There is one expression in scientific and commercial visualization that is heard more often than anything else: Virtual Reality. The sound of threse words evokes notions of cyberspace and visual landscapes unseen by man, promising the solution of all visualization problems by the fully immersive manipulation of information objects giving insights impossible to achieve by any other means. However, probably no other words from this field have also been so heavily abused. If one follows advertisements and also quite a number published papers, about every display program which employs some 3D rendering is now adorned with the magic words Virtual Reality (VR). This includes applications which conceptionally do not go beyond standard 3D graphics, which have been routinely used in molecular modeling and similar fields for more than two decades. So if virtual reality is supposed to be more than simple molecular graphics, what is it?
Applications of true virtual reality character have been implemented and practically tested with chemical problems. However, these applications are leading edge research and relatively scarce. A review of these few systems would not fill many pages. So in order to broaden the perspective of this review, we decided to relax the rigid definition of Virtual Reality, and include a wider range of innovative visualization approaches which employ ingenious 3D metaphors and innovative ways to generate and manipulate graphical objects. These applications earn their special character and appeal because they help to generate insight and knowledge by the manipulation of graphical objects, even if a normal color monitor is used as display device instead of a full VR setup. With this broadened definition of Virtual Reality, an interesting and inspiring set of visualization methods for a wide range of application fields is covered. We hope the following overview and selection of examples will give a good impression where VR in chemistry is headed at the dawn of the next millennium.
Most of the time the term Virtual Reality pops up, it is in the context of VRML, the Virtual Reality Modeling Language. However, while VRML can be a tool for VR, there is no direct relationship between VR and VRML, and certainly not every VRML visualization qualifies for the VR attribute, even in loose definitions. Nevertheless, VRML is such an important topic that a few comments about this technology and its chemical applications are in place.
VRML [2] is a portable, platform-independent and flexible file format for the transport of 3D graphical information, not more. There are two main variants of VRML: The old VRML 1.0, which could only be used to transport static scenes, and VRML 2.0 [3], which adds a whole world of animation and multimedia to VRML. Strictly speaking, VRML 1.0 and VRML 2.0 are not compatible, but all VRML 1.0 features can be expressed in VRML 2.0 syntax, and most VRML 2.0 browsers still support VRML 1.0. VRML is now developed by an international consortium [4] and many of the major computer manufacturers and software houses have become members.
VRML scenes are built in a tree-like fashion from basic building blocks, which specify lighting, camera (viewer) positions, geometrical primitives such as spheres, lines and surface patches, materials and textures mapped on the surface of geometric objects, and geometric transformation nodes. The scene graph is traversed from the root. VRML scenes are viewed either by stand-alone programs, which can be used as helper applications for Web browsers, or increasingly by plugins [5], which make it possible to render VRML scenes in windows which are embedded in HTML pages. VRML viewers and plugins are available for all major computer platforms. This can be attributed to the fact that VRML is not specific to chemical or scientific visualization but is also recognized as a valuable medium for on-line commerce and games. Browsers support different scene navigation modes which are suitable for different application areas. For chemical visualization, simple external rotation and zooming of the full scene is normally more useful than walk-through modes with gravity and collision detection, which are advantageous for architectural models etc.
The biggest progress in VRML 2.0 for scientific applications, compared to the previous version, is scriptability. In VRML 1.0, the relative geometric positions of node objects were fixed. VRML 2.0 introduces a whole set of new techniques for animation. [12] Among these are interpolator nodes, an event routing mechanism to make objects react when certain actions are performed by the user, collision detection among objects and, most important, script nodes. Script nodes can contain nearly arbitrary instructions in VRMLscript [13], a JavaScript variant ([14], soon to change the name again to ECMAscript [15]). A Java interface (EAI, External Authoring Interface, [16]) to control the status and behavior of VRML nodes has also been defined. However, on the Java side, the Java3D [17] API (Application Programming Interface), which has no relationship to VRML, is being touted as a better integrated and potentially faster alternative to VRML. A serious struggle for dominance between these two standards for animated, interactive Internet-based graphics should be expected as soon as full Java3D implementations become available.
For many application areas it is desirable to have script-animated 3D scenes which can be viewed from freely chosen viewpoints, or provide guided tours highlighting points of special interest in complex structures such as proteins. Video sequences, which have been used with some success to visualize chemical processes, do not provide the freedom to select the viewing position. Also, video does not offer any choice to remove parts of the scene for a clearer view or to change display attributes of scene objects, which are both useful features to clarify the image of chemical structures.
The next level of sophistication uses shutter glasses. These are transparent LCD panes which can be blackened separately for each eye by applying a voltage. The voltage is synchronized, typically by an IR signal, with the video display, which switches in rapid sequence between two different views of the scene, one for each eye. In contrast to parallel stereo images, the display can use the full area, and the head can be moved. Of course the change of the images must be fast enough to avoid flicker, and the display bright enough to produce a well-lit impression if viewed only half of the time by each eye.
Head-mounted displays (HMD) and binocular omni-oriented monitor (BOOM) systems do not use a single screen any longer. They consist of two CRT or LCD screens, one for each eye. A HMD is worn like a helmet with a face cover. The user movements are tracked by auxiliary devices so that the display generator is always informed about the head position and view direction. A BOOM is connected via a system of rods to a fixed pedestal. At the end of the extension, a view box with the screens is located. When the user pushes, pulls, tilts or otherwise moves the view box, not unlike a periscope, his/her current location and orientation is reported by sensors measuring the angle of the joints connecting the rods.
One of the most spectacular, but also most expensive setups for virtual environments is the CAVE (Cave Automatic Virtual Environment, a recursive acronym [28]). The first CAVE was installed in 1992. A CAVE is a box, measuring about three meters on all sides, installed inside a larger, dark room. The walls, floor and potentially also the ceiling consist of rear-projection screens. Images are projected from the outside by video projectors. The user steps inside the cave and wears shutter glasses for an additional enhancement of the scenic impression. More than one user can be inside at a time, but only one user determines the viewpoint, the others are taken along as in a roller-coaster ride. CAVEs usually also include 3D-audio processing equipment, so directional sonic stimuli can be used to support the 3D viewing experience.
Computer-generated holograms (CGH,[32]) are another 3D imaging technology which may become more widely used in the future, because no headgear is necessary for viewing. Computing the interference patterns which form a monochromatic or white-light hologram is not trivial but has been done even in real-time on massively parallel computers. The main problem lies in selecting suitable output devices. The resolution of the patterns on film or other media must approach the wave length of light the make the interference pattern work. Common output equipment such as printers, monitor screens or slide exposers provide resolutions which are magnitudes too low. Creating holographic prints or films currently involves using tools such as modified scanning electron microscopes to write the patterns with a tight electron beam onto film [33], or acousto-optic modulators, where ultrasonic waves are transmitted into a crystal whose refractive index is locally changed by compression. A wide laser laser beam shining into the crystal experiences interference. [34] Holography is the only imaging technique which gives all visual depth cues. Stereoscopic images to not provide ocular accommodation, and volume displays cannot provide occlusion. Holograms of chemical structures exist, but have not yet received recognition beyond showcase pieces.
Keyboards and mice do not adapt well to the immersiveness of VR, so advances in voice, gesture, and touch recognition will be essential to facilitate user input. Input elements which have been proven to work well in 3D environments are for example Space Mice and Wands. A Wand is basically a joystick plus a set of buttons, but in addition it transmits is position and orientation is 3D space to a receiver via radio signals. User and wand positions are registered by sensing disturbances of the electromagnetic field caused by pulsed magnetic coils in the device, or by receiving ultrasonic or radio impulses from the devices via multiple sensors or antennae. Tracking the user position is important because in projection-based VR, the positions and view direction of the head must be known to adapt the projected scenes.
Reality, virtual or natural, is of course more than a visual image of a scene, although it has been estimated that more than 90% of the volume of the data streaming into the cortex is coming from the retinae. If an attempt is made to model reality, other senses should not be excluded and could provide valuable auxiliary data channels for the representation of complex data relationships requiring a maximum bandwidth input. This is really an area of active research, but its results seem not to have yet applied in the chemical domain, except sonification.
Immersive VR for scientific visualization has evoked a lot of interest, [1]. A review book which contains some information about this subject has been published. [43] In order to attain the virtual reality effect, the system must deliver very high graphics performance. Specifically, requirements for a usable VR environment include a user feedback response time of less than 0.1 s for fast, accurate manipulation of the environment, a frame rate of at least 10 frames per second to avoid annoying flicker and image jumps, and finally the application environment must contain objects if a sufficient level of fidelity to allow the performance of meaningful tasks. For practical applications this means that for datasets of scientifically interesting complexity for the visualization part alone nothing short of multi-processor SGI Reality Engine as renderer alone will suffice to produce the desired level of detail and response time. Additionally, extra processor power must be assigned to model the system response.
The hardware price tag of the equipment used in these studies is usually impressive. In one representative study, two SGI Onyx (20 and 8 processors, 2 and 3 Reality Engines) for the graphics in two CAVEs, an IBM SP2 massively parallel computer for the docking computations, and an SGI Indy for voice processing. This setup was barely able to generate about 10 frames per second, although it must be mentioned that as an additional complication the data (geometry, tracking, video audio) was exchanged over a high-speed network between two sites in order to demonstrate the feasibility of remote collaboration in such environments.
In this study [44], which was carried out at the Cornell Theory Center, the docking of chlorpromazine and mepacrine with the trypanothiophane reductase of Trypanosoma cruzi, a parasite causing Chagas` disease, was studied. The primary focus of this project however appears to have been a test for the WorkSpace toolkit used to implement the 3D user interface with its various manipulatable objects such as molecular models, directories, video panes showing the collaborators at the other site and models of the parasite and its vector, a small bug Triatoma infestans. Figure 1 shows a snapshot from the modeling session.
The combination of visualization and molecular dynamics is very popular and found in most of the state-of-the-art chemical immersive VR setups. The rationale is that these systems enable molecular scientists to have a visual and auditory experience of a chemical system while manipulating its physical properties by steering, in real time, a simulation experiment on a supercomputer. As a result, scientists are immersed in a realistic representation of a chemical system. Most traditional molecular modeling environments are limited to qualitative information that is obtained from static 3D models. However, complete understanding of molecular interactions requires knowledge of both the dynamic and static features of molecular systems. This is achieved by merging real-time molecular dynamics computer simulations, which can be very computationally expensive, with an immersive display of the chemical system.
Finally, VR in chemistry is no longer confined to research. Pauling World [54] is the first chemical VR environment which was exclusively designed to teach challenging concepts in science to students. Pauling World runs on a 4-processor Onyx Reality Engine with a HMD, a magnetic positional sensor, stereo sound and a custom vest which delivers haptic sensations. Pauling world is only one virtual world from a larger collection of educational worlds which form the ScienceSpace. The rendering capabilities of Pauling World are currently limited to various rigid standard molecular display forms (ball& stick, wire frame etc.) which can be opened and closed to icons by pointing with the index finger. At the time of writing, there appears to be no other method of interaction or animation. The didactic background and the reason for the use of definitely expensive equipment to deliver conceptionally simple visualizations which do not go beyond those of standard 3D molecule viewers which run on every PC, is not evident to us, in spite of having access to about a dozen full-text publications from the Web site. The other worlds of this project offer significantly more interaction and VR experience, justifying the VR effort, so there may be interesting developments in the future.
Few researchers are so lucky to have access to a CAVE or similar equipment to experiment with immersive modeling. This type of environment is very demanding both from the visualization side and from the force field responsible for the structure optimization, and thus expensive. However, smaller systems which use comparable algorithms and run on standard computer hardware exist. They rely on a normal computer monitor for output, use ingenious markers and other visual hints to compensate for the less perfect three-dimensional impression and employ a radically simplified force field to provide interactive response times for molecules up to a few hundred or thousand atoms. This class of programs is probably the first application in chemical modeling where the primary information gain is from the direct 3D interaction with molecular structures. In these programs, the visualization is not simply a display of some final computational results, but the graphical representation and the user interacting with it is the continuous source of new information.
The best known among these programs is Sculpt [55][56], now available in a commercial version from Interactive Simulations. [57] In a Sculpt modeling session, the user pulls atoms with the mouse, for example while attempting to dock into a receptor pocket, and the program continually minimizes the potential energy of the structure. Visual cues are provided by highlighting good and bad van der Waals or electrostatic interaction as cup-shaped symbols, so the user sees immediately why the energy goes up or drops during a specific operation. In contrast to many classical fitting programs, no isolated atom positions or bond angles are changed. The whole molecule is continually reshaped according to the forces originating from the pull point and atomic interactions. Figure 4 is a snapshot of Sculpt modeling.
Information is at the heart of all chemical research. Results are generally not produced from a void, but by expanding existing knowledge. No project can be started without obtaining precise and complete information about related work which has been performed (and possibly patented) elsewhere. At the core, the whole process is concerned about finding comparatively few documents from an enormous body of literature, patents and other information sources, typically still in printed form. Information publishers and information consumers can both profit from improved access paths. If people fail to understand the full range and nature of information available, this can be costly and damaging for both sides, especially in an industrial environment. The established big abstract databases and full-text sources fail to deliver user-friendly advanced searching and navigation. Their typical access and query methods are cumbersome and no longer state of the art.
Document navigation is of course a topic which is of central interest for digital publishing and consequently for this journal. In their daily work, chemists rely heavily on access to documents, and the search for relevant literature can consume a significant part of the time allotted to a project. Digital Internet-based publishing makes the access to the full-text information easier, but does not necessarily provide a solution to the problems of finding the right papers in the first place. While the hyperlink model works well enough for linking digital documents to hypermedia attachments, and to build a hierarchy of access paths leading from the homepage of a publisher via its individual journal pages, indices of numbered issues to the single paper, this access path is very much geared to the subscriber who peruses every issue of a journal or the researcher who has obtained a citation and knows precisely which paper he or she wants to retrieve. Navigation within concept space is completely different. While it is certainly possible to locate interesting papers by author, keywords, or full-text search on the abstract or the complete textual corpus, there is always the danger of missing interesting papers due to nomenclature differences, missed keywords, unknown authors or boundaries between different sciences or fields within a science. Algorithms exist which cluster documents by statistically derived word/content relationships [58] or by analyzing Web linkages [59], and will be able to find relevant documents even when nomenclature varies, provided that documents are stored in the database which builds a bridge between different concept spaces and nomenclature systems. While it is possible to provide a textual query interface to this type of system [60], navigating a multidimensional document proximity space benefits from the 3D metaphor, as it has been shown in the Xerox Parc Z-GUI project. [61]
Since the mid-80s, it has been questioned more and more whether the ubiquitous WIMP (windows, icons, menus, pointers) user interface approach [62] would be able to cope with the increased amount of information which needs to be handled in the workplace of scientists, engineers and clerks. It has been complained that mind-numbing mechanical activity would get in the way of users spending time and energy on what is really relevant and what the users want to accomplish.[61] The current desktop model and its inherent navigation conventions are 2D - but is has been proposed that 2D is not the space we are living in and that from biological considerations alone 3D should be a much more natural space to navigate in. There has been a considerable interest in the development of intuitive 3D-widgets. One focus has been the navigation of document collections, which tend to become too large to be handled efficiently by the traditional interfaces.
Databases play a very big role in chemistry. Compared to other sciences, a much larger part of the relevant literature is extracted, and selected information stored and made available on-line. Chemistry has also developed a number of unique database types, such as structure and reaction databases, where molecular graphs and graph transformations are queried, or spectral databases, which employ spectral similarity operators. So chemistry has both a large number, and very diverse types of databases. Currently, access to almost all databases is limited to a one-dimensional command-line model or 2D graphical interface models, with response forms and 2D schemes. Hits are typically displayed in a sequential order, sometimes clustered to reduce the tedium of paging through very many similar hits.
The traditional 2D display model has become a limitation in many respects. 3D database interfaces are currently a major focus of research in the database community. This trend is especially observed with respect to object-oriented databases, when there is no longer a database structure with can be easily represented with tables and navigated by requesting query values in columns. Object-oriented databases are increasingly used, especially in bioinformatics, to cope with the rapidly growing amounts of very diverse data from genome analysis which is difficult to map onto classical relational table structures. The concept of 3D interfaces to databases is unrelated to the storage and retrieval of 3D atomic positions in structure databases or the provision for 3D search operators on the data, which is of course a well-known feature of numerous chemical databases. There is a rich literature on the problems of 3D structure searching, and this topic will not be part of this review.
The AMAZE project [68] is probably the first instance of a 3D interface to a complex chemical database. The developers of this system share the conviction that the WIMP interface paradigm with its desktop metaphor has outlived its usefulness, because it does not provide the expressive power software engineers require when faced by the currently observed shift in priority from functionality to usability. Since most chemists still relegate more complex database queries to experts, the chemical community which has been surrounded by powerful, but complex database systems can probably relate to that statement. The main object-oriented database of the AMAZE system contains protein structure data with object classes such as the basic atomic structure, helices, loops, strands and other descriptors known from the chemical literature. These data objects typically are part of one-to-many or many-to-many relationships and are of a type which is awkward to encode and search efficiently in a classical database. Specifying meaningful and syntactically correct queries on the right objects from a complex database structure (about 20 object classes) with a textual query language has proven to be very difficult. It was however found that biochemists, who are routinely using 3D graphics in their modeling work, became proficient very rapidly when a 3D interface was offered. In this interface, the conceptual schemes and returned sets are depicted as blocks, forming a result maze. Biochemists can navigate to blocks which are a result set and attach more query operators to this set until the set is reduced to members which exhibit a complex pattern of characteristics in the various descriptive database classes. Query operators and operator combinations are represented by various shapes. Editing the parameters of an operator box is performed by selecting a block and opening input panels.
A final system called LyberWorld [70] focuses primarily on interfacing full-text databases in a VR or classical screen environment. Chemists know that retrieval of papers from abstract databases is often a problem. Finding the significant entries with a free vocabulary is not easy, and the number of abstracts yielded as the result of broad-range generic queries often so large that a manual perusal is impossible. LyberWorld uses innovative and very detailed abstract graphical objects in 3D space, such as the Relevance Sphere and the Search Tree widgets, to display relevance rankings, content relationships and similar information. LyberWorld has been demonstrated as a 3D interface to the INQUERY probabilistic information retrieval system of the CORDIS database. CORDIS stores textual information about research projects, including chemistry projects, sponsored by the European Community. Currently LyberWorld is being integrated with the ReLiBase database in the Docking-D [71] project. ReLiBase [72] is implemented on top of an object-oriented database system (VODAK) and stores data from heterogeneous sources in an integrated receptor-ligand database. The main application of ReLiBase is posing associative queries for analyzing the available data, exploiting the object relationships derived from different databases and algorithms under the ReLiBase umbrella. The handling of large numbers of textual documents in LyberWorld is shown in Figure 5.
The systems described so far were designed for single users. The concept of Populated Information Terrains (PITS) [73] extends database technology by multi-user VR and Computer Supported Cooperative Work (CSCW). Various other models of telecooperation for chemistry are discussed in the Collaboration and Conferencing section of this review. The rationale for using PITS were twofold. First, traditional database interfaces provide ample support for sophisticated querying (if the users can master the grammar), but lack in support for browsing. Because of the larger information density in volumetric representation, navigation for browsing should be facilitated. The density of information comes from the fact that attributes of entries or tables can be mapped both to extrinsic dimensions (position) and intrinsic dimensions (color, shape, size, spin, texture) simultaneously and there is a gradual scaling of visibility, as opposed to paging through traditional 2D output, where a record and its attributes are either visible in detail or not at all. Second, many databases provide concurrent multi-user access, but do not offer any method to share the data as part of cooperative work. Using systems like Q-PIT or DIVE-Q, multiple users wearing HMDs (Head Mounted Devices) can cooperatively explore data relationships and manipulate their common workspace by attaching queries or performing other manipulations.
Closely related to the database interfacing issues outlined in the previous section, which focus on navigation, query construction and browsing, are the recent developments in the field of data(base) mining. [76] The term database mining describes the process of sifting through large amounts of data with the aim to detect significant relationships between individual data items and to create rules with predictive power from these observations. A scientist can work effectively with a few thousand observations with a small number of measurements each. Effectively digesting millions of data points, each with tens or hundreds of measurements, is another matter. A variety of techniques are used for database mining, including hypothesis generation and verification, unsupervised global search for interesting patterns, supervised goal-directed mining, and graphical visualization of data relationships. [77] Since the variable space is often of a very high dimension, or the structure of the results, such as decision trees, has some complex tree-like shape - convincing visualization of the findings of the mining process is of considerable importance. Walkable 3D-scenes for trees, rules and evidence patterns have been found to be a useful model for this type of application. Trees which are larger than what can be printed in readable form on a sheet of paper are difficult to view and navigate with traditional 2D-type tree displays. Database mining is a field in chemistry which has just begun to become systematically explored, although there is a considerable potential. Many companies possess extensive datasets, for example from QSAR studies originating from drug development projects. Until recently, QSAR studies were typically limited to small datasets of similar compounds. With the progress in available computational power and mining algorithms, large-scale reevaluations of the data collected over years has prospect to unearth hitherto overlooked patterns which might lead to reactivated projects leading into areas not explored in the original work. Another promising application area is the analysis of DNA and RNA sequence collections. VR environments, as well as novel non-immersive display metaphors, have been found to be very effective tools for the supervision and guidance of mining sessions.
Two new technologies for solid rapid prototyping have been successfully applied to build hardware models of proteins and other chemical structures with complex surfaces. [91] The first method, stereolithography, uses a vessel filled with liquid monomers. One millimeter or less below the surface a pedestal which can be lowered by a motor is installed. A switched UV laser scans the surface, inducing polymerization at the irradiated surface points. After each surface scan, the pedestal is moved downwards by the thickness of one layer, moving the polymerized surface parts deeper into the monomer bath and protecting it from further laser irradiation. Slice after slice is produced this way. The polymer of a new layer will form a homogenous material when regions overlapping the layer below are irradiated. Layers, which are intersections of the model, can consist of unconnected polymer patches, so arbitrary shapes can be produced, as long as there are holes allowing the monomer to flow from the interior of cavities.
On a much smaller scale, direct manipulation of atoms on surfaces has become a reality with Scanning Tunnelling Microscopes (STM) as sensor and effector. Instead of creating macroscopic real models of compounds for virtual manipulation, microscopic real entities are manipulated from macroscopic virtual models. A VR environment, the Nanomanipulator [93], is under development which translates actions of the chemist into voltage pulses send to the microscope needle which can be used to move single atoms on surfaces. The scientist wears a HMD and operates with a mechanical arm, not unlike those used in radioactive isotope processing plants. The chemist can feel the surface structure by moving a pointer with the mechanical arm over the virtual sample surface. This cases the sensor tip of the microscope to be moved, and the mechanical arm produces a force feedback according to the voltage at the sensor tip. The Nanomanipulator has been used to study the mechanisms of surface modifications on gold and silicon substrates, monolayers of various carbonic acids, and in a number of other projects.
However, virtual reality environments, transportable reactive or scripted 3D scenes, direct response modeling and 3D information navigation are all very recent developments, as can be seen from the dates on the cited literature references. Some of these technologies still require computational resources and special-purpose hardware far beyond the budget of ordinary users. As far as the necessary computer power is concerned, this is probably nothing which should be worried about in the long term. The performance numbers of desktop equipment continues to double every year and a half, so computations requiring a dedicated parallel processor costing millions of dollars today may be affordable mainstream desktop equipment in five years. Similarly, the performance of graphics boards is being pushed at a rate even faster than that of raw processor power. So these performance criteria which render many of the more advanced applications described in this paper unusable for most of the potential audience should not be considered a serious problem preventing VR to become a mainstream tool.
In constrast, the advantage of installing and using a fully immersive VR environment such as a CAVE, or other special equipment such as a VR workbench, is much more debatable. These systems must demonstrate their usefulness in a much more dramatic way in order to justify the investment in a cost-conscious environment. There are a number of obstacles to be overcome by this approach: The costs are very high and not likely to fall with the same speed as those of standard computer equipment, the size of the installation exceeds the space of a normal office, and the user has to leave his or her normal office to work with the system, disrupting the workflow. To make these systems a success, killer applications are required. Operated by a skilled and trained user, a data analysis or modeling performance must be achieved which without doubt outclasses 2D and simple 3D (computer screen projections, possibly combined with shutter glasses) technology for the same task. According to our judgement, this kind of performance still waits to be demonstrated for chemical applications. We feel that this is not impossible, since convincing demonstrations for other scientific fields such as fluid dynamics, oceanography [95], meteorological data analysis or military intelligence work [96] have been given. In the field of modeling, the described results obtained with immersive VR fail to impress from the modeling standpoint. We do not yet see how the results of the described docking experiments could not have been achieved in a more limited environment, provided the computer power for the forcefield updates with immediate response were available - or even with a more restricted forcefield on standard hardware.
If we claim that immersive VR modeling is an area where more work is needed on cogent demonstrations proving the power of VR, what about the promised advantages of using 3D metaphors in the other areas described in this review? The potency of using non-VR 3D graphics of all kinds for molecular visualization is probably unchallenged, since chemical structures are so obviously three-dimensional. In the area of 3D navigation and database queries, quantitative studies exist which were executed in a controlled environment and corroborate that 3D metaphors lead to a measurable improvement in the time required for tasks such as finding a document of given content or assembling complex database queries. Both are in fact everyday tasks of chemists, performed more often than modeling or 3D visualization. These tasks related to information retrieval exhibit specific chemical peculiarities but are certainly not unique to chemistry, or a subfield of chemistry. Modeling and molecular visualization however are domains of specialists. The potential number of users, and the market, for database access systems is much larger in comparison. We feel that once momentum is gained in these areas, a widespread introduction of the described class of systems for document navigation, database retrieval, and database mining can be expected within the foreseeable future.
The field of VR as a whole is of very recent origin, and scientist have to get used to the underlying ideas to be able to make creative use of its capabilites. Scientists in Wonderland [97] is a phrase which still describes much of the ongoing work. Making a definitive prediction about the future directions of VR is certainly difficult. Molecules and molecular interactions are inherently 3D, and chemistry has the largest document, factual and structure-related databases of all sciences. We are convinced that VR in general has promise to become an important tool in managing chemical information in the long run. The conceivable application areas in chemistry are very broad. They reach from the examination, modeling and documentation of the behavior of a single small molecule to distributed database navigation with millions of structures, spectra, sequences and documents. Exciting developments will most likely take place in the next years which have not yet been thought about.