In terms of disease diagnosis, our immune system is the best doctor we know. It carries out disease diagnosis with unmatched precision before any clinical symptoms arise. The past and ongoing battles with disease and infection are recorded in the form of immune memory, which is composed of a repertoire of immune cells that bear immune receptors that specifically recognize and neutralize invading pathogens.
We are now able to read these immune receptors using high-throughput DNA sequencing on small blood samples at a cost realistic for clinical use. However, we are not yet capable of understanding what we read. Specifically, we are not yet able to translate an immune receptor?s DNA sequence to which disease states are reflected by these immune receptors. Since the immune system is continuously responding to the presence of pathogens and other factors, learning the link between immune receptors and disease would enable continuous monitoring of disease state throughout life.
The pattern recognition capacity of machine learning makes it uniquely suited for translating the immunological sensor system into a human-readable account of disease. However, rather than lending themselves to the application of existing machine learning methods, immune repertoires have particularities that call for conceptual advancements in machine learning methodology.
We have in the early phase of the Doctor AI2 project already made good progress in three complementary directions:
1) we have made interesting theoretical observations related to multiple-instance learning, which is a key methodological approach in the project as it allows to model the immune state of large repertoires of immune receptors.
2) we have made good progress related to spatial modeling of interaction between immune receptors and antigens.
3) we have completed a first version of a machine learning platform for immune repertoires and described the platform in a manuscript currently available as a preprint.
Early diagnosis of disease is key to optimal treatment and in terms of diagnosis, our adaptive immune system is the “best doctor”. It carries out diagnosis with unmatched precision before clinical symptoms arise. There is a gold rush in academia and industry to develop artificial intelligence (AI) methods that exploit our immune system’s capacity to assist doctors in everyday diagnosis.
The adaptive immune system records each past and ongoing battle with disease. This immune memory is recorded by “immune receptors” - short genetic sequences specific for each disease. Immune receptors can today be sequenced at high-throughput. We have previously shown that similar immune receptors (similar: identity of entire genetic sequence or subsequence) arise in different individuals when faced with the same disease. Thus, the pattern recognition capacity of machine learning may be leveraged to detect disease-associated patterns in the genetic sequences of immune receptors. However, so far, machine-learning-based exploitation for immunodiagnostics of immune receptor sequence datasets has been rather poor. This is due to (1) a lack of machine learning approaches that can exploit the unique biological characteristics of immune receptor repertoires, (2) and a lack of ground truth data for machine learning benchmarking.(3) There exists currently no platform for the machine learning analysis of large-scale immune receptor datasets.
To resolve these knowledge gaps, we propose to develop novel AI methodology and implement a comprehensive software platform for immune receptor-based diagnostics. To validate our framework, we have access to the world-wide largest experimental and synthetic immune receptor datasets.
In the medium-term horizon, the transdisciplinary project Doctor AI^2 will move the research frontier in AI techniques for immune-receptor immunodiagnostics and contributes to the AI revolution in medicine by supporting clinicians in therapeutic decision making.