Constraints and requests

Stage 1 already produced a publishable number on workstation-scale data. Getting past it needs a genomic cohort larger than any public dataset offers, and a pipeline reprocessed at a scale a workstation can't reach on its own. Neither of those is something better modeling can substitute for — they are questions of access.

We're applying to NVIDIA Inception for two things, and both follow directly from what Stage 1 already showed us. Four additional hypotheses for improving on the baseline fitness predictor were tested, and all four scored worse than the training-fold mean — so the near-term limit isn't modeling technique, it's data. Public longitudinal cohorts for clonal haematopoiesis are already exhausted, and roughly half of all CHIP mutations — frameshift, splice, and indel variants, including ASXL1 frameshifts, among the strongest signals in myeloid disease — can't be embedded as single-residue substitutions under a protein-language-model approach at all. Closing that gap means a larger cohort and, eventually, a different embedding approach. Neither is something we can do alone.

The Stage 1 pipeline already runs on NVIDIA BioNeMo's ESM-2 protein language model, cross-checked against a HuggingFace backend, on the cohorts harmonised so far. Reprocessing that same pipeline against a larger, biobank-scale cohort is a question of compute headroom to scale, not a new modeling technique.

The next scale of validation requires biobank-scale genomic cohort access, which requires an institutional partnership — not a modeling breakthrough. We don't have a confirmed partner yet. What we're asking for is the credibility and network of NVIDIA Inception to help open that conversation.

Each stage is gated on the previous one producing a number we would be willing to publish.

01

Already produced — this is the completed baseline the rest of the roadmap builds on.

Lothian Birth Cohorts and SardiNIA harmonised into 591 clonal trajectories across 268 unique missense variants; signal enrichment measured at 4.4× (Lothian) and 4.1× (SardiNIA) over the noise floor; gene-level growth rates recovered blind, matching the published ordering in the literature on this same data; pipeline running on NVIDIA BioNeMo's ESM-2, cross-checked against a HuggingFace backend.

02

Gated on securing an institutional partner.

Reprocessing the same fitness-measurement pipeline against a genomic cohort far larger than any public longitudinal dataset offers, to test whether variant-level fitness prediction holds up at a scale that could eventually matter clinically.

03

Gated on the Stage 2 result.

Testing whether clonal fitness prediction changes how an individual carrier is actually monitored, rather than only whether it is measurable in a research cohort.