Confirmation that the electrode interface is the main roadblock to CMOS biosensor commercialisation at present, has been corroborated in [16], where they state:

‘The primary design challenge using CMOS technology is the interface design between assay and integrated chip (IC) which generally calls for additional post-fabrication steps to facilitate compatibility in detecting targets (e.g., analytes).’

Constructing an overall picture of biosensor research is impeded by the large quantity of published articles and patents combined with the diversity of the technologies and applications. However, a quantitative analysis of the literature shows the most commonly reported IC-based biosensors use cells as bioreceptors with optical fibre or piezoelectric transducers. The fibre optic element is not a true transducer in its own right since its role is more accurately described as a conduit to a transducer that is remote from the bioreceptor and analyte [17]; these remote sensors can use various spectroscopic techniques such as absorption, fluorescence, phosphorescence and surface plasmon resonance. Additionally, cell-based biosensors using piezoelectric, magnetoelastic or optical-based transducers [Raman, FTIR (Fourier Transform Infrared Spectroscopy), SPR (Surface Plasmon Resonance)] are not readily implemented in standard CMOS integrated circuits and therefore lie beyond the scope of this review. Conversely, electrochemical (amperometric, potentiometric, conductometric, impedimetric) transducers are most suited to manufacture using standard CMOS processes since electrodes in contact with an analyte can be readily formed on the surface of the integrated circuit (the formation of CMOS electrodes is discussed later in Section 4.3). These electrodes may be used in conjunction with the various types of bioreceptor discussed above.

The use of complex IC technology in a biosensor application naturally needs justification. Where, for example, a passive device is called for, requiring no transistors and where a single layer can define both the tracks and the electrodes, then this may often be achieved cost-effectively using a custom manufacturing process employing simple photolithographic methods. However, when a specification calls for circuitry close to the electrodes, such as low noise pre-amplifiers for neuronal recordings, the benefits of using CMOS are obvious. Research on biosensors using CMOS presently utilise mature fabrication processes, with the ability to define features of only ∼0.1 μm or larger. However, the industry is currently working toward features as small as 22 nm for 2011 [18]. The economies of scale resulting from volume manufacturing and the ability to pack data processing capabilities into very small areas of silicon chip enable multi-electrode arrays (i.e., the transducer element) with excellent spatial resolution to be manufactured. These advantages are also combined with a much lower device cost than can be achieved using a custom semiconductor manufacturing process.

Prior to discussing the modification of CMOS electrodes and the changes necessary to satisfy the requirements of a practical transducer for cell-based biosensors, it is essential to understand the metal surface which is naturally presented by aluminium pads to environments such as physiological solutions and cells. The nature of aluminium is determined by its physical and chemical characteristics which are fundamental to its performance as an electrode material. The environments which are of most interest here for cell-based biosensors are aluminium in air, water and physiological media. Aluminium, along with other metals used in medicine such as titanium, is very reactive—its surface reacts spontaneously with air to form a ‘natural’ oxide film of amorphous Al₂O₃. This reactivity is determined by the Gibbs free energy of formation, being highly negative for aluminium (−791.15 kJ·mol⁻¹) and titanium (−888.8 kJ·mol⁻¹) [26,27]. The instantaneous reaction with oxygen in air results in an oxide growth rate that is proportional to log-time: the thickness forms very quickly to approximately 10 nm after which it is self-passivating, preventing further reaction and film growth [28].

The stability of the metal and oxide film in a medium is defined by the electrochemistry of corrosion and is best illustrated using a Pourbaix diagram as shown in Figure 4. It is important to consider the pH not only of the bulk solution but also the localised conditions. The electrochemical potential, E, defined by the Nernst equation, will also vary due to local conditions such as alloying species, defects and contaminants. At low pH, aluminium dissolves to form Al³⁺ ions and at high pH it dissolves to form AlO₂⁻, these conditions both being the basis of corrosion. In deionised water corrosion should not occur, but in saline physiological medium, Cl⁻ chloride ions will be adsorbed to the surface which creates localised acidic conditions (i.e., dilute hydrochloric acid) under which the passivated surface deteriorates leading to corrosion of the underlying metal [29]. This process then accelerates (auto-catalytic) since the aluminium dissolution process causes a further increase in Cl⁻ concentration at the corrosion site. A saline physiological medium is therefore potentially damaging to an unprotected aluminium CMOS electrode [30].

In addition to the stability of an aluminium surface exposed to a biological medium, it is vital to understand the factors that determine the electrical characteristics of such an interface. For example and as discussed later, the electrical characteristics of electrodes for transducers sensing neuronal activity are determined primarily by the chemistry of the solid-solution interface. The basis for modern interface models originates from the theory devised by Hermann von Helmholtz in the nineteenth century, later refined by Gouy and Chapman. The basic premise of the models is the existence of an electrical double layer at the solid-solution interface which in turn forms the basis of electrical circuit models of the interface. These theories are presented in [27] and [31] and will not be discussed further in this review.

Any CMOS electrode in contact with biological cells, tissue and/or physiological medium needs to be compatible with its environment, must not alter the physiology of the analyte under investigation or being detected and must be non-toxic to all the biological components in the system. Unfortunately, one difficulty with assessing biocompatibility of IC materials is that no definitive list has been compiled [32] and therefore results must be taken from more loosely related applications such as orthopaedics or smaller evaluations.

Without modification of the CMOS pads, the surface presented by nearly all CMOS technologies is aluminium. The biocompatibility of aluminium and its oxide, alumina, have been thoroughly studied, much work having being done to evaluate in vivo performance for use with orthopaedic prosthetics [33,34]. The performance of aluminium also depends much on the adherent superficial native oxide layer (alumina) and corrosion. However, the in vivo use of alumina has generally been confined to orthopaedics. This is because of its poor compatibility with blood due to its thrombogenic action (i.e., its tendency to cause undesirable clotting). Frequently, aluminium is coated with titanium nitride to improve in vivo performance [33]. For applications ranging from prosthetics to CMOS transducers, the overall in vivo interaction with their environment is primarily governed by the natural chemistry of the body: simplistically, this is a NaCl aqueous solution of concentration ∼0.1 M with organic acids, proteins, enzymes, macromolecules, electrolytes, dissolved oxygen and nitrogenous compounds. The resulting pH is approximately 7.2, often decreasing to ∼5.5 in the vicinity of tissue damage. Possible interaction mechanisms between any material, such as an electrode, and the body include inert/bioinert (no reaction), biodegradation (gradual breakdown by biological or biochemical processes), bioresorption (removal by cell activity or by continuous ionic diffusion) and bioactivity (a specific behaviour of a material).

The degradation of the native oxide of aluminium (alumina) on a CMOS aluminium electrode in a physiological environment is limited by its natural corrosion resistance due to the metal being in its highest oxidation state. However, a concern is that either defects in the oxide film on an electrode may enable aluminium ions to leach into the body or that the alumina itself may degrade. In physiological conditions aluminium easily forms an insoluble Al(OH)₃ precipitate or a solution of AlCl₃. The toxicity of these and other aluminium salts (10–100 mM) has been evaluated and shown to have only a small effect on the viability of mammalian neuronal cells [35].Additionally, neurotoxicology research into Alzheimer’s, has not shown a causal relationship between the disease and aluminium exposure [36]. Walpole et al. [37] and Karlsson et al. [38] tested nanoporous alumina substrates for aluminium ion leakage and found the dissolution rate of ions into culture measured after nine days was sufficiently low to be considered as non-toxic. In the context of IC materials, an in vitro assessment in [39] showed an enhanced proliferation (vitality) of Caco2 epithelial cells on aluminium versus the glass controls. These results therefore suggest that aluminium with a stable native oxide may form a biocompatible surface as a transducer for cell-based biosensors, but only where the electrode remains electrochemically stable—a condition usually difficult to achieve in the context of biosensors.

Whilst we are primarily interested in the CMOS electrodes in the context of this review, the biocompatibility of the surrounding material must not be overlooked. CMOS ICs predominantly use silicon nitride as the surface (passivation). Receveur et al. [32] concluded that silicon nitride is biocompatible, as stated by references [40–42]. In [39] it was shown that silicon nitride was an excellent substrate for Caco2 cell proliferation.

Research into CMOS interfaces for neuronal recordings has centred mainly on two types of electrode: the electrolyte-oxide-semiconductor (EOS) FET and the metal electrode. These are explored in subsections (a) and (b) below while section (c) describes some more recent research using porous alumina based electrodes. It is maintained that the composition of the transducer interface is the primary obstacle to designing successful products; other aspects of a CMOS neuronal interface product, such as signal amplification, data processing and communication, can leverage capabilities that are already well-established in the semiconductor industry. In this respect, impressive CMOS MEAs have already been demonstrated [43–51] and so the design of CMOS circuitry is not a focus of this review.

The height of the cleft between the cell and electrode is a key parameter that influences the magnitude of extracellular neuronal signals and has therefore been a primary consideration in the development of state of the art neuronal sensors [66]. The use of adhesion proteins has been a main line of investigation as a tool to minimise the cleft by forming tight electrode junctions with the electrode/substrate. It was concluded in [67] that the proteins promote cell adhesion and that the cleft can be minimised using a coating of RGD peptide sequence (RGD = arginine–glycine–aspartic acid (Arg-Gly-Asp)—a laminin and fibronectin fragment). In [68] it was noted that RGD immobilises cells, with the exception of neurons. Polylysine was also discussed as an adhesion molecule but the cleft size was expected to be larger compared to a layer of RGD peptide. In [63] and [69] it was stated that the YIGSR peptide sequence—a laminin fragment—also promotes cell adhesion whilst minimising cleft size (YIGSR = tyrosine–isoleucine–glycine–serine–arginine). Other methods to promote good adhesion include the use of polyethylenimine (PEI) and laminin [70]. Whilst generally successful at producing adhesion, the cleft is wider and so they produce a less efficient electrical interface (Figure 7) (coupled voltage reduces with the square of the cleft distance). A MEMS approach to adhesion has also been investigated whereby a wafer was micromachined to provide pneumatic anchoring of rat cardiomyocytes [71]. This technique has been successfully incorporated into a family of single-use MEAs manufactured by Cytocentrics AG (Germany) [72]. Investigations by [73,74] showed that modification of a silicon surface by patterning (in the range of tens of nanometers to micrometers) can also assist with attachment.

Silicon substrates have been successfully modified by [76] and [77] to produce porous silicon. Scanning electron microscope (SEM) images suggested a tight junction, but these need to be confirmed by electrical characterisations. Unfortunately, the opening of windows to silicon introduces the same drawback as the EOS FET in that ionic contamination of the CMOS logic circuits is likely.

A second key parameter affecting the effective signal of a neuronal recording is the impedance of the electrode. The majority of publications reporting successful recordings from CMOS ICs (c.f. references in Section 4.3 above) use a coating on the electrode of platinum black to increase its effective surface area and hence decrease its impedance. This is a well-established method used for electrophysiology and other electrochemistry applications [78]. CMOS IC electrodes have been successfully coated with biocompatible platinum and platinum black [79]: the platinum pads were created by a lithographic patterning step, but the subsequent platinum black was an electrodeposition performed by biasing the stimulation circuitry. The platinum black was shown to reduce successfully the electrode impedance. Interestingly, the benefit arising here from decreased electrode impedance is not intuitive: the improvement comes not from an increase in signal amplitude (the signal magnitude relates to the proportion of an electrode that is covered by a cell, i.e., forming a potential divider. Decreasing the impedance per unit area using platinum black decreases the impedance under the cell but also decreases the impedance to the grounded bulk electrolyte and therefore the platinum black has little direct effect on signal amplitude) but instead from reducing the noise produced by the electrode itself—i.e., the benefit is an improved signal-to-noise ratio at the FET gate input. This occurs since the ‘root mean square’ of the instantaneous noise (r.m.s.) thermal noise (additional sources of noise exist in electrode-electrolyte interfaces [80] but thermal noise dominates in neuronal recording FET-based electrodes), V, produced in an electrode is proportional to the root of its resistance, R, where, k is Boltzmann’s constant, T is temperature and B is the bandwidth [81,82] (see Equation 1): (1) V = 4 k   T B R

Commercial MEAs used for mammalian neuronal recordings include the Multi Channel Systems and MED64 (Alpha Med Scientific, Inc.) products. Quality assurance data for MEAs purchased from Multi Channel Systems showed all electrodes had individual impedances of 39 k–41 kΩ at 1 kHz (the characteristic frequency of an action potential), that is in agreement with their documentation claiming 20 k–400 kΩ for their range of titanium nitride electrodes [83]. The MED64 electrodes are claimed to have impedances of 7 k–10 kΩ at 1 kHz [84]. To prepare CMOS IC electrode arrays for neuronal recordings at the University of Bath, it was decided that the 1 MΩ impedance at 1 kHz achieved with the gold deposition alone was unacceptably high. The standard technique of lowering the impedance by coating with platinum black was used, with a target impedance of 40 kΩ (at 1 kHz) for 30 μm diameter electrodes. Figure 8 shows a porous alumina and gold CMOS electrode after deposition of platinum black.

As outlined in Section 1, the potential applications of a silicon-electrode junction are diverse and so the particular type of cell being interfaced is relevant and must be taken into consideration when reviewing recent developments in the literature. For example, many of the published results demonstrating successful recordings from electrically active cells relate not to mammalian neurons but to human embryonic kidney (HEK) cells and cardiomyocytes [48,62,70,79,85–96]. Such cells may often produce signals of larger amplitude than achievable with mammalian neurons. Therefore, due to the low electrode impedance, cell positioning and small cleft requirements, producing reliable extracellular mammalian neuron recordings using CMOS electrodes is probably one of the most demanding applications. The above review of neuronal recording serves to illustrate the complexity of many CMOS transducers and cell-based biosensor applications where the neurons serve as the bioreceptor.

This section discusses MEAs for neuronal recording and stimulation applications where the electrode is acting as a transducer. Although generally suffering from limited spatial resolution and expensive to manufacture, passive MEAs have been manufactured commercially for some time. As a result, as already noted, much work was begun during the 1990s in order to leverage the potential benefits of planar semiconductor technologies—primarily CMOS—offering a potentially cheap source of electrodes, integrated signal processing and increasingly excellent spatial resolution (with on-chip signal multiplexing and processing, large CMOS arrays may be formed with spatial resolution limited by the cell culture density rather than the IC technology. For example, a 128 × 128 array of 10 μm diameter electrodes at 50 μm spacing would be a resolution of 4 × 10⁸ m⁻²). Early investigations, especially by groups such as Fromherz et al. at the Max Planck Institute of Biochemistry identified the adherence of biological cells to the IC electrodes as a major challenge [101]. Progress has been slow as the factors influencing cell-substratum adhesion are complex.

The principal manufacturer of MEAs is presently Multi Channel Systems (MCS) GmbH (Germany), supplying a range of MEAs with planar microelectrodes, typically an array of 60 recording electrodes, plus an optional large planar stimulation electrode [102]. MCS have developed a comprehensive set of amplification, data acquisition and data analysis tools to support their MEAs product line. MCS have identified that low impedance electrodes are necessary in order to achieve acceptable signal-to-noise ratios and they have therefore developed a high surface area titanium nitride (TiN) electrode which typically achieves an impedance of only 40 kΩ at 1 kHz for a circular electrode of 30 μm diameter (Figure 10). Whilst the electrodes are high performance, their durability is low: MCS state that the electrodes are re-usable but cleaning is difficult as the electrodes cannot be touched with cotton buds, etc, due to the fragility of the TiN. The electrodes also seem to degrade in ambient conditions and therefore might have a useful maximum lifetime of one year. Other suppliers of passive MEAs use conventional platinum and platinum black electrodes, probably due to their improved robustness. Ayanda Biosystems SA (Switzerland), manufacture MEAs with a footprint compatible with the MCS data acquisition systems [103].

Ayanda produce square (40 × 40 μm) platinum electrodes with typical impedances of 400–600 kΩ (at 1 kHz). In partnership with Bio-Logic SAS (France), Ayanda also produce passive MEAs with up to 256 electrodes [104]. Alpha Med Scientific, Inc.(Japan), manufacture the MED64 system which comprises 64 electrode MEAs and supporting data acquisition and analysis instrumentation [105]. All three manufacturers use glass substrates which have the benefit of allowing imaging of the cells using phase contrast microscopy. MEAs can be supplied optionally with indium tin oxide (ITO) tracks that are transparent at visible wavelengths and therefore further improve the clarity of optical microscopy. This is not possible with CMOS as the silicon substrate is opaque. The MED64 system uses square electrodes of 50 × 50 μm coated with platinum black to achieve a low impedance of only 7–10 kΩ (at 1 kHz). It should be noted that this low impedance is achieved at the expense of the platinum black creating tall (7.39 μm) dendritic growths above the substrate surface (Figure 11). It is questionable whether the body (soma) of dissociated neurons will readily cover such a tall feature and instead the cells may prefer to adhere only to the substrate.

A recent entrant into the MEA market is the company 3Brain [106], a spin-out of CSEM (Centre Suisse d’Electronique et de Microtechnique) and represents a first for the penetration of CMOS technology into the MEA market.

The 3Brain APSMEA chip (Figure 12) uses CMOS to form an array of 4,096 electrodes for high spatial resolution electrophysiology. At approximately £140, it appears competitively priced against the Multi Channel Systems passive MEA (£250–300 each). Each electrode is 21 μm × 21 μm with a pitch of only 42 μm. The IC includes on-chip address decoders and signal processing, and claims high sensitivity and low-noise [44]. Details of the CMOS post-processing and the electrode surface are undisclosed [44,107] and therefore a detailed appraisal of their process is not possible here.

Beyond neuronal interfaces, commercial CMOS-based products are presently few. However, new products are emerging, but are mainly beyond the scope of this article (i.e., not cell-based), such as the DNA sequencing chip from Ion Torrent (now part of Life Technology Corp.) [108]. However even here, the use of CMOS requires additional layers to be deposited on the top of the CMOS and patterned using photolithography to provide biocompatibility. Similarly, CustomArray Inc. also uses CMOS in their DNA synthesis chips where photolithographic post-processing is required to form the electrode array [109]. Bionas use CMOS ICs in a 96-well plate format [110]. This seems to represent the first commercial product using CMOS in such a format and is an indication of what it likely to become a considerable focus for this technology in the future. However, the Bionas ICs require considerable photolithographic custom-/post-processing steps to integrate the various type of sensor (pH, oxygen consumption and cell-substrate impedance).