From hackathon prototype to clinical validation

The original Project Rosie pipeline began with tumour variants, annotated the changes, generated altered peptides, predicted binding, ranked candidates and produced an mRNA construct design. It demonstrated how open scientific tools and a clinician-facing explanation layer could fit together.

Google recognized Project Rosie with the Health & Sciences Prize in the Gemma 4 Good Challenge. That recognition is a milestone in the prototype’s execution. It is not an experimental result showing that the selected targets will work in a patient.

The next development programme, Neoantigen Retriever, asks a more demanding question: which candidates have enough biological support to deserve laboratory testing?

Variant allele frequency is the fraction of sequencing reads containing a variant. It is useful evidence, but it is not a direct measure of the fraction of cancer cells carrying that mutation. Tumour purity and copy-number changes can alter the relationship.

For target selection, we want to understand whether a mutation is shared across much of the tumour or belongs to a smaller branch. Clonality estimates require additional evidence and assumptions. Their uncertainty should remain visible rather than disappearing into a single ranking number.

A DNA mutation can be present even when the altered gene is barely expressed. If the tumour is not producing the source protein, there may be little relevant peptide to present.

Tumour RNA sequencing can add gene-expression evidence and support for the mutant allele. It also introduces its own requirements for specimen handling, quality control and interpretation. Coordinating sequencing around the biological question is more valuable than adding another file type without a plan.

A peptide must pass through a sequence of biological steps before a T cell can inspect it. A binding prediction covers only part of that path. Antigen processing, stable presentation and recognition remain separate questions.

This distinction shapes the validation plan. Binding and stability assays address compatibility with presentation molecules. Immune assays such as ELISpot can measure aspects of a response. Tumour-recognition experiments ask a further question: do responding cells recognize the relevant tumour?

A useful positive result has to be interpreted alongside controls, DLA context, sample quality and the assay’s limits. We have not reported completion of these experiments.

The programme earns confidence through computational benchmarking, experimental validation and conditional clinical evaluation. Each stage has an objective and evidence that gates the next.

• In silico: test ranking against suitable public or curated datasets and established approaches, recording the limits of canine and allele-specific evidence.
• In vitro: evaluate prioritized candidates using defined binding, stability and immune-response assays with specialist laboratory partners.
• In vivo: consider a clinical study only when candidate evidence, manufacturing quality, clinical suitability and an authorized protocol support progression.

A future learning system needs to connect the input data, model version, selected targets, experimental results, product quality and clinical observations. Without that provenance, it is difficult to tell which prediction succeeded or why a candidate failed.

Clinical outcomes also have many contributing factors. An immune signal is not the same endpoint as tumour response or survival. Those questions need suitable study designs, defined endpoints and interpretation by the clinical and scientific team.

The public prototype remains part of the project’s history. It provides a basis for reviewing the original approach and a clear demonstration of execution. The next version needs better evidence handling, transparent ranking and independent validation.

Our immediate work is to develop the computational pipeline and coordinate a clinic-connected validation programme. Prospective collaborators contribute different parts of that effort. A conversation with a clinic, an agreed study and a completed assay are distinct milestones.