A technically sharp AI and biology paper published on April 14, 2026 is ImmUni, in Cell Genomics. The paper presents a unified transformer based framework for modeling three linked steps in neoantigen biology: HLA binding, antigen presentation, and immunogenicity. That matters because a lot of immuno oncology pipelines still treat these as separate prediction problems, even though the biology is sequential and tightly coupled. ImmUni moves in the opposite direction and tries to model the chain as one connected system. 
What makes this interesting from a technical perspective is the emphasis on cross task interpretability rather than just one benchmark number. According to the paper summary, the model is designed to learn across binding, presentation, and immunogenicity at the same time, which gives it a shot at exposing where shortcut bias appears and where the biological signal actually survives across tasks. That is more useful than a narrowly optimized classifier, because the real problem in neoantigen prediction is not only accuracy. It is understanding which features remain meaningful as you move from peptide chemistry to immune recognition. 
This matters because cancer immunology is full of partial proxies. A peptide can bind MHC and still fail to matter clinically. It can be presented and still fail to trigger a T cell response. A unified model like ImmUni is interesting precisely because it tries to reduce that fragmentation. Instead of stacking disconnected tools, it builds one representation that spans the whole path from candidate epitope to likely immunogenic signal. That is the kind of modeling move that can make AI in biology more mechanistic and less like a pile of disconnected scores. 
The broader signal is that AI in biology is getting more serious about composing biological processes instead of predicting one assay at a time. Early systems often stopped at affinity prediction or presentation ranking. Work like this points toward models that try to represent biological causality in stages, even if imperfectly. If that trend continues, some of the most useful models in computational immunology will be the ones that connect the layers rather than winning one isolated subtask. 
Sources
https://www.cell.com/cell-genomics/fulltext/S2666-979X(26)00076-5
https://www.sciencedirect.com/science/article/pii/S2666979X26000765
https://www.cell.com/cell-genomics/newarticles