Virtual Reality in Biology: could we become virtual naturalists?

15 Virtual Reality (VR) and derived technologies which mixes the real world with some aspect of 16 virtuality (e.g., Augmented Reality, Mixed Reality) has emerged as a powerful allied in 17 education. Recent technological advances made Virtual and Mixed Reality (VMR) accessible 18 at our fingertips. However, only recently VMR has been explored for the teaching of Biology. 19 In this review, we describe how VMR applications are useful in Biology education, discuss the 20 caveats related to VMR use that can interfere with learning, and look into the future of VMR 21 applications in the field. We then propose a conceptual model for the future of VMR in 22 Biology, which envisage to combine VMR with Machine Learning and Artificial Intelligence 23 to provide unprecedented ways to visualise how species evolve in self-sustained immersive 24 virtual worlds, thereby transforming VMR from an educational tool to the centre of biological 25 interest. This conceptual paper aims to stimulate debates on how new technologies can 26 revolutionise teaching and learning across scenarios, which can be useful for improving 27 learning outcomes of biological concepts in face-to-face, blended, and distance learning 28 programmes. 29


Introduction 33
With increasing computational power, technologies that were costly or impossible to 34 implement in the past have now become accessible in laptops and mobile phones (Kish 2004;35 Waldrop 2016). These technologies are now revolutionising the ways we interact with the 36 world, how we learn, and how we teach (Veletsianos 2010). Virtual and Mixed Reality (VMR, 37 see Box 1 for terminology) is one of these technologies which has gained increasing attention 38 in the academic and teaching communities (Mazuryk and Gervautz 1996). In fact, over the last 39 decade, there has been an exponential increase in the publication of papers in topics involving 40 Virtual and Mixed Reality in Education (Fig 1a). VMR can be defined as an alternate world 41 filled with computer-generated entities that interact with human sensory and motor systems to 42 cause a sense of 'presence' (psychological state) in the subject through the use of an 43 'immersive' technology (i.e., technology that simulates an environment that is not necessarily 44 real) (Yoh 2001). Presence can be defined as 'a state of dissociation from reality in which 45 people feel the subjective experience of existing in the digital environment (Slater 2003). ' 46 Although presence and immersion have been used interchangeably (Barbot and Kaufman 47 2020), experiences that increase presence do not necessarily increase immersive feelings, and 48 vice versa [see meta-analysis by (Cummings and Bailenson 2016)], suggesting that, although 49 these terms refer to the feeling of 'being there', they are not necessarily equivalent. 50 Nevertheless, both are important in VMR applications. According to the Oxford dictionary, 51 Virtual Reality is defined as 'images and sounds created by a computer that seem almost real 52 to the user, who can interact with them by using sensors', which highlights that presence and 53 immersion are key aspects of virtual reality [see e.g., (Lombart et al. 2020)]. According to 54 Milgram et al. (1995)'s taxonomy, immersive experiences are achieved through a complex 55 continuum in reproduction fidelity of both the real and virtual environments, whereby the 56 limitations of approaches as well as hardware (e.g., devices and displays) can influence the 57 degree of immersive experience and presence available to the user . Note, 58 however, that the feeling of presence and immersion may not necessarily be the ultimate goal 59 of VMR technologies [ ; but see (Sheridan 1992;Robinett 1992)], 60 although the increasingly more realistic displays can eventually result in one feeling complete 61 immersion and presence (Naimark 1991). 62 63 VMR has been around for decades, and it thought to have its origin when, in the 1960s, Morton 64 Heilig created one of the first immersive multi-sensory simulator that included stimuli such as 65 sound, scent, wind and vibration (called 'Sensorama') (Heilig 1962; Mazuryk and Gervautz 66 discussions below). Having defined the scope of this conceptual paper, we hope that this paper 101 will help guide future developments in VMR applied to biology in a constructive manner, 102 stimulating collaborations across fields (e.g., Computer Science and Gaming) to develop new 103 teaching technologies to facilitate and enhance students' learning experiences. Science Topic query of publications (left) and citations (right) that involves VMR and 107 education (orange) and VMR and biology (red). WoS searches were conducted on 12-May-108 2020 with search term queries '(virtual AND augmented) reality AND education' or '(virtual 109 AND augmented) reality AND biology'. For each search, reviews and proceedings of 110 conferences were excluded. In total, there were 6, 443

VMR uses in Biology education 118
While VMR in education has gained exponential attention of the academic community, VMR 119 in biology has advanced at a slower pace, comprising ~ 5% of academic publications in the 120 field (Fig 1a). Nonetheless, VMR has gained important applications in both secondary and 121 The potential misuses of VMR 169 As for any new technology, we are still discovering the limitations of VMR applications as 170 educational tools. VMR applications are attractive because they contain a wide variety of 171 sensory stimuli that give the participant a sense of presence. However, too many stimuli -such 172 as colours, shapes, characters, movement -can distract the participant and have detrimental 173 effects on learning, a phenomenon that has been acknowledged in the literature and commonly 174 referred to as cognitive overload (Whitelock et al. 2000). A recent study has shown that 175 university students learned less and experienced higher cognitive overload when they 176 experienced a science lab in a fully-mounted VMR headset as oppose to the VMR scenes 177 played on 2D displays, in spite of higher feeling of presence in the VMR scene as opposed to 178 the 2D screen display (Makransky, Terkildsen, and Mayer 2019). This suggests that, in some 179 cases, the very same attributes that make VMR attractive can make VMR applications 180 ineffective. Nonetheless, it is possible to mitigate cognitive overload by controlling the 181 information flow that students are exposed to as part of the design of the environment, the tasks 182 needed to be completed, as well as the degree of interactivity with the environment at any given 183 Interestingly, a recent meta-analysis suggested that the feelings of presence and motion-196 sickness likely trade-off -with higher presence resulting in lower motion-sickness (Weech,197 Kenny, and Barnett-Cowan 2019), supporting the idea that motion-sickness can be mitigated 198 (or eliminated) with thoughtful design and user-experience. energy and spatial resource, respectively [as described in (Ray 1992)]. As a result, artificial 227 Tierra entities become progressively more adapted to exploit one another in order to gain 228 advantage over the use of CPU and memory (Ray 1992;Thearling and Ray 1994;Ray 1994; 229 Ray and Xu 2001). The outcome of this self-sustained virtual evolutionary world is remarkable 230 given that the system evolves differences in entity sizes, ecological specialisation (e.g., 231 parasites) and population dynamics processes (e.g., extinction) (Ray, 1992(Ray, , 1994

Can VMR and Artificial Intelligence (AI) revolutionise artificial evolutionary systems? 243
As discussed above, VMR is a powerful and appealing educational technology to teach 244 biology. This is because students and educators respond rationally as well as emotionally to 245 the educational material in the immersive experience, which can accentuate learning (de Jong,  The technological advances that allowed VMR to become an accessible tool has also allowed 252 for the feasibility of powerful statistical models of Machine Learning and Artificial Intelligence 253 (AI). Machine Learning are algorithms that process and learn with huge amounts of data in 254 order to perform a task without necessarily being explicitly programmed to do so (Bishop 255 2006). AI attempts to simulate human intelligence in machine systems; this includes machine 256 learning but also (bio-inspired) robotics, ethics and philosophy associated with AI development 257 (Russell and Norvig 2002 The answer to the first question is, in our opinion, a sounding 'yes'. We strongly believe that 264 future technological advances have the potential to create an immersive virtual world that 265 reproduces the forces of evolution, which can allow us to visualise and measure how species 266 have evolved, how ecosystems are formed, how species adapt to their environment, how we 267 can anticipate effects of adverse climatic conditions across ecosystems in our changing world. 268 In a sense, we could become 'virtual naturalists' that explore evolution in a simulated (virtual) 269 world in the same sense that naturalists explore the natural (real) world. The learning benefits 270 are unprecedented given that students can experience inaccessible and inhospitable 271 environments, observe evolution, adaptation, trophic interaction, parasitism and many more 272 biological processes without stepping outside the classroom (Learning affordance #2; Table 1). 273 Moreover, the freedom given to the students within these BioVRs forms the perfect ground for 274 inquiry-based learning and engagement, where the students will observe and explore the  Table 1). Ultimately, the BioVR could be used to supplement and stimulate collaborative 278 learning tasks, where students may compare and contrast the evolution of individual BioVR 279 evolutions and identify similarities (e.g., parallel evolution) and divergences (e.g., 280 specialisation traits) between virtual entities (Learning affordance #5; Table 1 BioVR and without it, the system does not have the evolving entity. The ancestor entity 300 is equivalent to the ancestor species which gave origin to life on Earth, and is a common 301 feature of artificial life systems [e.g., (Ray 1994)]. In other words, the ancestral entity 302 is the first 'living' inhabitant of the virtual planet. 303

Gather a large empirical dataset of environment-traits-species interactions as a basic 304
starting-point for determining how different species evolve in different ecosystems 305 (e.g., evolutionary convergences, divergences, character displacement) -this could be 306 called 'rules of evolution'. This will allow the system to 'know' which adaptations are 307 more likely to yield fitness advantages in a given environment. For instance, heat 308 tolerance (or traits related to coping with high temperatures) is a likely to increase 309 fitness of populations living in virtual habitats that resemble arid regions. The actual 310 evolution of traits following the 'rule of evolution' will depend on the variability 311 present in the ancestral entity populations and we envisage that, in a BioVR simulation, 312 this parameter can adjustable according to the purpose or the needs of the virtual world. 313 One way in which evolutionary rules could be extracted from this dataset is using, for 314 example, supervised learning and/or clustering algorithms (see Box 2) to extract general 315 rules as to how species evolve (morphologically and behaviourally) across different 316 environments, commonality between functional traits across species in the same 317 environments, as well as the number, distribution, and behaviour of different species 318 within the same environment. For instance, species living in warm habitats are likely 319 to share similar adaptations to high temperatures and thus, one could expect that virtual 320 entities inhabiting warm virtual habitats should resemble follow similar patterns. Once 321 these rules of evolution are estimated (or guessed) across all habitats of the virtual 322 world, ancestral entities can evolve and differentiate accordingly. For example, imagine that a species evolves a remarkable adaptation to convert virtual 326 resource A into B. This transformation should feedback into the system so as to allow 327 new evolutionary rules, perhaps favouring other species to adapt and utilise virtual 328 resource B (which is being produced) instead of virtual resource A (Fig 2). 329 5. Given this self-sustained cycle of interaction between entities and the environments, 330 and the iterative system that modulates virtual evolutionary rules, BioVR can become 331 an artificial ecosystem, fully accessible for exploration through VMR in inquiry-based 332 learning quests. This allows students and researchers to experience and study evolution 333 in this immersive environment, comparing the outcomes of evolutionary forces within 334 different environments within a BioVR and across BioVRs with different setups. 335 Furthermore, since data visualisation is key for understanding biological processes 336 [e.g., (Karr and Brady 2000)] and is an essential component of affective learning, the 337 use of VR to create BioVR worlds will allow VR to transcend the status of an 338 educational tool that helps learning and teaching in Biology to become the main 339 technology for experiencing and learning about virtual biological phenomena. 340

341
We provided these steps in order to quick-start ideas around the practical challenges necessary 342 to realise the conceptual proposition made in this paper. We are not presumptuous of our algorithm is implemented to empirical environment-trait-species datasets in order to extract the 361 patterns (or 'rules') of evolution across environments. Meanwhile, the initial settings for the 362 BioVR world and the ancestral AI entity are also set. The settings include physical and 363 environmental conditions, as well as patterns of lifespan, movement, and reproduction of the 364 AI entity. Next, the 'rules of evolution' are incorporated into the BioVR and AI entity with 365 original settings, and the BioVR is allowed to evolve. Note that the evolution patterns in the 366 BioVR are then fed-back to the machine learning model, which is updated. This way, the only 367 input from empirical data is at the initial states, and BioVR are allowed to evolve independently 368 afterwards. As a result, we can measure and visualise species evolution as it happens, in an 369 immersive experience of the BioVR.

Evolving BioVR
Initial empirical rules of evolution

Conclusion 373
The use of VMR has provided promising results for consolidating learning across secondary 374 and tertiary biology education. With increasing technology, the combination of VMR with 375 Machine Learning and AI has the potential to create a self-sustained evolving virtual world 376 (BioVR) that allow us to uniquely explore how life as we know evolves and responds to 377 extreme climatic conditions. Thus, the use of new technologies and innovative applications can 378 provide better learning outcomes and student experiences in face-to-face, blended, and distance 379 learning contexts by amalgamating biological principles in immersive classrooms. 380 381

Funding details 382
This study was not funded by any funding agency. 383 384

Conflict of interest 385
The authors have no conflict of interests to declare. 386

Box 2 -Supervised and unsupervised machine learning. 675
Machine learning models can be broadly classified into supervised or unsupervised learning 676 algorithms, depending on the structure of the data (Mitchell, Michalski,  here, see e.g., (Zhu and Goldberg 2009) for details). Unsupervised learning algorithms use data 679 in which the outcome is not yet labelled or identified, and therefore the algorithm cannot 680 'know' the outcomes in advance. The algorithm then learns how to classify and predict the 681 outcome from new observations based on the inherent structure of the data at hand. An example 682 of unsupervised learning is the clustering of groups within a dataset (Fig 4a). Conversely, 683 supervised learning algorithms uses data in which the outcome is known, and the algorithm 684 learns how to predict the outcome of future observations based on what was learnt from the 685 information and outcomes obtained from previous data. An example of supervised learning is 686 the classification (or prediction, in the case of regression models) of a new observation between 687 two categories based on n number of characteristics or variables (Fig 4b).