In March, the BSI Oxford Immunology Group (BSI OIG), in collaboration with the University of Oxford Immunology Network, presented its annual one-day symposium to showcase the fantastic immunology research happening across Oxford. Researchers from a range of disciplines met at the University of Oxford Mathematical Institute to present their work and explore opportunities for cross-disciplinary collaboration.
Advances and challenges for AI in immunology
The final session of this year’s symposium tackled the role of artificial intelligence (AI) in immunotherapy discovery. Professor Hashem Koohy, of the MRC Translational Immune Discovery Unit, began by giving a brief history of statistical inference, machine learning and artificial intelligence, before setting out the major advances and critical challenges for AI in immunology.
The discussion drew on our panel’s expertise in multi-modal single-cell analyses, ethics and regulation, and translation to industry. One fundamental challenge was to establish if AI or machine learning algorithms have truly advanced our understanding of biology and immunology. For example, in structural biology, AI models such as AlphaFold have enabled inference of protein folding for hundreds of millions of protein complexes, but it is not clear whether this has been translated into a deeper understanding of structural biology.
We discussed the use of AI and machine learning in the data-rich field of single-cell biology and the challenge of interpreting increasingly complex models, given that the machine learning field has historically favoured accuracy over biological plausibility.
We agreed that there had been a skew towards a ‘snapshot’ approach to new software development, where one model is said to outperform another in a narrowly defined context. However, for the development of immunotherapies, the proof is always in the ‘immunological pudding’, and both academia and industry now require solid functional data to support the adoption of new models.
Collaboration and transparency in innovation
As algorithms become more advanced, there is a risk that the resources required to develop them limits their use to a few large technology and pharmaceutical companies. Our panel highlighted the role of international competitions, consortia and anti-trust legislation in ensuring the fair use of these algorithms, as well as the potential of international collaborations to develop and publish the large datasets required by the models.
We discussed the challenges of applying existing regulatory and ethics frameworks to AI and machine learning when used to improve our understanding of health and biological systems. There is a clear need for responsible research that abides by data protection laws, as well as transparency in development to ensure the quality of data sets used to train systems, and to limit the effects of existing biases. The panel agreed there is an important role for the academic community in developing AI regulation alongside industry, regulatory bodies and government departments. Small biotechnology firms are also key for driving innovation, as the era of model-first innovation wanes and focus shifts to favour companies creating novel datasets, new discovery modalities, or new lenses into immunological mechanisms.
Future challenges
In summary, while AI and machine learning have clearly advanced the general state of immunological knowledge, and have enabled us to ‘automate the boring’, there remain significant challenges if AI is to provide a deeper understanding of biological systems.
Integration of prior knowledge in the form of Bayesian models, and mechanistic modelling in the form of ordinary differential equations within an end-to-end model architecture, are expected to be budding areas of research.