The BAD project: data mining, database and prediction of protein adsorption on surfaces

Abstract

Protein adsorption at solid–liquid interfaces is critical to many applications, including biomaterials, protein microarrays and lab-on-a-chip devices. Despite this general interest, and a large amount of research in the last half a century, protein adsorption cannot be predicted with an engineering level, design-orientated accuracy. Here we describe a Biomolecular Adsorption Database (BAD), freely available online, which archives the published protein adsorption data. Piecewise linear regression with breakpoint applied to the data in the BAD suggests that the input variables to protein adsorption, i.e., protein concentration in solution; protein descriptors derived from primary structure (number of residues, global protein hydrophobicity and range of amino acid hydrophobicity, isoelectric point); surface descriptors (contact angle); and fluid environment descriptors (pH, ionic strength), correlate well with the output variable—the protein concentration on the surface. Furthermore, neural network analysis revealed that the size of the BAD makes it sufficiently representative, with a neural network-based predictive error of 5% or less. Interestingly, a consistently better fit is obtained if the BAD is divided in two separate sub-sets representing protein adsorption on hydrophilic and hydrophobic surfaces, respectively. Based on these findings, selected entries from the BAD have been used to construct neural network-based estimation routines, which predict the amount of adsorbed protein, the thickness of the adsorbed layer and the surface tension of the protein-covered surface. While the BAD is of general interest, the prediction of the thickness and the surface tension of the protein-covered layers are of particular relevance to the design of microfluidics devices.

---

1 Introduction

Protein adsorption at solid–liquid interfaces is critical to a large number of industrial and biomedical applications. In the biomedical arena alone the adsorption of proteins on surfaces is relevant to areas as diverse as biomaterials, protein microarrays and lab-on-a-chip devices. This large span of interest is paralleled by the diverse intent of the quantification of protein adsorption on surfaces. For instance, for implantable biomaterials the protein adsorption has to be minimised to avoid bio-incompatibility. For protein microarrays, one needs to find the optimum balance between higher protein concentration on surfaces, which leads to an increase in overall sensitivity; and protein denaturation, which leads to sensitivity decrease. Finally, for lab-on-a-chip devices protein adsorption may be deleterious to the transport of biofluids in micro-channels and could complicate device design and operation, but the immobilisation of proteins in micro-chambers is essential for analytical functions.

The efforts aiming to understand and predict protein adsorption are usually based on the modelling and simulation that uses a combination, in various degrees, of either a discrete, molecularly-orientated approach, or a continuous, thermodynamically-orientated approach. At one extreme, the discrete approach is based on the description of the protein as a set of interconnected objects (atoms, or groups of atoms), but the sheer complexity generated by this approach precluded the development of any model with a significant level of generality. Firstly, treating the problem by computational chemistry, e.g., by molecular dynamics simulations,^1–3 is at present not feasible due to the large numbers of atoms involved and consequently the prohibitive computing resources required. Consequently, most models^4–7 approximate the protein by a collection of rigid objects, with simple shapes and uniform charge. This simplification, although able to deliver accurate predictions in a number of cases,⁷ is not helpful when attempting predictions for proteins with more complex shapes. The difficulty of describing accurately and efficiently the protein object is compounded by the failure to account for the large number of interactions and fluid-flow effects which affect the adsorption process.^8,9 Fundamentally, and computational problems aside, the discrete, molecularly-based models have an inherent difficulty in taking into account even the simplest adsorption parameters, e.g., concentration of the protein in solution.

The progress of the alternative approach, which describes protein adsorption in thermodynamic terms, has been more substantial, largely because the relevant theoretical framework had been already developed for other adsorption processes and because this approach had a higher empirical content than molecularly-based approaches. This thermodynamic-orientated approach can be connected more easily than molecularly-based models to experimental efforts, which can be used for the estimation of model parameters. Indeed, a wealth of publications developed models for a finer description of the protein adsorption, e.g., its kinetics,^7,10–12 and the correlation between adsorbed amount and protein concentration in solution in the form of adsorption isotherms^13,14 with parameters estimated from, or with results compared with experimental data. Despite its versatility, and in contrast with molecularly-based models, the thermodynamic models have an inherent difficulty in describing the impact of process conditions on protein adsorption, in particular their impact on adsorbed protein bioactivity. Indeed, as small changes in the pH, temperature or ionic strength of the solution can induce considerable changes in protein conformation, the adsorbed amount of protein and its bioactivity can change dramatically.

Despite this general and acute interest in protein adsorption and much work done towards its modelling and simulation, the complexity of the process translates into the incapacity of predicting it with a reasonable degree of accuracy,¹⁵ even if precise information regarding the protein structure, fluid environment and surface parameters is available. A similar problem of untamed complexity faced by other scientific areas, e.g., drug discovery,^16,17 material discovery,^18,19 is tackled by combinatorial approaches, which comprehensively map the relevant experimental ‘space’ and then seek to correlate input parameters (e.g., structure, composition) with output parameters (e.g., bioactivity, material properties).

To this end, we describe a Biomolecular Adsorption Database (BAD), freely available online, which aggregates published data regarding protein adsorption. The database can be used for the selection of materials or operation conditions for microfluidics devices.

2 Methods

2.1. The BAD concept and data collection

The Biomolecular Adsorption Database (BAD) was conceived as a web-orientated database comprising the experimental data reported in the literature. While the purpose of the database is to be a depositary of information regarding the adsorption of biomolecules in general, at the moment only data regarding the protein adsorption is present, due to its relative importance and to the complex behaviour of proteins at surfaces.

The database has been intended to be used primarily for the design of microfluidics devices, in particular for the selection of materials for device fabrication, and for the selection of the optimum operating conditions of lab-on-a-chip devices. Consequently, the BAD comprises only literature data that comprehensively report quantitatively the protein, surface and fluid environment descriptors. Unfortunately, only a small proportion of the literature reports on the full extent of protein adsorption data, but the large amount of the general literature counterbalanced this drawback.

An essential feature of the BAD project was the open and free access to the database, as well as the possibility of continuous upgrade, either by database maintenance and inclusion of newly published data, or entries submitted by individual researchers, similarly to e.g., Protein Data Base, PDB.²⁰ The first tentative on-line version of the BAD was mentioned before²¹ in the context of the calibration of micro-channels for microfluidic devices. Recently, we have reorganised and supplemented this database with additional entries and implemented it on a MS-Access platform. Where necessary, the primary data have been refined, and brief descriptions of the experimental techniques are provided. This contribution reports on the present design and capabilities of the BAD, which is freely available at http://dbweb.liv.ac.uk/bad/. An image of the entry portal to the BAD is presented in Fig. 1.

Fig. 1 Screenshot of the portal to the Biomolecular Adsorption Database.

The BAD contains experimental results from adsorption isotherms, plateaus and pseudo-plateaus of adsorption kinetics experiments and single adsorption experiments (when adsorbents are incubated in a protein solution with a known initial protein concentration). The primary data has been collected from the open literature (see ESI†) using the major literature search engines (e.g., Scopus, Wiley, Science Direct, ISI Web of Science, etc.) and using several combinations of relevant keywords, e.g., (protein adsorption)AND(contact angle). This initial search was followed by the critical analysis and augmentation of the published data.

Presently, the BAD comprises 768 records of protein adsorption experiments. In several instances, the pair of values (protein concentration in solution) vs. (protein concentration on the surface) had to be estimated from the original contributions through graphical interpolation of data points in the original figures. 80% of the data collected in the BAD come from adsorption isotherms data, 16% from adsorption kinetics experiments and 4% from single adsorption experiments.

2.2. Protein adsorption variables reported in the BAD

The database reports on several input parameters of the protein adsorption process, i.e., related to the protein, surface and fluid environment, together with additional information; the output parameter—mass of adsorbed protein on the surface; and some auxiliary information, i.e., method of measurement and relevant reference.

Protein variables. Presently the BAD comprises data regarding the adsorption of 25 representative proteins, namely: albumin (HSA and BSA) 21%; fibrinogen 17.7%; lysozyme 14%; immunoglobulin G 10.4%; alpha-lactalbumin 9.8%; myoglobin 4%; fibronectin 3.1%; ribonuclease 3%; cholesterol esterase 2.6%; alpha-chymotrypsin 2.2%; insulin 2.1%; beta-casein 2.1%; cutinase 1.7 %; human growth hormone 0.8%; immunoglobulin M 0.8%; alpha-2-macroglobulin 0.8%; alpha-s1-casein 0.8%; beta-lactoglobulin 0.8%; protein C3 0.5%; alpha-1-acid glycoprotein 0.4%; hemoglobin 0.4%; collagen 0.3%; gamma-globulin 0.3%; mucin 0.3%; protein A 0.3%. Structural files (entries in the Protein Data Base, PDB) are provided for 15 proteins, which represent about 90% of total entries in the BAD. The proteins with molecular structures available as PDB files are highlighted in italics. The BAD also provides information regarding the protein primary descriptors, i.e., the composition (number of residues and their fraction in the protein composition), and secondary structure motifs (alpha helices, beta sheets and S–S bonds). For a single case, i.e., fibrinogen, two structures are provided (the structural file of the whole molecule comprises three domains). Furthermore, using the protein sequence in FASTA format²² and a Matlab script, the BAD also provides the secondary descriptors, i.e., molecular weight, overall hydrophobicity, isoelectric point and the estimated surface tension of the protein. The protein descriptors are calculated from the individual parameters of the amino acids, i.e., mass, pKa, hydrophobicity (White & Wimley hydrophobicity scale²³) and surface tension (estimated using SciPolymer software from SciVision), weighted in accordance to the respective amino acid mass contributions. We also calculated the standard deviation of the amino acid hydrophobicity within the protein.

Surface variables. The BAD presently comprises 9 types of surfaces on which protein adsorption has been measured, in the following proportions: polymers 49%; oxides (including unmodified silica) 22.8%; modified silica 11.6%; silicon wafer 5.2%; phospholipids 4.2%; glass 3.9%; self-assembled monolayers (SAMs) 2%; gold 1.2%; mica 0.3%. The central surface descriptor is surface hydrophobicity. The surface hydrophobicity can be quantified by either surface tension or the contact angle. Despite the fundamental nature of surface tension, as opposed to the practical but empirical nature of the contact angle, the latter has been chosen as primary data in the BAD because of its widespread use. The relationship between the empirically measured contact angle and surface tension has been also described (see also ESI† and section 3.4.). When the original publication does not report the contact angle, its value is assumed to be equal to that reported by the same author(s) in a related paper; or assumed to be equal with an average of close values reported elsewhere for the same surface, but only for common types of surfaces such as glass, silica, etc. When both the advancing and receding contact angles are reported, the BAD reports their average.²⁴ The BAD reports the surface chemistry where available, e.g., specialised polymers, and assumes that adsorbent surfaces are smooth.

Fluid media variables. Presently, there are 11 different buffer solutions with distinct composition represented in the BAD, for which adsorption experiments were performed. The buffer descriptors recorded are pH, ionic strength and, where available, temperature. If the protocol for the buffer is available, the pH and the ionic strength are calculated using a web-based Buffer calculator available at http://www.liv.ac.uk/buffers/.²⁵ For experiments citing “room temperature” the value was assumed to be 22 °C. Phosphate buffers with no added NaCl are designated by PB, and those with added NaCl as PBS.

Protein concentration on the surface. The data regarding protein concentration on the adsorbing surface, C_surf, (mg/m²), has been collected from three different types of experiments, i.e., (i) adsorption isotherms data, (ii) adsorption kinetics experiments and (iii) single adsorption experiments. An explanatory note is provided in the BAD for each experiment type. 12 different measurement methods have been used for the quantification of the amount of protein adsorbed on surfaces, with the proportion for each technique as follows: UV adsorption 31.42%; radio-labelling 27.99%; ellipsometry 17.7%; Lowry method 5.86%; quartz micro balance (QCM) 4.55%; scanning angle reflectometry (SAR) 3.78%; total internal reflection fluorescence (TIRF) 3.77%; sedimentation field-flow fractionation (SdFFF) 2.47%; reflectometry 1.17%; bicinchoninic acid assay (BCA) 1.04%; surface plasmon resonance (SPR) 0.52%; optical waveguide lightmode spectroscopy (OWLS) 0.39%. A brief description for each of the techniques is provided as a note.

Protein concentration in solution. The protein concentration in the fluid media, C_sol, (mg/ml), has slightly different meanings for each of the above types of experiments, i.e., (i) the equilibrium concentration in solution after adsorption for adsorption isotherms data, (ii) the bulk protein concentration for adsorption kinetics experiments, and (iii) the initial protein concentration (decreasing during adsorption) for single adsorption experiments.

The BAD also provides links to additional information about the proteins, buffers, surfaces, and methods of measurement.

2.3. The organization of the BAD

The primary data was organized along several central tables (web pages). Each table contains the unique index for each row (unique or primary key). The tables are connected through the use of foreign keys that are referential constraints between relations. This approach avoids the redundancy of data in the BAD, maintains its coherence, and allows for the easy extraction of the relevant information. The structure of the information presented in the BAD is schematically presented in Table 1. Fully detailed information regarding the database is presented in the ESI.†

Table 1 Information provided by the BAD and its organisation

Main BAD table/web page

Protein name—link to Protein table

Surface Acronym—link to Surfaces table

Contact angle (°)

Surface tension (mJ/m²)

Buffer—link to Buffer table

Buffer pH

Ionic strength (M)

Surface protein concentration (mg/m²)

Solution protein concentration (mg/ml)

Experiment type

Measurement method—link to Methods table

Reference—id leading to Reference table

Notes—comments on specific entry

Protein table:

Protein name

Source—information about protein source

PDB ID (linked Protein Data Bank Identifier)

Molecular weight (KDa)

Isoelectric point

Protein surface tension (mJ/m²)

Residues: number of residues of the protein

Residues in alpha helices: number of residues in the alpha helices structures

Residues in beta sheets: number of residues in the beta sheets structures

Disulfide bonds: number of disulfide bonds of the protein

Composition of amino acids

% of EACH amino acid in the protein sequence

Protein hydrophobicity

More info: PDF file containing details about the selected protein

Surfaces table:

Surface full name

Surface Acronym and surface abbreviation

Description

Reference: link to reference id and further information

Buffer table:

Buffer ID/acronym

Description, full buffer name

Composition: composition of buffer solution

Reference: link to reference id and further information

Methods table:

Measurement method name

Acronym

Description

Reference: link to reference id and further information

Link to external source related to the relevant methodology

Reference table:

ID—reference identification

Author/s

Title

Year

Source

Link to external source related to the relevant reference

---

2.4. Statistical analysis

The BAD has been tested through the application of regression and neural networks analysis, for the statistical soundness of the variables and overall database extent, respectively.

The piecewise multi-linear regression with breakpoint has been applied to a subset of the BAD representing 420 data points (67% of the overall BAD). This subset comprises data that is fully characterised, i.e., complete protein-, surface- and fluid-descriptors, and excludes substantial outliers, e.g., measurements with the quartz microbalance, protein molecular weights above 150 kDa, extremely high protein concentration in solution. Further, the regression has been applied separately to data representing adsorption on hydrophilic (contact angle lower than 45 deg; 201 cases) and hydrophobic surfaces (199 cases).

The regression analysis used a least-squares penalty function as implemented in the software package Statistica™ (from StatSoft Inc.). Whilst the Statistica package is capable of enabling the algorithm to estimate the breakpoint automatically, manually tuning the breakpoint provided a comparatively better fit. The chosen breakpoint was the value at which the correlation between the observed and predicted values was at its highest. We used several estimation algorithms: quasi-Newton, Hooke–Jeeves, Simplex, Rosenbrock and combination of Rosenbrock pattern search and quasi-Newton. Of these, only the first and last were successfully able to converge to a satisfactory solution and both methods gave practically identical results. The maximum number of iterations was set to 1000 and the convergence criterion was set to 10 (the optimization stops when the changes in the parameters from iteration to iteration are no more than the convergence criterion).

The above subset was then analysed by neural networks procedures, also as implemented in Statistica. First, the 420 dataset described above was divided in three sets: training (50, 100, 150, 200, 250 and 300 data points), selection (or validation, 60 data points), and testing (60 data points). The size of the training, selection and test sets has been automatically suggested by Statistica. For each of these combinations we carried out 10 separate runs, each run fitting several types of neural networks using the above sets. Each run uses sets of identical size, but with different composition, prepared through a random process of selection from the BAD 420-subset. For each run we recorded the results, i.e., training, selection and test error as averages, standard deviations and minimum values, only for the networks that improved during neural network procedure. Each run produced an average of 84 improved networks. Second, we run the same calculation plan but with double the size of test sets (120 data points). Consequently the training set could be only up to 240 data points. Finally, we divided the 420 data set in two quasi-equal data sets (211 and 209 data points), each related to different surfaces, i.e., hydrophilic (up to 45°) and hydrophobic. The same procedures were run as before on these two reduced size data sets, but only up to 150 data points for the training set (the remainder being used for the selection and test sets). These surface-specific neural networks calculations produced an average of 50 and 80 improved networks per training set, for hydrophilic and hydrophobic surfaces, respectively. In all calculations, the protein surface concentration was the continuous output and the continuous inputs were protein concentration in solution, ionic strength, contact angle, absolute difference between the pH of buffer and the isoelectric point of the protein, protein hydrophobicity and its standard deviation. In all calculations, both the linear and logistic regression output encoding have been used. The criterion used to select the retained networks was the balance error against diversity. The types of the selected neural networks are linear, probabilistic, general regression, radial basis function and 3 and 4 layer perceptron.

2.5. Prediction of protein adsorption-relevant parameters

Analysis of different network types and architectures, which used the maximum network performance and minimum error criteria, has revealed that the multiple percepton neural network (MLP) gives the optimal results. For the purpose of the further analysis a 3-layer MLP was chosen for fitting the data for hydrophobic surfaces and a 4-layer MLP for hydrophilic surfaces, respectively, and trained using the back propagation method as implemented in Statistica. The most improved network models were further selected. For the hydrophobic surfaces the 7 : 10 : 1 architecture composed of 7 neurons in the first input layer, 10 neurons in the second hidden layer and 1 neuron in the third output layer was selected with the mean-squared error of 0.031. For the hydrophilic surfaces the 7 : 11 : 11 : 1 architecture was fitted with the mean squared error of 0.025. In both cases the hyperbolic function was used as an activation function in hidden layers and the logistic function in output layers, respectively. The performances of the selected neural networks were further analyzed and implemented in Matlab, which was used as the platform for the prediction routines available on the BAD portal.

3 Results and discussion

3.1. The BAD and the prediction of protein adsorption for lab-on-a-chip design

Despite the interest in protein adsorption for nearly half a century, one could not yet predict with engineering accuracy the adsorbed amount of a particular protein, on a particular surface in a particular fluid environment. Aside from the previously discussed complexity of the process, this inability also has its roots in the evolution of the interest in protein adsorption. At the beginning much of the interest in protein adsorption was mainly theoretical, which resulted in important theoretical advances—many in use even today. Later, the advent of implantable materials, first metallic, then increasingly polymeric, paralleled by the emergence of more precise analytical equipment, shifted the focus from processes to materials and surfaces. More recently, the development of biomedical microdevices, e.g., biosensors, bioMEMS, micro- and nanoarrays and lab-on-a-chip devices, comes with new interests, challenges and opportunities. While the BAD is intended to help the practitioners involved in any of the above mentioned areas, i.e., colloid and surface science, biomaterials and biomedical microdevices, we expect that the latter community will benefit the most. Indeed, in microfluidics systems the flow of the fluids is well within the laminar regime, and this would make the design of lab-on-a-chip devices trivial if not for a number of complications, of which two could be critical:

(i) Change in device geometry. In small sized features (micro-channels, corners, etc.) the attachment of the biomolecules on the walls of the device could induce important changes of their dimensions, at least relatively, thus changing the hydrodynamic resistance to flow in the device.

(ii) Change in surface properties. In laminar regime essentially all the hydrodynamic resistance is concentrated at the walls and this resistance is used to estimate the power required to operate the microfluidics device. But what are the walls? When biomolecules, mostly proteins, attach parasitically and indiscriminately on the walls the roughness of the wall and its nature change, at times dramatically. Then the hydrodynamic resistance changes as well.

The BAD and the BAD-based prediction of protein adsorption could help the lab-on-a-chip designers to address these difficulties.

3.2. Comprehensive mapping of protein adsorption experimental ‘space’

An alternative approach regarding the prediction of protein adsorption, entirely opposite to the approaches based on models, either molecular or thermodynamic, is to comprehensively map the multidimensional experimental ‘space’, i.e., the [protein parameters] × [surface parameters] × [solution parameters]. A similar comprehensive, largely parallel experimentation has been routinely used for more than a decade in the form of combinatorial chemistry for drug discovery,^16,17 but also for materials discovery.^18,19 Although not explicitly combinatorial, this approach has been attempted, most notably when interfacing representative proteins with ‘model’ surfaces made of self-assembled monolayers (SAMs).²⁶ The complementary effort to combinatorial techniques, denominated as Quantitative Structure-Activity Relationship (QSAR)²⁷ has been also attempted, but in an ‘inverse’ manner, i.e., using neural networks techniques²⁸ to optimise the properties of polymers for fibrinogen adsorption, rather than predict the protein adsorption on any given material.

Decades-long research regarding protein adsorption would have offered hope that enough agreement was reached following the large amount of experimental data generated. The—so far—intractable difficulty of the process of protein adsorption, as well as its ubiquitous importance, generated a large amount of empirical data scattered through several decades of scientific literature. This contribution aims to collect, filter and organize this large amount of this unintended combinatorial data regarding protein adsorption in a database that can be used for further QSAR-orientated studies.

3.3. Distribution of the protein adsorption descriptors in the BAD

Proteins. The distribution of the molecular weights and of the isoelectric points of the proteins represented in the BAD is shown in Fig. 2a and 2b, respectively. While the proteins represented in the BAD are small, medium and large sized, the large majority of data (approximately 72%) are cases for proteins with small molecular weight (between 6 and 85 kDa). This distribution is the result of the over-representation of albumin, lysozyme and fibrinogen. A very small percentage of cases refer to proteins with high molecular weights, (e.g., 0.8% for immunoglobulin). The distribution of the cases versusisoelectric points of the proteins is also the result of the overrepresentation of lysozyme (IP = 11) and albumin (IP = 4.7) and fibrinogen (IP = 5.4), the latter two resulting in a larger cluster for IP between 4 and 7.

Fig. 2 Distribution of the properties of the proteins present in the BAD: (a) molecular weights, (b) isoelectric points.

The molecular weights, isolectric point, hydrophobicity and surface tension have been estimated from the amino acid descriptors, weighted by the relative amino acid composition of the protein. The molecular weight and the isoelectric point are parameters that are independent of the tertiary and quaternary structure of the protein and therefore their estimation from the primary structure is uncontroversial. The protein hydrophobicity and its surface tension are however properties that should manifest on the molecular surface of the protein, at least as adsorption on surfaces is concerned. Consequently, their estimation should be—ideally—derived from the molecular surface, which in turn depends on the tertiary and quaternary structure of the respective protein. Alternatively, all protein descriptors can be included in the BAD as reported in the literature. Despite all of these considerations, the estimation of the protein descriptors as implemented in BAD was dictated by practical considerations. First, the protein adsorption literature very rarely reports the molecular weight and isoelectric point of the protein used, and almost never its hydrophobicity and surface tension. Second, the unavailability of PDB structures for a large proportion of cases in the BAD would have drastically reduced the size of the statistical population and compromised the quality of the prediction of protein adsorption. Third, even if the molecular structure of the proteins were available, the quantum mechanics calculations will be prohibitive even for medium size systems. Fourth, the molecular weight and the molecular surface-related parameters are statistically correlated: at higher molecular weights the overall hydrophobicity will be lower than at lower molecular weights, due to the creation of the hydrophobic core. For all these practical reasons, the data as implemented in the BAD used the primary structure for the estimation of protein descriptors. Further developments, e.g., availability of more protein PDB structures, estimation of protein properties at sub-molecular scale, will be constantly implemented and thus improve the quality of the BAD-based predictions.

Surfaces. From the histogram in Fig. 3, which shows the distribution of the contact angle data of the surfaces present in the BAD, it is evident that the hydrophobicity of the adsorbing surfaces aggregate in two distinct clusters: hydrophilic surfaces with contact angles between 0 to 45°; and a flatter cluster with contact angles between 70 and 120°, with few cases bridging these two main clusters.

Fig. 3 Distribution of the contact angle of surfaces in the BAD.

Fluid media. The histogram in Fig. 4a represents the distribution of pH of the entries in the BAD. The large majority of reported experiments (700 cases) were performed in the neutral pH region (pH = 6.6–7.7) with the total pH range spanning from 2.75 to 11. In the absence of any information regarding the charging of the adsorbing surfaces, we preferred to record the absolute difference between the isoelectric point and the pH of the buffer, with the distribution presented in Fig. 4b (mean value = 2.31). The distribution of the ionic strength of the buffer (presented in the Fig. 4c) indicates the experimental preference for either buffers with very low content of ions, or an ionic strength around 0.15. The overall ionic strength range spans from 0.001 to 0.42.

Fig. 4 Distribution of the parameters of the fluid media in the BAD: (a) pH of buffers, (b) difference between pH of buffers and isoelectric points of proteins, (c) ionic strength of buffers.

Protein concentrations. The distributions of the protein concentration in solution and on the adsorbing surface are presented in Fig. 5a and 5b respectively. As expected, much of the data in the BAD is a reflection of protein adsorption experiments at low concentration in solution (91.4% up to 2 mg/ml), and consequently a lower concentration of the surface (79.43% up to 5 mg/m²). We should note however that the overall range of protein concentration in solution and on the adsorbing surface spans several orders of magnitude.

Fig. 5 Distribution of proteins concentration: (a) in solution, (b) on the surfaces.

3.4. How representative is the BAD?

A question relevant to the prediction of protein adsorption based on the BAD is whether this database is large enough to be statistically representative for the included proteins. Neural networks and regression analysis can help to provide an answer to this question.

First, for the neural network analysis applied to the whole BAD (420 cases; all taken from adsorption isotherm experiments) the error of estimation (test error) of the best neural network reaches a minimum (around 7%) around 150–200 data points, which suggests that a few hundred data points suffice to produce a reasonably accurate prediction for the proteins represented in the BAD. Even more interesting is the observation that less data points (a result of the splitting the data in two sub-sets) are needed to achieve the same or lower errors for the sub-sets describing protein adsorption on hydrophilic and hydrophobic surfaces, respectively. The results of the neural networks analysis (details in the ESI†), strongly suggest that the BAD is large enough to be representative and that further improvements would be the result of additional data for other proteins, surfaces and fluids, and/or more precise measurements, rather than duplication of the present data.

Second, the regression analysis using linear regression with break point resulted in good statistical fits, i.e., a correlation coefficient, R2 of 84.75%, 84.33% and 86.68% for the whole BAD, the hydrophilic, and the hydrophobic surfaces database subsets, respectively. This good fit strongly suggests that the input variables (protein-, surface- and fluid-descriptors; and protein concentration in solutions), as they are represented in the BAD, are in good correlation with the output variable (protein concentration on the surface). These regression results (presented in the ESI†) are not appropriate however to devise relationships that can be used for the prediction of protein adsorption, because, for multiple linear regression, the position of the breakpoint is obtained at the intersection between the regression planes rather than lines—and this does not usually result in a unique point in the multidimensional descriptors “space”. Conceptually, these results suggest that the BAD is also representative with regard to the correlation between the protein adsorption parameters.

3.5. Prediction of lab-on-a-chip design parameters

Prediction of the amount of adsorbed protein. Further optimisation of the neural networks allowed the identification of a set of networks that accurately predict the amount of protein adsorbed on the surface. The quality of the neural networks-based prediction is remarkable: correlation coefficients of 0.95 and 0.97, for data regarding hydrophobic and hydrophilic surfaces, for comparisons forcing the regression through the origin (more detailed information in the ESI†). These neural networks-based routines are implemented in the BAD, where the user can estimate the amount of adsorbed protein when specifying the protein name, solution pH, isoelectric point, water contact angle and protein concentration in solution. It should be noted however that this accuracy could be the sign of over fitting by neural networks, in particular for proteins that are under-represented in the BAD. Consequently, the user should check at all times the validity of the predictions, e.g., attempting to build a Langmuir-type isotherm curve.

Prediction of the thickness of adsorbed protein layer. The development of increasingly complex lab-on-a-chip devices, with an increasing level of integration and density of microfluidics components per chip area, translates in the shift from micro- to nano-fluidics structures. Before getting to the point where nanofluidics are required to handle single, but large biomolecules,^29–31 the narrowing of the channels following the deposition of the adsorbed biomolecules on the walls start to have a sizeable impact over the flow and consequently the design of the chip for nominal widths below 2 µm.²¹ The prediction of the amount of the adsorbed protein allows for the estimation of the thickness of the protein layer adsorbed on surfaces. This procedure, described in detail in the ESI,† is implemented in the BAD as an applet that allows the user to estimate the protein layer thickness as a function of protein-, surface- and fluid-descriptors. While the vast majority of the data in the BAD represents protein layer thicknesses up to one monolayer, the prediction can estimate higher values. The values estimated for protein layer thickness are minimum values, as we assumed the closest packing of proteins and ignored the inherent uptake of water in the protein layer.

Prediction of the surface tension of protein-covered surfaces. While the utility of the prediction of protein layer thicknesses might be restricted to the design of nano-sized channels, the impact of the change of surface tension of the surfaces of the micro/nano-fluidics structures following protein adsorption should have a wider impact on lab-on-a-chip design. The predicted change of the surface tension of the material the lab-on-a-chip device is made of, during its operation, should provide optimum designs for the whole life of the device. Similarly with the two predictive tools described above, the prediction of the surface tension of protein-covered surfaces is implemented as an applet of the BAD website, which allows the estimation as a function of protein-, surface- and fluid descriptors.

We need to qualify the prediction of the surface tension of the protein-covered surfaces. The estimated surface tension of protein-covered surface is calculated from the contributions of the bare surface and the pure protein, weighted according to their relative coverage of the total surface. The estimated surface tension of the pure proteins (by a procedure detailed in the ESI†) has provided results that are lower than the experimental data published in the literature,³² which reports the surface tension of pure proteins adsorbed on the surface (full coverage) from saline contact angle measurements on thick layers of the hydrated proteins. This difference can be explained by the fact that while our estimation of the surface tension of pure proteins takes into account all amino acids, the hydrophobicity of the protein manifests more on the molecular surface of the protein. Furthermore the amino acids present on the protein molecular surface are likely more hydrophilic than those inside the protein core. Also our prediction could not include the inherent uptake of the hydration water. With all these qualifications, the estimated surface tension of the surfaces of lab-on-a-chip devices in operation is certainly closer to the working reality, and therefore should be preferred in the design calculations to values for bare surfaces.

3.6. Perspectives and future work

This contribution presented a database regarding the protein adsorption and—using a sub-set of the database—also a predictive method to estimate the parameters relevant to the lab-on-a-chip design. A more comprehensive database and a more accurate prediction could have been produced, if not for several problems related to the measurement and reporting of published protein adsorption data. Despite these shortcomings, the present and the near future offer many opportunities to greatly improve the quality of the database and prediction of protein adsorption, as follows.

Standardisation. The literature on protein adsorption is very large, with the major literature databases listing several thousands of titles with “protein adsorption” as a key word, e.g., PubMed 1800; Scopus 4000; ISI Web of Knowledge 6000, etc. Despite this enormous body of knowledge and despite several years of the authors' filtering through thousands of articles, only a very small fraction of the total literature could be useful for the BAD. The main difficulty in integrating this large amount of data lies in the under-reporting of the experimental conditions, i.e., protein, surface and fluid environment conditions. While the protein descriptors are perhaps more difficult to quantify, the surface, and certainly the fluid environment descriptors are far easier to report. Mutually agreed guidelines, if not standardization, similarly with the common practice for other databases (e.g., PDB) would greatly improve the capacity of the scientific community to adsorb the enormous experimental knowledge regarding protein adsorption. Our on-line data submission portal proposes a tentative guideline for reporting protein adsorption data (see ESI†).

Protein descriptors are the most under-reported set of parameters related to protein adsorption, despite their general availability in multiple databases, e.g. PDB, FASTA, etc. Should these descriptors be reported as a matter of common practice, the relationship between protein characteristics and protein adsorption would become much easier to characterise and quantify. Furthermore, the present work correlated the protein adsorption (outputs) with protein descriptors (inputs), but these have been calculated from the protein primary structure rather than from the properties on the molecular surface. It would be expected that the properties manifested on the protein molecular surface, in particular the hydrophobicity, would have a stronger relationship with protein adsorption (although some protein descriptors, e.g., the charges, would be parameters that are ‘bulk’-based rather than molecular surface-based). Consequently, the quantification of some of the protein descriptors on its molecular surface almost surely would improve the prediction of protein adsorption. However, in order to reach this desideratum, new algorithms and methodologies need to be developed, as the present quantification of the molecular surfaces is tailored for drug discovery applications, i.e., using high resolution (water or solvent molecule, Å-range) probing of the protein surface, while protein adsorption is essentially the result of the interaction between the protein surface and adsorbing surface at—relatively—low resolution (surface nanotopography, few nm-range).

Surface descriptors are also under-reported in the literature. This is only partially justified by the experimentally-problematic measurement of the contact angle, e.g., its dynamic nature, but these measurements are relatively trivial compared with the measurements of protein concentration of the adsorbing surface. The development of biomedical microdevices, in particular micro/nanoarrays and lab-on-a-chip devices, makes the impact of the surface more relevant than ever. Several developments are critically needed for a better characterisation and prediction of protein adsorption. First, the surfaces will need to be fully characterised from the electrical point of view, e.g., point of zero charge (pzc), zeta potential, especially in the context of lab-on-a-chip applications. Second, the nanotopography of the surface, often in the range of several nanometres, is likely to interact at the molecular scale with the proteins that have similar dimensions, especially in solutions with low concentrations—as is the case for most biomedical microdevices. Somehow related, the extended exposure of the polymer surfaces to the fluid media surface can cause changes in the macromolecular structure, in which case the protein-surface interactions become even more complex. Third, a full description of the chemistry, which was available in the BAD only for a small number of cases, would allow the automatic calculation of the physico-chemical properties of the adsorbents (e.g., surface potential, surface charge density, pI (or pzc), Hamaker constant). This capability will be added in future versions, in connection with web-based programmes (e.g., MODEL–Molecular Descriptor Lab;³³ Polymer Database PoLyInfo³⁴ and other methodologies used to estimate these properties from the chemical structure of polymer materials.^35,36 Fourth, as the immobilisation of proteins on the surface is increasingly done for the benefit of ultra-sensitive, single molecule detection devices, there is an imperative need to fully characterise the surface morphology and physico-chemistry with nanometre-level resolution. The mature development of Scanning Probe Microscopies, in particular Atomic Force Microscopy, makes this desideratum an achievable target.

Fluid descriptors are usually the easiest to report, but are still rarely fully described. Lab-on-a-chip applications in particular, where for instance the ionic strength of the buffer could be a critical element in the design of appropriate fluid flow, require the full characterisation of the fluid environment conditions. Whenever possible the reporting of the correlation between the flow conditions and the structure or bioactivity of the immobilised proteins will be of particular relevance to lab-on-a-chip applications.

Combinatorial experiments in microfluidics. Most, if not all of these challenges and additional requirements are counterbalanced by the opportunities offered by lab-on-a-chip devices.^37,38 Conceptually, the large number of the combinations of different surfaces would suggest running combinatorial experiments in a microarray format, but the long contact times between the surface and the protein solution, as well as the large number of combinations of the fluid (protein concentration, pH, ionic strength) recommends a microfluidics coupled with a microarray approach. Moreover, several methods for the absolute measurement of the mass of adsorbed proteins, e.g., ellipsometry have been miniaturised together with inherently microscopy based methods (e.g., confocal microscopy). Indeed a combinatorial experiment using modern scanning ellipsometry, which requires an approximate 200 × 200 µm footprint, would result in approximately one thousand experiments on 1 cm². Certainly such a combinatorial experiment would require nontrivial logistics, design, operation of the chip, and data management of the information, but to put things in perspective this experiment alone would more than double the amount of protein adsorption experiments recorded in the BAD.

Conclusions

While protein adsorption at solid–liquid interfaces is important to many applications, but its impact to the design and operation of lab-on-a-chip devices is critical due to the small dimensions involved and due to the transient regime in which these devices usually operate. Despite the general interest and a large amount of research in the last half a century, the complexity of protein adsorption phenomenon precluded a design-orientated prediction. The online, free-access Biomolecular Adsorption Database (BAD) presented here aims to alleviate this gap in the engineering knowledge. Furthermore, the data present in the BAD allowed for the derivation of predictive tools that can estimate the amount of adsorbed protein, the thickness of the adsorbed protein layer, and the surface tension of the protein-covered surfaces. Lab-on-a-chip device design will require further advances in protein adsorption research, but equally can contribute to an efficient experimentation plan for the benefit of all interested areas.

Contributions of the authors

ENV compiled and upgraded the data from the open literature and performed physico-chemical calculations; EP designed and implemented the BAD webpage, including the online prediction tools, and performed the neural networks analysis; DVN Jr contributed to the initial concept, the design of the database, including online versions, and contributed to the regression and neural networks analysis; DVN designed and directed the project, performed the regression analysis, and contributed to the neural networks work.

Acknowledgements

The research reported here was sponsored over the years by grants from Defence Advanced Research Projects Agency (DARPA), Australian Research Council (ARC) and European FP6 Programme. DVN wishes to acknowledge the initial work of Dr Jinan Cao, Florin Fulga and Gerardin Solana from Swinburne University of Technology; and the very fruitful discussions with Dr Anantha Krishnan of DARPA and Prof. Mauro Ferarri from Ohio State University. The authors would like to thank Kelly Dorgan for many comments and suggestions that have considerably improved the manuscript.

References

1. R. Braun, M. Sarikaya and K. Schulten, J. Biomater. Sci., Polym. Ed., 2002, 13, 747–757.
2. T. E. Cheatham III and P. A. Kollman, Annu. Rev. Phys. Chem., 2000, 51, 435–471.
3. M. F. O'Donohue, E. Minasian, S. J. Leach, A. W. Burgess and H. R. Treutlein, J. Comput. Chem., 2000, 21, 446–461.
4. J. Garcia De La Torre, M. L. Huertas and B. Carrasco, Biophys. J., 2000, 78, 719–730.
5. O. Byron, Biophys. J., 1997, 72, 408–415.
6. J. Liang, H. Edelsbrunner, P. Fu, P. V. Sudhakar and S. Subramaniam, Proteins: Struct., Funct., Genet., 1998, 33, 1–17.
7. R. C. Chatelier and A. P. Minton, Biophys. J., 1996, 71, 2367–2374.
8. M. Malmsten, J. Colloid Interface Sci., 1998, 207, 186–199.
9. F. Höök, J. Vörös, M. Rodahl, R. Kurrat, P. Böni, J. J. Ramsden, M. Textor, N. D. Spencer, P. Tengvall, J. Gold and B. Kasemo, Colloids Surf., B, 2002, 24, 155–170.
10. P. R. Van Tassel, Materialwissenschaft und Werkstofftechnik, 2003, 34, 1129–1132.
11. P. Déjardin and I. Cottin, Colloids Surf., B, 1995, 4, 111–120.
12. M. A. Brusatori and P. R. Van Tassel, J. Colloid Interface Sci., 1999, 219, 333–338.
13. M. R. Oberholzer and A. M. Lenhoff, Langmuir, 1999, 15, 3905–3914.
14. A. Ladiwala, K. Rege, C. M. Breneman and S. M. Cramer, Proc. Natl. Acad. Sci. U. S. A., 2005, 102, 11710–11715.
15. T. A. Horbett and J. L. Brash, eds., Proteins at interfaces II: Fundamentals and applications, American Chemical Society, Washington, DC, 1995.
16. M. A. Gallop, R. W. Barrett, W. J. Dower, S. P. A. Fodor and E. M. Gordon, J. Med. Chem., 1994, 37, 1233–1251.
17. E. M. Gordon, R. W. Barrett, W. J. Dower, S. P. A. Fodor and M. A. Gallop, J. Med. Chem., 1994, 37, 1385–1401.
18. J. C. Meredith, A. P. Smith, A. Karim and E. J. Amis, Macromolecules, 2000, 33, 9747–9756.
19. R. R. Bhat, M. R. Tomlinson and J. Genzer, J. Polym. Sci., Part B: Polym. Phys., 2005, 43, 3384–3394.
20. H. M. Berman, J. Westbrook, Z. Feng, G. Gilliland, T. N. Bhat, H. Weissig, I. N. Shindyalov and P. E. Bourne, Nucleic Acids Res., 2000, 28, 235–242.
21. D. V. Nicolau Jr, F. Fulga and D. V. Nicolau, Biomed. Microdevices, 2003, 5, 227–233.
22. W. R. Pearson and D. J. Lipman, Proc. Natl. Acad. Sci. U. S. A., 1988, 85, 2444–2448.
23. S. H. White and W. C. Wimley, Biochim. Biophys. Acta, 1998, 1376, 339–352.
24. E. L. Decker, B. Frank, Y. Suo and S. Garoff, Colloids Surf., A, 1999, 156, 177–189.
25. R. Beynon and J. Easterby, Buffer solutions, Taylor & Francis, 2003.
26. A. Sethuraman, M. Han, R. S. Kane and G. Belfort, Langmuir, 2004, 20, 7779–7788.
27. R. Perkins, H. Fang, W. Tong and W. J. Welsh, Environ. Toxicol. Chem., 2003, 22, 1666–1679.
28. J. R. Smith, V. Kholodovych, D. Knight, J. Kohn and W. J. Welsh, Polymer, 2005, 46, 4296–4306.
29. R. Austin, Nat. Nano, 2007, 2, 79–80.
30. A. van den Berg and M. Wessling, Nature, 2007, 445, 726–726.
31. R. Mukhopadhyay, Anal. Chem., 2006, 78, 7379–7382.
32. C. J. Van Oss, D. R. Absolom, A. W. Neumann and W. Zingg, Biochim. Biophys. Acta, 1981, 670, 64–73.
33. Z. R. Li, L. Y. Han, Y. Xue, C. W. Yap, H. Li, L. Jiang and Y. Z. Chen, Biotechnol. Bioeng., 2007, 97, 389–396.
34. Y. Takaeda and K. Yagi, Polym. News, 2003, 28, 352–355.
35. D. W. Van Krevelen, Properties of polymers, 3rd edn., Elsevier, Amsterdam, 1990.
36. A. A. Askadskii, Computational materials science of polymers, Cambridge International Science Publishing, 2001.
37. A. Lionello, J. Josserand, H. Jensen and H. H. Girault, Lab Chip, 2005, 5, 1096–1103.
38. K. R. Hawkins, M. R. Steedman, R. R. Baldwin, E. Fu, S. Ghosal and P. Yager, Lab Chip, 2007, 7, 281–285.

---

Footnotes

† Electronic supplementary information (ESI) available: Supplementary Information on 44 references of BAD; detailed description of database organisation; quality of Neural Networks analysis and prediction; prediction of protein layer thickness; prediction of surface tension of protein covered surfaces. See DOI: 10.1039/b813475h

‡ ENV, EP and DVN Jr contributed equally to this work.

---