docs: propose agentic discovery + inference engineering additions to CV

.gitignore

@@ -9,3 +9,6 @@ _site/
 # Jekyll metadata
 .jekyll-metadata
 
+
+# Vale linter — fetched via `vale sync`, not for commit
+.vale-styles/

.markdownlint.jsonc

@@ -0,0 +1,11 @@
+{
+  // CV and prose use LaTeX-style ventilated prose: one sentence per line.
+  // Disable hard line-length cap; sentences run as long as they need to.
+  "MD013": false,
+  // Ordered list flexibility for publication numbering.
+  "MD029": false,
+  // Allow inline HTML (we use angle brackets for emails).
+  "MD033": false,
+  // Don't require H1 at top of file (Jekyll layout supplies title).
+  "MD041": false
+}

.vale.ini

@@ -0,0 +1,7 @@
+StylesPath = .vale-styles
+MinAlertLevel = warning
+Packages = write-good, proselint
+Vocab = ctr26
+
+[*.md]
+BasedOnStyles = write-good, proselint

docs/sanger-gdm-cover-letter.md

@@ -0,0 +1,85 @@
+# Cover Letter: Google DeepMind Fellow in Genomics and AI
+
+Dr Craig T. Russell
+London, UK · <craig.russell.phd@gmail.com> · [ctr26.github.io](https://ctr26.github.io)
+
+To the Fellowship Review Panel,
+Wellcome Sanger Institute & Google DeepMind Academic Fellowship Programme
+
+---
+
+## Why Sanger, and why now
+
+A growing call for the virtual cell sits across the field, with two paths in view: top-down or bottom-up.
+Bottom-up means simulating biology from molecular kinetics, in the spirit of Karr et al.'s whole-cell Mycoplasma model (Cell, 2012); by conservative estimates this would need compute on the timescale of the universe itself, absent a quantum leap.
+Top-down means learning the cell from data at the scales now becoming available; this *might* be enough to resolve disease without first bridging to bottom-up.
+I want to take the multimodal foundation-model approaches I have built in industry and ground them in the world's richest genomics environment, to address fundamental biological questions.
+Recursion's Maps and Sanger's Atlases are complementary geometries of biology: perturbation-response at industrial scale versus cell-type and tissue coverage at population scale.
+Sanger is the only UK site where atlas-scale single-cell data (Cellular Genomics), a programme committed to predictive biology (Generative Genomics), and a strategic commitment to AI for science sit in one building.
+I want to spend three years there.
+
+I am an ML research scientist at Valence Labs / Recursion Pharmaceuticals.
+I lead projects exploring agnostic representation learning of biology through images, in collaboration with Google Cloud, alongside research efforts on the Virtual Cell initiative, thinking carefully about how to fine-tune multimodal LLMs over knowledge graphs, free text, transcriptomic and phenotypic imaging data to predict cell state and drug response.
+I co-developed TxPert (arXiv:2505.14919), a transcriptomic perturbation predictor, and contributed to the Boltz2 proteome-scale virtual screening pipeline.
+Before Recursion I led AI engineering at EMBL-EBI's BioImage Archive and Uhlmann Group, supervised six PhD students, and released `bioimage_embed` (self-supervised learning for biological images) and `shape_embed` (cell-shape representation, arXiv:2507.01009) as community resources.
+My PhD (Cambridge, 2018) was an instrument-building project: I designed and built a light-sheet microscope from scratch.
+I work as comfortably at the data-generation end of the stack as at the modelling end.
+
+## My Grand Challenges
+
+I would primarily focus on Grand Challenge 4: *Harnessing multimodal generative AI to study the interplay of genetics with spatial transcriptomics, proteomics and morphology in human tissues*, led by Mo Lotfollahi and Muzlifah Haniffa.
+The downstream goal of predicting genetic variation from routine clinical pathology slides sits at the intersection of the multimodal foundation models I have built at Recursion (TxPert; the Virtual Cell initiative integrating knowledge graphs, transcriptomics, free text and phenomics; Boltz2 components) and the imaging methods I built at EMBL-EBI (`bioimage_embed`, `shape_embed`, MIFA).
+Sanger's >200 million curated cells across H&E, Xenium, scRNA-seq and ATAC give a training corpus at the right scale for cross-modal generative architectures.
+
+I would secondarily consider Grand Challenge 5 (*A Foundation sequence model of the Cis-Regulatory Code in developing human tissues*, Lotfollahi/Taipale) and Grand Challenge 2 (*Predictive and Mechanistic AI for Longitudinal Multi-Omics in Complex Disease*, Anderson) as complementary directions, but Challenge 4 is the best fit for my background.
+
+The technical contribution has four parts.
+First, integrate Sanger's H&E + Xenium + scRNA-seq + ATAC corpora under a single multi-expert architecture with a shared latent space, using knowledge graphs as structural priors and curated literature as semantic anchors.
+The architectural lineage is the multimodal LLM-over-omics work I have built at Valence; the data scale is what makes Sanger the right home for it.
+Second, evaluate not only held-out reconstruction but counterfactual prediction.
+Does the model give scientifically reasonable answers when asked 'what would the spatial transcriptome and morphology look like for this individual's variant in this tissue?'
+Third, evaluate and correct for ancestry-shift and tissue-shift in the latent space, so that variant-to-pathology predictions are anchored in data the model has seen rather than extrapolated unsafely from a single reference cohort.
+This connects directly to my outreach plan below.
+
+Fourth, the area I most want to push: build an agentic discovery layer on top of the foundation model, walled in by inference-time engineering so that hallucination stays within scientifically tolerable failure modes.
+Naive agentic LLMs in biology fail loudly.
+The agent invents a gene name, fabricates a citation, proposes a perturbation that exists in no cell line.
+The engineering response is to refuse free generation over the things that matter.
+Constrained decoding restricted to a closed vocabulary of valid Ensembl IDs, HGNC symbols and ChEMBL identifiers; tool-call verification against Sanger's own structured resources (Cellular Genomics atlases, Tree of Life genomes, the human variant catalogues), so that every claim round-trips through real data before the agent emits it; retrieval grounding to a curated literature corpus with citation-level attribution; verifier-model ensembles that score each hypothesis for biological plausibility before it leaves the agent.
+The result is an agent that proposes follow-up experiments (variants to validate, tissues to oversample, perturbations to run) but only inside a sandbox where every leaf claim is checkable.
+This is what separates an LLM that talks about biology from one that can act as a research collaborator.
+GDM mentorship is most directly useful here: the inference-engineering and constrained-generation work at DeepMind on AlphaFold's confidence calibration and the AlphaProof verifier loop is the lineage I want to bring into genomics.
+Beyond single-agent verification, I am keen to push further into agentic swarms: ensembles of agents with strong inference programming, using tools and counterfactuals to surface novel biology, with cross-agent disagreement as an additional verification signal.
+
+Three deliverables in year one.
+(i) A Sanger-anchored multimodal foundation model trained jointly on Cellular Genomics atlases (H&E + Xenium + scRNA-seq + ATAC), with public benchmarks for cross-modal counterfactual prediction.
+(ii) A variant-to-pathology benchmark co-designed with Lotfollahi's and Haniffa's groups that scores models on ancestry-stratified validation cohorts rather than convenience samples.
+(iii) An open-source agentic discovery harness with built-in hallucination suppression (constrained decoding, tool-call verification, retrieval grounding, verifier ensembles) that other Sanger researchers can wrap around their own foundation models.
+Years two and three extend this to the cis-regulatory code (Challenge 5) and to longitudinal cohorts (Challenge 2), where the same foundation-model backbone applies.
+
+For secondary supervisor I suggest Prof Pietro Liò (Cambridge, Computer Laboratory) for his work on graph foundation models for biology, with Dr Virginie Uhlmann (EMBL Barcelona, formerly EMBL-EBI) as a complementary option from prior collaboration on bioimage representation learning.
+Neither is a precondition; I am open to whichever AI/ML academic the panel matches me to.
+
+## Outreach: capacity-building for Genomics and AI in LMICs
+
+The outreach fund is the part of this Fellowship I most want to use.
+I will run a recurring 'Foundation Models for Genomics' workshop series at LMIC partner institutions, twice a year, one week each, producing graduates who can fine-tune and evaluate ML models on local genomic data.
+The teaching pattern is the one I already run at EMBL-EBI for 40+ researchers a year, adapted for compute-constrained settings: containerised pipelines, transfer learning from publicly hosted weights, and datasets sourced where possible from the host country's own collections.
+
+The first workshop I would scope sits with the H3ABioNet / H3Africa community, who have spent over a decade building African genomic infrastructure and have been systematically underserved by the foundation-model wave.
+Year-two and year-three editions expand to South-East Asia (probably anchored at the MORU collaboration in Bangkok or NUS in Singapore) and Latin America.
+Each edition leaves behind a containerised teaching environment that runs on modest hardware, a small number of fine-tuned community models on locally relevant tasks, and a peer cohort of trainers who can run later editions independently.
+I would budget the fund for travel cost-sharing for participants who cannot self-fund, because that is where most 'open' programmes silently fail.
+
+Beyond the workshops I would use the ambassadorial role to push for honest evaluation practices in published genomics-AI work, specifically benchmarks that include ancestry- and tissue-diversity stratifications, in line with the Grand Challenge framing above.
+I have done versions of this community-shaping work already: founding the Virtual Cell Journal Club at Valence, contributing to napari and the Bioimage Model Zoo, and co-authoring the AI4LIFE €5M federated bioimage AI infrastructure grant.
+
+## In closing
+
+I am a research engineer.
+I came up through instrument-building, moved to industry to build multimodal foundation models at scale, and want to return to do this work in the open.
+Sanger is the one UK site where the data, the strategic commitment to AI, and the global-health framing sit in one building.
+I want to spend the next three years there.
+
+Yours sincerely,
+Craig T. Russell, PhD