' is now ubiquitously used in daily life [1,2,4] and covers a wide array of topics and fields, including pigments, medicines, sensors, actuators and light-emitting and photovoltaic devices, which have all been boosted by fundamental advances in photoluminescence research. Although Jablonski predicted infinite excited states for all optical materials, emission (fluorescence or phosphorescence) almost always comes from the lowest excited state of a standard organic photoluminescent system; hence, emission is independent of excitation, with higher energy excitons essentially useless for photoluminescence. This limitation was first stated by Kasha in the 1950s, is referred to as 'Kasha's rule' [5] and is supported by extensive studies, especially in pure organic systems. [6][7][8] Various version of Kasha's rule exists, with the most famous corollary being Vavilov's rule, which states that the luminescence quantum yield is generally independent of the excitation wavelength. [9] Kasha's rule states that all emissions come from the same equilibrium structure of the first singlet excited state, irrespective of excitation, and all excited dynamics are independent of the shapes of the potential energy surfaces of states higher than S 1 or T 1 , which implies that most of the nature of an emission, including its maximum wavelength, quantum yield and fluorescence lifetime, is unaffected by the excitation energy.
A higher energy exciton experiences internal conversion (IC) prior to emission. In most cases, IC (often 10 −12 s) is much quicker than fluorescence emission (often 10 −9 s), with excitons sufficiently internally converted and relaxed back to the equilibrium structure of the lowest excited state, which is the S 1 state for fluorescence, prior to luminescence in most cases. Emission from a higher excited state should be possible when the IC rate is comparable to the fluorescence rate, [10,11] and generally, IC needs to be suppressed to produce an anti-Kasha system; [12,13] however, the IC rate is difficult to determine for a specific macroscopic system. Empirically, adiabatic energy gaps play crucial roles that negatively correlate with IC rates. [6,14,15] Non-adiabatic coupling between electronic states is of great importance in such a process. [16] The Fermi golden rule states that the transition rate for an isolated energy level is proportional to the density of states between the initial and final states of the system: [17] where ⟨f |H ′ |i⟩ denotes the matrix element of the perturbation H′ between the two states and (E f ) is the density of states, which is the derivative of the number of continuum states with respect to the potential energy. If coupling occurs between two excited states, the result Γ i→f is proportional to the IC rate (k IC ). Equation (1) reveals that the IC rate correlates negatively with the adiabatic energy gap (ΔE adia S2−S1 ), which can be obtained by calculating the energy difference between the equilibrium structures of the S 2 and S 1 states. Modulating the S 2 /S 1 adiabatic energy gap to a relatively large value would lower the IC rate to a level comparable to the rate of fluorescence, which would lead to emission from S 2 . Attracted by the possibility of the involvement of higher excitation states, many attempts to break Kasha's rule have been made since it was first stated. [11,[18][19][20][21][22][23][24][25][26][27][28] The most famous higher-state emitter is azulene (shown in Scheme 1C), which contains an unsaturated fused seven-membered/fivemembered ring system and is a specific example of Hückel's rule of aromaticity. [12,13,29,30] To meet the 4n+2 π electron requirement, the ground state of azulene has a highly uneven charge distribution and a coincidingly very small transition energy from the ground state to the S 1 state, which also increases the energy difference between the S 1 and S 2 states. In addition, after 70 years, despite the discovery of occasional examples that display emissions from higher excited states, there is no universally applicable guiding principle or general rule for the design and use of such materials. Hence, a rational approach to anti-Kasha behaviour is highly sought after because it would lead to a new generation of photoluminescent materials with limitless possibilities that emit from higher excited states.
The adiabatic energy gap is best regulated by introducing a donor-acceptor (D-A) structure, as previously mentioned. [31,32] The adiabatic energy gap between a ground and S 1 state can be tuned using a single D-A couple. Similarly, the adiabatic energy gap between S 2 and S 1 is tuneable when two D-A couples are combined in one conjugated system (Scheme 1D). With this in mind, in this study, we designed a series of pure organic molecules, each with a single electron acceptor and two different electron donors, to finally achieve anti-Kasha behaviour and produce significant emissions from S 2 in dilute solution. In addition, the anti-Kasha behaviour of the synthesised compounds was found to be easily tuned by the polarity of the environment. This behaviour was used to map endocellular polarity and classify cells based on mapping textures.

Design and synthesis
Using the D-A-D' model shown in Scheme 1D, we built an extensive base of compound candidates by choosing various electron donors and acceptors. A larger difference between the electron-donating abilities of the two donors should result in a larger gap (ΔE adia S2−S1 ). For ease of synthesis, p-(dimethylamino)benzene), tetrahydroquinoline, p-(diethylamino)benzene, and p-(diphenylamino)benzene (DPAB) were chosen as stronger donors (D), while pmethoxybenzene (MB), p-hydroxybenzene (HB), and benzene (B) were chosen as weaker donor (D') candidates. Twelve possible structures exist after permutation. To determine those with promising anti-Kasha behaviour, we needed to build a simple screening scheme for selecting the most suitable molecules.
Escudero et al. introduced an appropriate protocol for predicting anti-Kasha emissions from azulene derivatives. [16] Inspired by their basic idea, we simplified the calculational steps to make the approach faster and more practical. The PBE0 hybrid functional is highly cost-effective in dealing with the transition energies of small organic molecules. [33] S C H E M E 1 (A) Simplified Jablonski diagram of a Kasha system and an anti-Kasha system (IC stands for internal conversion). (B) Electron patterns and energy gaps of a specific system at the Frank-Condon point. (C) Schematic view of the transition densities of the S 0 , S 1 and S 2 states of azulene under PBE0-D3/def2-TZVP level (the orange and purple surfaces denote the induced and reduced spaces of electron density from ground state to excited state, respectively). (D) Schematic diagram of the energy levels of a molecule with one electron acceptor and two different electron donors. H and L denote the highest occupied molecular orbital (HOMO) and LUMO orbitals; S 0 , S 1 and S 2 denote ground state, first singlet excited state and second singlet excited state. D and A denote electronic donor and acceptor S C H E M E 2 (A) Electron donor screening and the calculated ΔE calc S2−S1 of the various donor couples (the pink shade highlights the largest ΔE ver S2−S1 values with same weaker donors). (B) Schematic depicting chemical syntheses As shown in Table S1, a small benchmark indicated that PBE0 is proper for the time-dependent (TD)-density functional theory (DFT) calculation of this specific system. We optimised the 12 structures at the B3LYP/def2-SVP level of theory using Grimme's D3 correction and then performed TD-DFT calculations under PBE0/def2-TZVP level, the results of which are shown in the table in Scheme 2A. The red shaded area highlights that DPAB has the largest ver-tical energy gap (ΔE ver S2−S1 ) at the Frank-Condon point (FC) for each weak donor tested, which is regarded as an indicator of ΔE adia S2−S1 . Knoevenagel condensation is particularly valuable for the construction of (C = C) π-bridged D-A systems. [34] The various systems were prepared as shown in Scheme 2B. Acetophenones containing weak donors were reacted with malononitrile under base catalysis in ethanol. The first-step products continuously react with aldehydes containing strong donors to yield the final product under base catalysis. The experimental procedures were facile, and the yields were promising (up to 80%). The final products, DPAB-MB (TMCB), DPAB-HB (TPCB) and DPAB-B (TBCB), were chosen for further investigation. [35,36]

Polarity-triggered anti-Kasha behaviour
The emission and absorption spectra of TPCB are shown in Figure 1A. The TPCB solution exhibits different emissions when excited differently. The photoluminescence features of TMCB, TPCB and TBCB in solution were recorded by acquiring their emission spectra in dilute organic solu-tions (acetonitrile), as shown in Scheme 3C; they all display two absorption bands in their excitation spectra. Shorterwavelength maxima were located at 380-400 nm, while longer-wavelength maxima were found at 540-560 nm. The emission wavelength is independent of the excitation wavelength for most cases that follow Kasha's rule. However, in our system, we observed different emissions when excited at shorter and longer wavelengths. TMCB, which has the smallest ΔE exp S2−S1 , emitted at 574 nm when excited at 395 nm but emitted at 657 nm when excited at 541 nm; the emission ratio (Φ S2 /Φ S1 ) was determined to be 65. With larger ΔE exp S2−S1 values, TPCB and TBCB exhibited even larger emission ratios (85 and 86, respectively), which is near the introduced design principle, which proposes that a larger ΔE results in a more obvious S 2 emission. We confirmed the absence of aggregation by dynamic light-scattering technique. However, in the Tuneable bands are easily achieved by restricting or promoting twisting through solvent polarity. Stronger D-A couples are more sensitive to changes in polarity than weaker ones. Therefore, anti-Kasha behaviour can also be switched by changing solvent polarity. We attempted to explore the emission features of TPCB in various solvents but found that excitation-dependent emissions were not observed in dilute solutions of TPCB in less-polar solvents ( Figure 1A). As shown in Table S2, S 2 emission was observed when excited at a shorter wavelength of approximately 380 nm in acetonitrile and dimethyl sulfoxide (DMSO); however, such an emission was not observed in solvents such as dichloromethane (DCM), tetrahedronfuran (THF) and toluene (with lower Δf values).
Detailed investigations were performed using acetonitrile and THF. As shown in Figure 1B, the emission wavelengths and fluorescence lifetimes are all dependent on the excitation wavelength in acetonitrile ( Figure S1 includes all four solvents). Emission was maintained at approximately 580 nm, with slight deviations, when excited in the shorter absorption wavelength band. When excited at a wavelength in the critical region between the two absorption bands, the emission wavelength shifted dramatically from 580 nm to around 660 nm and remained constant as the excitation wavelength was further extended into the absorption band. Furthermore, the fluorescence lifetime of TPCB in acetonitrile also varied; while its fluorescence decay rate could not be fitted using a first-order exponential function when excited at the band with higher energy; a second-order exponential function yielded a decent correlation coefficient (R 2 = 0.99548). The two lifetimes extracted from the second-order function are 3.569 and 0.250 ns. Similar results were obtained in DMSO ( Figure S2). The first-order exponential function perfectly fitted the data (R 2 = 0.98582) when the excitation wavelength was varied to 500 nm, and only one lifetime (0.711 ns) was obtained. However, in THF, the emission and fluorescence lifetimes remained unchanged with varying excitation energy.
Despite all synthesised compounds displaying excitationdependent emissions, it is possible that these observations are not due to emissions from higher excited states. Other phenomena, including: (1) the presence of an impurity or other species, (2) the red-edge effect reported by Weber in 1960, [37] which is another well-known mechanism for generating excitation-dependent emissions, and (3) transition dipoles in different parts of the molecule, may be responsible.
High-performance liquid chromatography (HPLC) was used to disregard the first possibility, which verified the presence of only one component in this system ( Figure S3). For the second possibility, the literature reports that the red-edge effect is found in rigid media and is due to a 'hot distribution' of molecules; relaxation is sluggish in some cases, which leads to the longer excitation wavelengths exciting 'hot' molecules. Theoretically, the excitation-emission wavelength of a red-edge effect system monotonically increases from the excitation wavelength, with maximum intensity reached at infinite wavelength. [38] However, the excitationemission wavelength correlation of our system (displayed in Figure 1B and Figure S1) exhibits a typical logistic correlation with two flat regions and a dramatic shift region at specific wavelength transitions between the S 0 → S 2 and S 0 → S 1 bands. A comparison with reported systems that exhibit the red-edge effect [39][40][41] led us to disregard this mechanism due to the abovementioned inconsistencies.
For the third possibility, the logistic correlation between excitation wavelength and emission wavelength, which led us disregarding the red-edge effect, is orthogonal for emitters with two transition dipoles in different parts of the molecule. Such systems partially inspired our dual-intramolecular charge transfer (ICT) model contribution to multiple isolated dipoles. However, our model differs in principle from these systems because the excitation-dependent emissions of our compounds are polarity dependent rather than universal and do not behave in the solid state. These dipoles exist regularly in any environment in a standard system with two dipoles and exhibit two different emissions. Even if the two isolated dipoles emit at the same wavelength in a low-polarity solvent, they still yield different fluorescence lifetimes. We did not observe any trace of a second lifetime in either subnanosecond or picosecond fluorescence-decay experiments (shown in Figure 1E and Figure S2) when THF or DCM was used as the solvent.
Due to its middle position among the three synthesised compounds, TPCB, was chosen for our computational investigations. The structure of TPCB in its ground state in acetonitrile was optimised using the DFT method at the B3LYP-D3/def2-SVP level. The absorption spectrum of TPCB in acetonitrile was calculated using the TD-DFT method at the PBE0/def2-TZVP level with D3 correction and the SMD solvent model, as it performs better in transition-energy-related calculations. [33,42] The experimental and simulated absorption spectra are shown in Figure 2. The calculated spectrum reproduced features of the experimental ones; there is an absorption band at around 500 nm corresponding to the S 0 to S 1 transition and a shoulder at 350 nm corresponding to the S 0 to S 2 transition. Equilibrium structures on the S 1 and S 2 surfaces were optimised at the B3LYP-D3/def2-SVP level. All optimised structures are reported in Cartesian coordinates in Supplementary Appendices 1-3. The HOMO of the ground state (point FC in Figure 2A) is generally located on the stronger electron donor (triphenylamine) component, with HOMO-1 located on the weaker electron donor (phenol) part, while the LUMO is located on the electron acceptor (dicyanovinyl) unit. TD-DFT calculations revealed that the S 0 to S 1 transition is significantly due to HOMO to LUMO excitation, while the major S 0 to S 2 transition is due to excitation from HOMO-1 to the LUMO ( Figure S4). The optimization calculation implies that the red part displayed in Scheme 3A is not fully co-planar due to the spatial conflict between the cyano group and the proton on the benzene ring. The twisted skeleton could also inhibit the couplings between the two above mentioned ICT states. These results are also consistent with the introduced design principle, which proposed that the S 0 to S 1 transition occurs from the stronger donor to the acceptor, with the S 0 to S 2 transition occurring from the weaker donor to the acceptor. The transition densities also lead to a similar conclusion ( Figure 1B and Figure S5). [43,44] From Einstein's radiative transition equation: f is oscillator strength, and E is vertical transition energy.
Besides, k IC is the IC rate constant and could be empirically obtained by energy gap law: In which ΔE S2−S1 (in m −1 ) is the energy gap between S 2 and S 1 state and could be also obtained by experiment and calculation. Since the energy gap law, which is an empirical statement, only reflects the density of states term in Fermi's golden rule, we also calculate the oscillator strengths between the singlet excited states ( Figure S6) to check whether the electronic couplings have effects on the IC process. From the calculation result, we can be inferred that the transient couplings between S 2 and S 1 state is promising and could be neglected, implying that the energy gap plays dominating role in the IC process from S 2 to S 1 state. We can get that k IC ≅ 10 10 s −1 . So, the number of S 2 excitons that return to the S 1 state via IC is 83 times greater than the number of excitons emitting in S 2 state. However, the fluorescence quantum yield of S 2 state is at least 86 times higher than that of the S 1 state. The quantum yields of corresponding systems are low in high-polar environments due to the twisted intramolecular charge-transfer (TICT) effect. Addition of some restriction of rotation of electron donors may enhance the fluorescence efficiency. Considering the quantitative limit of the instruments, especially the quantum yield measurement, and the error of the calculation, at least we can propose that the mentioned effect offsets the IC rate to same extent and make the S 2 emission observable. Figure S7 displayed excitation spectra monitored on S 1 and S 2 emission wavelengths and also indicated that not all S 2 excitons would emit from S 2 state, and there will be a big portion internally converted back to S 1 state.
The equilibrium structures of S 1 and S 2 (points ES1 and ES2, respectively, in Figure 2A) are different. For the former, transition from the triphenylamine to the dicyano group dominates the state, and torsion angle θ tends to be 90 • owing to the TICT effect, which is also responsible for the low quantum yields of the S 1 emissions of all products because their oscillator strengths are significantly lower than those of the vertical geometries. However, in this process, the potential energy of the S 2 state increases due to a reduction in conjugation. Furthermore, the dicyano groups can also be aligned vertically to the main structure with various angles φ; the energy of the S 2 state at 120 • (the original φ was 150 • ) can be lowered by overcoming a small energy barrier (3.62 kcal/mol) to a low energy point (ES2). Figure 3A,B displays the proposed energy surfaces and the calculated minimum energy pathway at the PBE0-D3/def2-SVP level using the SMD model for acetonitrile. The calculated ΔE adia S2−S1 is Finally, we demonstrated the excited state dynamics of the S 2 excitons. In FC structure, there is also a 30 degrees angle between the phenol group and the buta-1,3-diene skeleton, which may also diminish the couplings between S 2 and S 1 states. After excitation, the S 2 excitons relax to ES2, which is kinetically and thermodynamically permitted; IC is still partially possible. However, the emissions from the S 1 state are so weak that they are possibly covered by S 2 emissions. Shoulder peaks are observed in the emission spectra when excited at a shorter wavelength.
Considering that ΔE is the critical factor for anti-Kasha behaviour, we listed values in various solvents in Table S2 for discussion. We found a noticeable trend, in which ΔE decreases with decreasing solvent polarity factor (Δf). Theoretically, the vertical excitation energy is linearly related to Δf, with a specific slope for a single ICT system, and stronger D-A couples providing a more significant slope. We derived mathematical expressions to better understand the relationship between ΔE and solvent polarity. From the literature, [45][46][47][48] the S 0 to S 1 transition is described by (Lippert-Mataga equation): abs, S1 = −m abs g ( e, S1 − g while the S 0 to S 2 transition is described by: where: and μ g , μ e, S1 and μ e, S2 are the dipole moments of the ground, S 1 and S 2 states at their FC structures, respectively, while is the equivalent spherical radius of the solute (Onsager radius), and 0 is the vacuum permittivity. h and c stand for Planck constant and vacuum light speed. Combining Equations (5) and (6) yields: abs, S2 − abs,S1 = −m abs g ( e,S2 − e,S1 ) with: Because the S 0 to S 2 transition is from the weaker D-A couple, μ e, S2 is smaller than μ e, S1 ; hence, Equation 5 reveals that ΔE exp S2−S1 and Δf are linearly correlated in a positive manner. The calculation provides a value of approximately 1.14 for the theoretical slope of Equation 5, using the approximate equation: slope ≃ g ( e − g )∕ 3 as reported by Il'ichev and co-workers. [49] The experimental data are plotted in Figure S8, which reveals a highly linear relationship (R = 0.99511) with a slope of approximately 1.03, consistent with the theoretically determined value, implying that ΔE is positively correlated with solvent polarity. Therefore,

Cell-imaging and classification
Our in-depth study showed us that solvent polarity can be used to control the anti-Kasha behaviour of our system, and the ICT behaviour of the compounds themselves is also influenced by polarity. Although a gap exists between chemical polarity and cellular polarity, more information can be obtained by monitoring TPCB emissions in cellstaining experiments compared to regular bio-imaging agents. [50,51] MTT assays revealed that TPCB is highly bio-viable ( Figure S10). HeLa cells were stained with TPCB for 15 min; confocal laser scanning microscopy showed different signal distributions in the yellow and red regions when excited at 405 nm (for S 2 ) and 500 nm (for S 1 ) ( Figure 3A,B). The red signals provide detail, with shapes identified in the red chan-nel when excited with less energy. On the other hand, the yellow signals are more general and are particularly strong in some endoplasmic membrane structures and lipid droplets. The different distributions of the yellow and red signals were generally mapped using the red channel to divide the two signals after normalisation ( Figure 3C). The divided image provides the S 2 and S 1 emission intensity ratio (I y /I r ) in every single pixel, which reveals the microenvironment of each specific position in each cell.
This method was also applied to human lung fibroblasts (HLF) and COS7 cells. As shown in Figure 3E-J, we obtained different spatial distributions according to cell emission ratio. A high emission ratio was observed for HeLa cells in lipid droplets and some membrane structures. However, high emission ratios were only observed in regions surrounding the nuclei of HLF cells, and they exhibited a different pattern than the HeLa cells. A high emission ratio was mainly observed for COS7 cells in lipid droplets; nevertheless, these lipid droplets mainly accumulate around the nuclei instead of the cell edges. The S 2 emission map could also be regarded as internal standard to enhance the contrast of cell imaging photos. As displayed in Figure 4, in either S 1 , S 2 channel or bright field, the intercellular structure could not be observed constantly. Nevertheless, in Figure 4D,F, these structures are clearly displayed.
While humans can identify the abovementioned morphological changes, they are challenging to digitise. However, a convolutional neural network (CNN) is a well-developed technique for classifying images. We built a training dataset of the various cell types (HeLa, HLF and COS7) containing pictures with labelled image ratios (about 200 pictures for each type). A validation dataset was also built with 150 pictures (50 pictures for each type). A four-layer CNN model was then established using the PyTorch. [51] The first two layers is a two-dimensional convolution layer that is used to obtain the edges of the input images. The third is a reshape layer in which two-dimensional data are transformed into linear data. Then three linear layers are followed by the last layer, which is a softmax layer that returns an array of three probability scores that sum to 1 ( Figure 5). [52] The cross-entropy between the predicted and expected assignations, is calculated using the validation dataset during the learning process. The model converged to reach minimum cross-entropy after 200 epochs and gave an excellent sorting capability that was more than 88.0% accurate (95% confidence interval: ±5.5%). However, if the training set was from the red channel photographs, the accuracy on prediction is around 82.0% (95% confidence interval: ±7.0%). This result indicates that ratio map contains more comprehensive information and features of corresponding types of cells than simple one channel pictures, which only reflected the cellular textures.

CONCLUSION
In this study, we introduced a ΔE-oriented principle for obtaining pure organic emitters that exhibit anti-Kasha behaviour. Following this principle, we designed and synthesised a series of compounds (TMCB, TPCB and TBCB), all of which exhibit typical anti-Kasha behaviour in dilute solutions with excitation-dependent fluorescence and significant S 2 emission. In addition, solvent polarity was found to significantly impact anti-Kasha behaviour by influencing ΔE; we found that only a polar solvent can trigger excitationdependent emission. The environmental sensitivity of this anti-Kasha effect enabled TPCB to be used as a cell-imaging reagent to provide information about the microenvironmental polarity distributions in different types of cells. Moreover, we built a highly accurate (>90.0%) CNN model for cell-type classification purposes using cellular polarity distribution information provided by the two-channel signals from TPCB. Although the utilization of 405 nm light is not suitable for bio-imaging, we still provided a proof of concept of a real-world application of such system that will motivate the continual research for the development of compounds with better biocompatibility. The anti-Kasha system has been a purely academic concept for decades, with no generally applicable design rules. Our work provides rational access to anti-Kasha systems and highlights further potential applications, thereby expanding possibilities and providing encouragement for this field.