Recent field notes
Early storylines are taking shape around public data and AI-assisted workflows. These notes are not finished claims. Each one is held deliberately small: a question, a working boundary, and the kind of external expertise that would make the next test stronger.
May 2026
AlphaScience Lab is using public datasets and AI-assisted execution to sketch a set of careful, evidence-led research directions. The common thread is not automation for its own sake. It is the attempt to make uncertain analyses auditable enough that outside collaborators can decide what is worth testing next.
01
One story starts in Alzheimer disease atlas data, where high-pathology donors do not always collapse into a simple marker narrative. The current working thread is a restrained Sst-centred inhibitory gate: a candidate way to organize single-cell, spatial, and molecular signals without treating the pattern as causal proof.
Useful collaborators: AD biology, single-cell or spatial analysis, proteomics, and external cohort validation.
02
Another story follows a failed simple predictor. Public dependency and drug-response resources suggest that anti-ferroptosis defense can mean different things at the expression and dependency layers. The useful claim is still bounded: a computational architecture that needs experimental and pharmacological pressure tests.
Useful collaborators: ferroptosis biology, cancer dependency, drug response, and experimental validation.
03
UMAP, t-SNE, and related maps are powerful because they make high-dimensional data visible. They are risky for the same reason. This project asks a narrower question: when a visual claim appears on a two-dimensional map, what high-dimensional evidence actually supports it?
Useful collaborators: single-cell, spatial omics, bioinformatics, visualization, and claim-auditing workflows.
04
Cross-disease atlases invite the hope that cell states can transfer across contexts. We are treating that hope cautiously. The question is whether there are bounded, transferable representations that survive disease, cohort, and platform shifts, and what must be rejected as overgeneralization.
Useful collaborators: statistics, machine learning, single-cell atlas integration, and external cohort validation.
05
The toxicogenomics story is intentionally small: a source-resolved mini-atlas around BDE-47 and related public evidence. The goal is not to make a broad environmental health claim from thin material, but to organize heterogeneous omics signals into a more inspectable evidence map.
Useful collaborators: toxicology, environmental health, omics evidence synthesis, and exposure-aware interpretation.
06
The method story sits underneath the science stories. We are building a workflow that keeps track of prompts, code, data choices, failed routes, figures, checks, and human judgement. The aim is a cleaner record of how AI-assisted work becomes a bounded scientific claim.
Useful collaborators: AI for science, reproducibility, research integrity, and human-AI collaboration.
Invitation
If one of these storylines maps onto your expertise, the most useful next step is a bounded conversation: what would count as a serious external check, what should be removed from the claim, and what data or experiment would make the story worth pursuing.