Landscapes of disease
We have used popular AI models to generate landscape images associated with infectious and neglected tropical disesease
It’s been a little more than four years since we (Miquel and Gemma, Ersilia’s co-founders) shifted our attention as scientists from cancer research to infectious diseases, knowing little more than our undergraduate lessons in microbiology and previous stints working with ISGlobal and CIDRZ. As we delved in the fascinating biology of pathogens that have coexisted with humanity for centuries, some of which exhibit delicate, complex life-cycles that transition between humans, animal reservoirs and insects as transmission vectors, we also realised how little space is given to the complications these pathogens may cause in humans. Chances are that, if you ask a passerby in Barcelona (where Ersilia has its headquarters) what cancer is, you’ll get a more or less accurate description in the line of “abnormal cell growth, dangerous tumour, metastatic spread”. However, names like onchocerciasis, leishmaniasis or schistosomiasis will, at most, elicit some memories of a holiday trip to an exotic, tropical country for which they had to get a vaccine or take a pill.
Our mission at Ersilia is, first and foremost, to build a more sustainable research ecosystem in global health that covers the needs of populations that are often neglected by the current paradigm of research, which tends to happen in and for the global north. To achieve this, we cannot limit ourselves to carrying out scientific research in isolation. We need to build awareness around the existing imbalances in global health, understand neocolonialist practices and how to revert them, and engage a wider audience in the fight against neglected diseases, also commonly spelled as the diseases of poverty:
Trachoma would ultimately disappear if one waited for general socioeconomic development to occur in every household in every village around the world, but with good conscience, we cannot sit and wait for this to happen sometime in the next 100 years. — Taylor, 2009
From the specific perspective of an AI-first organisation, early attempts from our end to cover this include interrogating Large Language Models (LLM) around global health, or reflecting on the perspectives of experts working in the field. We have now had a unique opportunity to further this necessary research. We have spent 10 days at Konvent, a laboratory for creators, artists and technologists in the countryside of Catalunya, in the framework of the AI Residency program. Our goal was to leverage LLMs and text-to-image AI models to reflect on our views of neglected tropical diseases, highlighting the preconceptions associated with them and hoping to create spaces of awareness and discussion. In this research, we have taken inspiration from the great assay “Illness as Metaphor” by Susan Sontag.
In short, we have undertaken (a) a massive literature review, focusing on the first western scientific texts describing neglected tropical diseases (most of which come from the imperialist era, when military doctors had to treat and prevent the scourge of infections among the armies and the native workforce) and (b) a large-scale interrogation of LLMs (GPT) to gather a summary of neglected tropical diseases organised in chapters such as their related medical concepts, the biology of the pathogens, the pre-colonial and colonial historical background, associated social factors, etc. In this exercise, rather than striving for 100% accuracy from the LLMs (which we know are still prone to errors), our priority was to critically inspect the responses generated by the LLMs, as they merely sample the common knowledge available on the web. With this wealth of information, we have automatically generated text-to-image prompts that describe the landscape associated with such diseases. We have intentionally chosen landscapes as a visual depiction to avoid the unnecessary grimness of poverty-stricken patients that the AIs tend to produce. In addition to the text prompt describing a landscape, we have offered the AI a base image of the pathogen in question to get inspiration from, to obtain an overlay between landscape and disease. Conceptually, we found this blend to be an effective way to bring together the romantic, adventurous views of Africa with the stigma associated with poor hygiene, poverty and risks.
In this first iteration, we have generated an AI-based summary of 15 diseases, of which we have chosen 5 to proceed with: Schistosomiasis, Onchocerciasis, Trachoma, Dengue and Leishmaniasis. With that, we have generated more than 2000 pathogenic landscape images using Midjourney, with a somewhat satisfactory result. You can see below an example, and explore a selected image collection in here.
This project is far from finished. We aim to continue it as a side-project to our scientific endeavours. Any feedback or suggestions meanwhile is highly appreciated: hello[at]ersilia[dot]io
All the code developed for the project is available under a GPLv3 licence, and the generated images are released under a CC-BY-4.0 License and can be downloaded here.
