Published writing on immunology, bioinformatics and AI in biology

The immune system is often described as two arms, but in practice it behaves as one continuous defence. Innate immunity acts within minutes: physical barriers, complement proteins, macrophages, neutrophils and natural killer cells detect broad molecular patterns shared by many pathogens.

Adaptive immunity takes days to mount, yet it is precise. B cells and T cells recognise specific antigens through highly variable receptors, expand clonally, and leave behind a memory population that responds far faster on re-exposure.

The handoff between them is where most of the interesting biology happens. Dendritic cells sample tissue, migrate to lymph nodes, and present processed antigen to naive T cells — turning a generic alarm into a targeted, long-lived response.

BLAST answers a deceptively simple question: which known sequences look like mine, and how confident can I be about that resemblance? Start by retrieving a sequence in FASTA format from NCBI and choosing the right program — blastn for nucleotides, blastp for proteins.

When reading results, the E-value matters more than percentage identity alone. A low E-value means the alignment is unlikely to have occurred by chance given the database size; high identity over a short stretch can still be noise.

Query coverage is the third number worth checking. A hit with 98% identity across only 12% of your query is telling you about a shared domain, not a shared gene.

Most of an immunologist's time is not spent at the bench — it is spent reading. Language models are genuinely useful for narrowing thousands of abstracts into a shortlist worth reading properly, provided every claim is traced back to the source.

Prediction tasks are the second front. Epitope prediction, TCR–peptide binding and single-cell annotation are all pattern-recognition problems with enough labelled data for models to be helpful, if imperfect.

The honest limitation is validation. A model can propose a hypothesis in seconds; confirming it still requires an experiment. The value of AI here is throughput of ideas, not replacement of evidence.