codex/add-column-for-cell-by-protein-visualization

[huanlity]

Summary

• add protein metadata loader and selection state mirroring the existing gene utilities
• extend the landscape UI to show a protein column with top-protein bars and callbacks that fetch cbp parquet data
• clear and synchronize protein selections across deck.gl interactions and reset protein bars on viewport changes

Testing

• npm run lint:js

---

https://chatgpt.com/codex/tasks/task_b_68fa1ede6bec8331a305d7785df2f839

[cornhundred]

Review notes: protein section in control panel

Overall: The parallel-section approach is right. Proteins are a separate modality
that still maps onto the same cell-coloring mechanism, so a sibling section next to the
gene section is the natural model. The code mirrors the gene path (meta_gene /
selected_genes / bar callbacks) closely, which makes it easy to follow. The cleanest
part is generalizing update_cell_exp_array to take a feature_store + { dataType }
and switch the cbg/cbp directory — that's the correct shared abstraction. Issues to
address before this is the final shape:

1. No backend/data support exists yet

run_pre_processing.py and dega.pre only produce cbg/ and meta_gene.parquet.
Nothing generates cbp/ or meta_protein.parquet. This branch is frontend-only against
files that don't exist yet — need to decide whether the data-prep side lands first/together.

2. The protein section always renders, even with no protein data

There's no is_protein flag analogous to the existing is_nbhd pattern. set_meta_protein
swallows its error and sets empty arrays, and ui_containers.js unconditionally appends
protein_container. Net effect: every existing dataset now fires a meta_protein.parquet
request that 404s and renders an empty "PRT" section. Most important fix — gate the whole
section on a viz_state.proteins.is_protein flag set from a landscape parameter, mirroring
is_nbhd.

3. Mutual exclusion via scattered manual resets is fragile

"Selecting a protein clears the gene/cat selection" is implemented by sprinkling
update_selected_proteins(viz_state.proteins, [], obs_store) into ~12 call sites (every
gene/cat/nbhd/dendro/search/cgm callback). This was already a smell with genes↔cats; a third
modality makes it combinatorial and easy to miss a site. Consider a single
clear_feature_selections(viz_state, { except }) helper or a single "active feature" concept.

4. Defensive ?. / || [] / || {} everywhere signals uncertain init

Guards like if (viz_state.proteins && ...), top_protein_counts || [], and
if (!viz_state.proteins.svg_bar_protein) suggest uncertainty about init. It is always
initialized (both landscape_ist and landscape_sst). Guarantee init once and drop the
defensive noise — it hides ordering bugs rather than preventing them. The
if (viz_state.obs_store.selected_proteins) guards are also dead since create_obs_store
always creates those observables now.

5. landscape_sst gets the state object but no section

landscape_sst.js initializes viz_state.proteins = {...} but never calls set_meta_protein
and never builds a protein bar. Dead state in the SST path — either wire it up or don't
initialize it there.

6. Smaller things

• Inconsistent UI construction: the PRT label is a hand-styled inline div, while
  cell/TRX/NBHD use make_button(...). Reuse make_button; also its blue collides with
  the TRX button's blue.
• No visual gene/protein distinction: get_cell_color returns red [255,0,0] for any
  expression, so protein-colored cells look identical to gene-colored ones. color_dict_protein
  is loaded but only used for bar colors, not cell coloring — decide whether that's intended.
• max_exp ... ?? 1 fallback in update_cell_exp_array silently changes the gene path:
  a missing meta_gene[gene].max previously threw, now it normalizes to 1.
• No model sync: selected_genes.js has sync_selected_genes (pushes to the Python widget
  model); there's no sync_selected_proteins. Fine if not needed yet, just flagging the asymmetry.
• No protein search box to parallel the gene search — bars only.

Bottom line

Keep the parallel-section direction and the feature_store/dataType generalization. Before
merge: (a) an is_protein gate so non-protein datasets are untouched, (b) a centralized
selection-reset rather than scattered manual clears, and (c) a decision on whether the data-prep
side ships alongside. Items 4–6 are cleanup.