6.2 Bioelectronic Devices Based on Bacteriorhodopsin

6.2 Bioelectronic Devices Based on Bacteriorhodopsin

Bacteriorhodopsin is the light-transducing protein in the cell membrane of an archae bacterium called Halobacterium salinarium, an organism originally called Halobacterium halobium (Oesterhelt and Stockenius 1971; Balashov, Litvin, and Sineshchekov 1988; Mathies et al. 1991; El-Sayed 1992; Lanyi 1992; Ebrey 1993; Luecke et al. 1999). Upon absorption of light, a covalently attached chromophore (light-absorbing group) initiates a complex series of molecular events that converts light energy to chemical energy. Bacteriorhodopsin-based devices exploit this complex series of spectrally distinct thermal intermediates (Zimanyi and Lanyi 1993; Lanyi and Varo 1995). The excellent holographic properties of the protein derive from the large change in refractive index that occurs following light activation. Furthermore, bacteriorhodopsin converts light into a refractive index change with remarkable efficiency (approximately 65%). The protein is ten times smaller than the wavelength of the stimulating light. Therefore the diffraction limit of the optical geometry, rather than the "graininess" of the film, determines the resolution of the thin film. Because the protein can absorb two photons simultaneously with an efficiency that far exceeds other materials, it can store information in three dimensions using twophoton architectures. Finally, the protein was designed by nature to function under conditions of high temperature and intense light (Shen et al. 1993; Lukashev and Robertson 1995), a necessary requirement for a salt marsh bacterial protein and a significant advantage for photonic device applications.

Photonic Properties of Bacteriorhodopsin

When the protein absorbs light in the native organism, it undergoes a complex photocycle, which generates intermediates with absorption maxima spanning the entire visible region of the spectrum (figure 6.4). Most current devices operate at ambient or near-ambient temperature and utilize the following two states: the initial green-red absorbing state (bR) and the long-lived blue absorbing state (M). The forward reaction only takes place via light activation and is complete in ∼50 μsec. In contrast, the reverse reaction can be either light activated or can occur thermally. The lightactivated M → bR transition is a direct photochemical transformation. The thermal M → bR transition is highly sensitive to temperature, environment, genetic modification, and chromophore substitution. This sensitivity is exploited in many optical devices based on bacteriorhodopsin. Another reaction of importance is a photochemical branching reaction from the O intermediate to form P. This intermediate subsequently decays to form Q, a species that is unique in that the chromophore breaks the bond with the protein but is trapped inside the binding site. The Q intermediate is stable for extended periods of time (many years) but can be photochemically converted back to bR. This branching reaction provides for long-term data storage, as discussed below (Birge et al. 1994).

Figure 6.4: Spectra of select intermediates during the bacteriorhodopsin photocycle. The outlined arrows indicate photochemical transitions, and the solid arrow represent thermal transitions. The insets represent the conformation of the retinal in that state. N, nitrogen; X, Schiff base nitrogen (in P) or carbonyl oxygen (in Q).

Associative Memories

Associative memories take an input data block (or image) and, independently of the central processor, "scan" the entire memory for the data block that matches the input. In some implementations, the memory will find the closest match if it cannot find a perfect match. Finally, the memory will return the data block in memory that satisfies the matching criteria or it will return the address of the data block to permit access of contiguous data. Some memories will simply return a binary bit indicating whether the input data are present or not present. Because the human brain operates in a neural, associative mode, many computer scientists believe that the implementation of large capacity, high-speed associative memories will be required if we are to achieve genuine artificial intelligence. We have implemented the design proposed by Paek and Psaltis (1987) by using thin films of bacteriorhodopsin as the photoactive components in holographic associative memories (Birge et al. 1997). The memory is shown schematically in figure 6.5.

Figure 6.5: Schematic diagram of a Fourier transform holographic (FTH) associative memory with read/write FTH reference planes using thin polymer films of bacteriorhodopsin to provide real-time storage of the holograms. Note that a partial input image can select and regenerate the entire associated image stored on the reference hologram. Although only four reference images are shown, an optical associative memory can store many hundreds or thousands of images simultaneously. This memory can also work on binary data by using redundant binary representation logic, and a small segment of data can be used to find which page has the largest association with the input segment. Selected components are labeled as follows: FL, Fourier lens; FVA, Fresnel variable attenuator; H1,H2, holographic films; PHA, pin-hole array; SF, spatial filter; SP, beam stop.

Both the reference and input images are entered into the system via a spatial light modulator (input SLM) and are focused via Fourier lenses (FL) onto the two holographic films, H1 and H2. Fourier association at H1 results in preferential illumination of the pinhole corresponding to the reference image that has the highest correlation (similarity) to the input image, or partial image. The radiation passing through that pinhole illuminates the selected image on H2, which is then transferred out of the associative loop onto a CCD detector. Thresholding is handled electronically, rather than optically in this implementation; however, optical thresholding can improve performance (Paek and Psaltis 1987; Gross, Izgi, and Birge 1992; Birge et al. 1997). As the example in figure 6.5 shows, only a partial input image is required to generate a complete output image.

The utility of the associative memory regarding a data search can be greatly enhanced by rapidly changing the holographic reference pattern via the single optical input while maintaining both feedback and thresholding. In conjunction with solidstate hardware, a search engine of this type can be integrated into hybrid computer architectures. The diffraction limited performance of the protein films, coupled with high write/erase speeds associated with the excellent quantum efficiencies of these films, represents a key element in the potential of this memory. The ability to modify the protein by selectively replacing one amino acid with another (see discussion below) provides significant flexibility in enhancing the properties of the protein (Hampp et al. 1994).

Three-Dimensional Memories

Many scientists and engineers believe that the major impact of molecular electronics on computer hardware will be in the area of volumetric memory. There are three different types of protein-based volumetric memories currently under investigation: holographic (d'Auria et al. 1974; Birge 1990; Heanue, Bashaw, and Hesselink 1994; Psaltis and Pu 1996), simultaneous two-photon (Parthenopoulos and Rentzepis 1989; Chen et al. 1995; Dvornikov and Rentzepis 1996) and sequential one-photon (Birge et al. 1994; Stuart et al. 1996; Birge et al. 1999). We have already described a holographic memory based on bacteriorhodopsin. Thus we will confine our discussion to the latter two architectures. These memories read and write information by using two orthogonal laser beams to address an irradiated volume (10–200 μm³) within a much larger volume of a photochromic material. Either a simultaneous two-photon or a sequential one-photon process is used to initiate the photochemistry. The former process involves the unusual capability of some molecules to capture two photons simultaneously. The sequential one-photon process requires a material that undergoes a branching reaction, where the first photon activates a cyclical process and the second photon activates a branching reaction to form a stable photoproduct. The three-dimensional addressing capability of both memories derives from the ability to adjust the location of the irradiated volume in three dimensions. In principle, an optical three-dimensional memory can store roughly three orders of magnitude more information in the same size enclosure relative to a two-dimensional optical disk memory. In practice, optical limitations and issues of reliability lower the above ratio to values closer to 300. Nevertheless, a 300-fold improvement in storage capacity is significant. Furthermore, the two-photon or sequential one-photon approach makes parallel addressing of data possible, which enhances data read/write speeds and system bandwidth.

The simultaneous two-photon memory architecture has received a great deal of attention in the past few years, and because bacteriorhodopsin exhibits both high efficiency in capturing two-photons and a high yield of producing photoproduct after excitation (Birge and Zhang 1990), this material has been a popular memory medium. But more recent studies suggest that the branched-photocycle memory architecture may have greater potential. This sequential one-photon architecture minimizes unwanted photochemistry outside of the irradiated volume and provides for a particularly straightforward parallel architecture. Above, we discussed the use of the P and Q states for long-term data storage. The fact that these states can only be generated via a temporally separated pulse sequence provides a convenient method of storing data in three dimensions by using orthogonal laser excitation. The process is based on the sequence shown in figure 6.6, where K, L, M, N, and O are all intermediates within the main photocycle, and P and Q are intermediates in the branching cycle. The numbers underneath the letters give the wavelengths of the absorption maxima of the intermediates in nanometers (e.g., bR has a maximum absorbance at 570 nm, or yellow-green region; O absorbs at 640 nm, in the red). The terms all-trans, 13-cis, and 9-cis below the letters refer to the isomerization state of the chromophore while the protein is in the particular intermediate form. Figure 6.7 shows a schematic representation of the operations required for implementing the branched-memory architecture, which are described in detail below.

Figure 6.6: Data storage based on the branched photocycle reactions of bacteriorhodopsin.

Figure 6.7: Schematic diagram of the branched-photocycle three-dimensional memory. The top diagram shows the write process, which is based on a page operation followed a few milliseconds later by an orthogonal write operation.

The reading and writing process starts by selecting a very thin region (∼15 microns) inside the data cuvette by a process called "paging". In this process, the paging lasers (there are two, one on each side of the data cuvette, but only one is shown for clarity), with a wavelength in the region 550–640 nm, initiates the photocycle within a ∼15 micron slice of the memory medium. The photocycle will return to the resting state (bR) in about 10 msec, the time window during which subsequent writing or reading must take place. In the absence of secondary laser stimulation, the protein within the paged region will simply return to the resting state. Thus a page operation by itself is nondestructive.

A parallel write is accomplished by using the sequential one-photon optical protocol. The paging beam activates the photocycle of bacteriorhodopsin and after a few milliseconds, the O intermediate approaches maximal concentration. The data laser and the SLM are now activated (ƛ = 680 nm, at ≈ 3 msec) to irradiate those volume elements ("voxels") into which 1 bits are to be written. This process converts O to P in only the simultaneously activated and irradiated locations within the memory cuvette. After many minutes, the P state thermally decays to form the Q state (the P → Q decay time, τ_P, is highly dependent upon temperature and polymer matrix). The write process is accomplished in ∼10 msec, the time it takes the protein to complete the photocycle. Because the decay time of the P state is variable, our memory architecture must treat both P and Q as bit 1, and have the capability to erase both of these species simultaneously when the voxel is to be reset to bit 0.

The read process takes advantage of the fact that light around 680 nm is absorbed by only two intermediates in the photocycle of light-adapted bacteriorhodopsin, the primary photoproduct K and the relatively long-lived O intermediate (see figure 6.4). The read sequence starts out in a fashion identical to that of the write process by activating the 568 nm paging beam. After two milliseconds, the data timing (DTS) and the data read (DRS) shutters are opened for 1 msec, but the SLM is left off, allowing only 0.1 percent of the total laser power through. A CCD array (clocked to clear all charges prior to reading) images the light passing through the data cuvette. Those elements in binary state 1 (P or Q) do not absorb the 680 nm light, but those volumetric elements that started out in the binary 0 state (bR) absorb the 680 nm light, because these elements have cycled into the O state. Noting that all of the volumetric elements outside of the paged area are restricted to the bR, P,or Q states, the only significant absorption of the beam is associated with O states within the paged region. The CCD detector array therefore observes the differential absorptivity of the paged region, and the paged region alone. This selectivity is the key to the read operation, and it allows a reasonable signal-to-noise ratio even with thick (1–1.6 cm) memory media containing >10³ pages. Because the absorptivity of the O state within the paged region is more than one thousand times larger than the absorptivity of the remaining volume elements combined, a very weak beam can be used to generate a large differential signal. The read process is complete in ∼10 msec, which gives a rate of 10 MB/sec. Each read operation must be monitored for each page, and a refresh operation performed after ∼1000 reads. Although data refresh slows the memory slightly, page caching can minimize the impact.

Data erasure can be accomplished a single page at a time by using a blue laser with output near 410 nm or globally by using incoherent light in the 360–450 nm range. The former method has the advantage of selectivity. The first commercially available blue diode lasers at 405 nm can soon be implemented to perform this type of erasure process. The current prototype uses the global erasure process, which is an acceptable alternative, and is certainly much less expensive. These and other issues are more clearly understood with reference to an actual prototype, as discussed below.

Prototyping and Reliability Studies

The 3-D volumetric memory based on the branched photocycle is in a second generation of prototyping, and the current prototype is shown in figure 6.8. The prototype has been designed in a modular fashion so that individual components can be rapidly replaced and new component subsections can be tested quickly. The key features are discussed below.

Figure 6.8: A close-up picture (a) and a schematic (b) of the current Level II prototype of the branched-photocycle volumetric memory. Key components of the prototype are shown and labeled in the schematic (b). This prototype uses a global erasure of the entire data cuvette via activation of a pair of cylindrical UV lamps. The lamps are located at the focus of two cylindrical hyperbolic mirrors positioned so that the entire memory cuvette can be irradiated (insert c). Dual data lasers coupled to the active matrix liquid crystal spatial light modulator (AMLC SLM) via a holographic diffuser are used to generate non-coherent light so that the imaged data beam does not contain diffracted artifacts (see text). (Reproduced from Birge et al. 1999.)

Dual paging lasers are used to provide homogeneous page activation. Although 570 nm lasers would be optimal, we are using two 60 mW, 630–635 nm lasers because relatively inexpensive diode lasers are available with this wavelength. We hope to convert ∼25 percent of the molecules within the irradiated page to O state, in preparation for either writing or reading.

Holographically coupled dual data lasers are implemented to deal with an interesting problem associated with small, pixel-sized active matrix liquid crystal spatial light modulators (AMLC SLMs). The use of such SLMs to control coherent light sources generates pixel aperture interference that is otherwise absent with incoherent sources. This problem has been reduced to acceptable levels by using two data lasers that are coupled via a pair of lenses to a holographic diffuser, as shown in figure 6.8. The coupling of two lasers removes the coherence inherent in single laser sources, and the holographic diffuser provides an even intensity distribution across the surface of the modulator. We use two 50 mW 680 nm diode lasers. The high power is required because of the low quantum efficiency of the O → P photochemistry (see discussion above). If the native protein is used, only a very small amount of protein can be converted from bR to Q. We can repeat the write process as many times as we wish until the necessary amount of Q state is present for the read operation. Organic cation analogs or Glu-194/Glu-204 mutants yield conversions that are 5–10-fold higher, and approach the conversion levels that are necessary for reliable and efficient operation in one write sequence.

Although it may appear inefficient to move the data cuvette rather than the optics, paging through cuvette translation is the preferred method when speed and reliability are taken into account. The data cuvettes are light (4.5 g plastic, 9.1 g BK-7) and the kinematic mount and the Peltier device dominate the inertial weight. The prototype uses a high-speed linear actuator (Newport 850F-HS) to locate the cuvette. This device allows the cuvette to be positioned, on average, in 60 msec, a value sufficient for the present testing purposes. Ultimately, we will use linear magnetic motors to achieve 10 msec latencies.

Global erasure of the entire data cuvette is achieved via activation of a pair of cylindrical UV lamps (All Electronics, UV-325 + BXA 12576), which have outputs of 463 40 nm. This soft UV light couples adequately to the absorption spectra of both P and Q, and sets all the bits within the entire memory cuvette to state zero (bR) in a few minutes. The lamps are located at the focus of two cylindrical hyperbolic mirrors and are collimated, as shown in the inset (c), so that the entire memory cuvette is irradiated when properly positioned.

Multiplexing of the data to achieve higher storage density can be implemented by using gray-scale multiplexing. This scheme stores 5 bits per voxel. (If we included polarization, we could store ten bits per voxel; Birge et al. 1999.)

Multiplexing is an important component of 3-D data storage, as it achieves viable storage densities without the need to operate at or near the diffraction limit. The spatial light modulator (Kopin, 320C) and the CCD detector (Cohu 1100) are designed to provide 8-bit gray-scale capability, which yields the potential of multiplexing 16 bits into each voxel via polarization doubling. In practice, reliability limits the writing and reading processes to 5 bits per polarization, leading to a current bestcase scenario of 10 bits/voxel. In addition, a number of bits along the edges must be allocated to alignment and checksum information. With these features fully implemented, a standard 1 1 3 cm³ data storage cuvette can store approximately ten gigabytes. This storage capacity translates to the use of between 10⁴ and 10⁵ protein molecules per bit for typical protein concentrations utilized here, and provides an adequate number of molecules per bit to yield a statistically relevant ensemble, as required to maintain reliability (see below) (Birge, Lawrence, and Tallent 1991; Birge et al. 1999).

Error analysis can be carried out by measuring read histograms (Lawrence et al. 1998; Birge et al. 1999). A detailed discussion of error sources and error analysis for 3-D memories is beyond the scope of this chapter (we direct the reader to Lawrence et al. 1998 for a more comprehensive discussion). Briefly, we note that a read operation carried out via differential absorption is measured on a normalized scale where 0 is low intensity reaching the detector (bit 0, O state absorption), and 1 is high intensity reaching the detector (bit 1, P and Q present and less O state absorption). By using a differential read process, the detector and associated circuitry normalize the signals across the detector array so that when no gray scaling is done, you get peaks at only two locations, 0 and 1. When gray scaling is used to increase bit density, you get peaks corresponding to each of the allowed levels, which correspond to bit assignments (see discussion in Birge et al. 1999). Read errors invariably increase the width of these peaks, and can even cause the peaks to overlap. The error rate is proportional to the overlap, and the goal is to keep such error at a level that can be handled via error-correcting codes.