--- license: cc-by-4.0 pretty_name: The Well for The Cell tags: - biology - simulation - time-series - whole-cell - E. coli size_categories: - 100K_file`, `_row_offset`, `_n_rows`) are null for data not in this repo, so `index["_file"].notnull()` (e.g. `index.expression_file.notnull()`) is the authoritative availability check. `load_dataset(..., "", split="")` fails for a family that has not been uploaded (e.g. `load_dataset(..., "expression", split="single_knockout")` until that family is uploaded). ## Configs Each config is one data product. `summary` is one row per trajectory and joins to `index.parquet` on `sample_id`: `n_steps` and, for each of the 18 scalar channels in the order of the product schema, `_mean`, `_min`, `_max`, `_first` and `_last` over the trajectory's 1 s `core` rows (int32 for `mrna_counts` and `n_oric` except `_mean`, float32 otherwise; the `growth_rate` statistics skip its null row 0, so `growth_rate_first` is the value at t = 1). For every family it also has 13 pathway proteome fractions, `pathway_translation_ribosome`, `pathway_transcription_rnap`, `pathway_stringent_response`, `pathway_glycolysis`, `pathway_tca_cycle`, `pathway_pentose_phosphate`, `pathway_respiration_atp`, `pathway_aa_biosynthesis`, `pathway_amr_efflux`, `pathway_oxidative_stress`, `pathway_global_regulators`, `pathway_proteostasis` and `pathway_cell_division` (float32, unit `fraction`): for each pathway of the Cellarium pathway panel, the pathway's protein counts over all protein counts at each 10 s `expression` row, averaged over the trajectory (Cellarium's definition; the gene lists are in Cellarium's `pathway_panel.py`). Every other config but `charging` (see tRNA charging) carries `sample_id`, `t`, `t_abs` and the 18 scalar channels; the matrix configs add per-gene or per-reaction columns. `expression` and `metabolism` keep the rows with `t % 10 == 0` plus each trajectory's last row (point samples, not averages); `expression_1s` and `metabolism_1s` keep every row. The split of every config is the family. `expression` and `metabolism` are exact row subsets of `expression_1s` and `metabolism_1s`; they are kept for loader compatibility and will be reconsidered at 1.0.0. | config | rows every | matrices (active columns) | trajectories | rows | size | |---|---|---|---:|---:|---:| | `summary` | trajectory | none: `n_steps` and `_mean`, `_min`, `_max`, `_first`, `_last` of each scalar; 13 `pathway_` fractions | 117,174 | 117,174 | 35.0 MB | | `core` | 1 s | scalars only | 117,174 | 316,751,845 | 18.6 GB | | `expression` | 10 s | `mrna_counts_matrix` (2,773), `protein_counts_matrix` (4,174) | 117,174 | 31,830,394 | 82.7 GB | | `metabolism` | 10 s | `reaction_flux_matrix` (3,974), `exchange_flux_matrix` (82) | 117,174 | 31,830,394 | 90.3 GB | | `expression_1s` | 1 s | `mrna_counts_matrix` (2,773), `protein_counts_matrix` (4,174) | 117,174 | 316,751,845 | 327.4 GB | | `metabolism_1s` | 1 s | `reaction_flux_matrix` (3,974), `exchange_flux_matrix` (82) | 117,174 | 316,751,845 | 880.3 GB | | `charging` | 1 s | `aa_pool_size` (21), `ntp_pool_size` (4), `trna_charged_fraction` (21) | not in this release | | | Matrix columns are pruned to the columns that are non-zero at least once anywhere in the campaign (active columns: mrna_counts 2,773 of 3,133, protein_counts 4,174 of 4,310, reaction_flux 3,974 of 9,612, exchange_flux 82 of 87). The full column lists, with an `active` mask and each active column's position in the product (`product_column`), are in `provenance/reference/.parquet`. ## tRNA charging The `charging` config has one row per 1 s step, like `core`, but no scalars: `sample_id`, `t`, `t_abs`, `aa_pool_size` (fixed_size_list[21], molecules), `ntp_pool_size` ([4], molecules) and `trna_charged_fraction` ([21], fraction of each amino acid's tRNAs that is charged). Join it to `core` on (`sample_id`, `t`). Column k of a list is the row with `column == k` of `provenance/reference/charging_.parquet` (21 amino acids, ATP/CTP/GTP/UTP). It holds only trajectories with index `trna_channels` true: runs converted from rc2 on (0 trajectories); every rc1 row has none, and never will (their raw output is deleted). Coverage: none in this release. Row 0 (`t` = 0) of every list is NaN: the simulator's listener writes an all-zero first row before its first step (checked at export), like `growth_rate`, which is null there. ## index.parquet One row per trajectory, sorted by family (tier order T1..T5), cell_id, seed, generation. | column | type | meaning | |---|---|---| | `sample_id` | string | trajectory id, `lineage_id + "/g{generation}"` | | `lineage_id` | string | `cell_id + "/s{seed}"`: one simulated lineage | | `cell_id` | string | the experimental cell (design x condition), e.g. `T2/argG_KO/basal`; see Ids | | `parent_id` | string | sample_id of the mother generation; null at generation 0 | | `experiment_name` | string | the campaign's experiment name, verbatim (wrong for T4 transcription-factor cells; see Ids) | | `family` | string | one of the five families (the split names of every data config) | | `tier` | string | campaign tier, T1..T5 | | `variant_type` | string | wcEcoli variant name | | `variant_index` | int32 | wcEcoli variant index | | `variant` | string | gene symbol(s), `wildtype`, or the variant type | | `condition` | string | the campaign's condition as designed; ignored by the simulator for TF-activity and sinusoidal cells (read `media_schedule`) | | `timeline` | string | the media schedule that actually ran, `"0 minimal, 1200 minimal_acetate"` (simulation seconds); null for sinusoidal media | | `seed` | int8 | simulation seed | | `generation` | int8 | generation number within the lineage (0 = first cell) | | `knockout_genes` | list | knocked-out gene symbols | | `knockout_gene_ids` | list | knocked-out gene ids (EcoCyc) | | `knockout_tu_ids` | list | transcription units whose expression was set to zero (a knockout removes a whole TU) | | `media_schedule` | list> | media changes, lineage-absolute seconds (`t_abs_sec`) and medium | | `medium_at_birth` | string | medium at the first row of this generation | | `medium_at_division` | string | medium at the last row of this generation | | `media_a` | string | first medium of a two-medium schedule | | `media_b` | string | second medium of a two-medium schedule | | `period_sec` | int32 | period of a sinusoidal schedule, s (null otherwise) | | `tf_target` | string | symbol of the transcription factor that was actually perturbed (regulation family) | | `tf_target_id` | string | EcoCyc id of that transcription factor | | `tf_state` | string | imposed TF state: active, inactive, or control | | `ppgpp_setpoint_um` | float | ppGpp set point, uM (regulation family) | | `ppgpp_factor` | float | multiple of the basal ppGpp concentration (0.2..2.0) | | `aa_shift` | string | amino-acid shift label, e.g. `+ARG`, `-ARG`, `rich->6aa`, `control` | | `perturbation_text` | string | one-sentence plain-language description of the perturbation | | `n_steps` | int32 | rows of this trajectory at 1 s | | `divided` | bool | the cell divided (= not terminated) | | `terminated` | bool | the lineage ended in this generation | | `termination_class` | string | none, time_limit, negative_counts, fba_infeasible, equilibrium_unstable, other | | `translation_collapse` | bool | Cellarium's translation collapse, `ribosome_elongation_rate_mean < 1` (aa/s) in the `summary` config; null for a trajectory without a summary row (see Outcomes) | | `termination_reason` | string | the simulator's termination message | | `trna_channels` | bool | the run's raw output recorded `aa_pool_size`, `ntp_pool_size` and `trna_charged_fraction` (false for every run converted before rc2; see Provenance) | | `division_time_sec` | float | `n_steps - 1` when divided, else null | | `final_mass_fg` | float | cell mass at the last row, fg | | `growth_rate` | float | mean instantaneous growth rate over the generation, 1/s | | `lineage_terminated` | bool | the lineage ended in some generation | | `lineage_generations` | int8 | generations present for this lineage | | `condition_effective` | string | the condition that actually ran: the campaign condition of the medium at t = 0, with `glc_20mM`/`glc_5mM`/`glc_2mM` folded into `basal`; null for sinusoidal media (see Limitations) | | `medium_fitted_for` | string | medium of the condition the first cell was initialised (fitted) from; `minimal` for every wild-type and knockout cell whatever its label | | `shift_at_t0` | bool | the medium at t = 0 is not `medium_fitted_for`: a cell fitted to one medium switched into another at t = 0 (null for sinusoidal media) | | `unfitted_shift` | string | `t0` (= `shift_at_t0`) or `mid_run` (a later switch into a medium other than `medium_fitted_for`, glucose doses as minimal, placeholders excepted) when the cell meets a medium its expression was not fitted for; a shift experiment, not an adapted cell | | `partial_knockout` | bool | a targeted gene kept a transcription unit: the knockout did not silence it | | `knockout_target_tus_missing` | list | transcription units of the targeted genes that were not zeroed (null without a knockout) | | `collateral_genes_full` | list | other genes every transcription unit of which was zeroed (silenced with the target) | | `collateral_genes_partial` | list | other genes that lost some but not all of their transcription units | | `superseded` | bool | replaced by a re-run (every partial knockout); kept for the record | | `superseded_by` | string | cell_id of the re-run that replaces this cell (null until it is published) | | `withdrawn` | bool | a placeholder medium with no mechanism in the model is in effect in this generation | | `withdrawn_reason` | string | why the row is withdrawn | | `design_id` | string | hash of everything the simulator was given (zeroed TUs, fitted-for medium, schedule with glucose doses folded, sinusoid, TF state, ppGpp factor, amino-acid shift); twins share it | | `canonical` | bool | true on one cell per design (keep `canonical` rows to count each experiment once) | | `short_trajectory` | bool | `n_steps <= 10` | | `viable` | bool | divided, not `translation_collapse` and not `short_trajectory`; null when that depends on a null `translation_collapse` | | `split_by_lineage` | string | train/val/test over a design's seeds: every lineage of every cell of a design with one seed is in one split, so twins never cross | | `split_by_tu` | string | train/val/test holding out whole transcription units; a row takes the most held-out split of its `knockout_tu_ids`; non-knockout rows are train | | `job_id` | int32 | campaign job id | | `attempt` | int8 | campaign job attempt | | `started_at` | string | simulation start time | | `run_dir` | string | campaign run directory | | `h5_group` | string | group in the run's channels.h5, e.g. `seed0/gen0` | | `git_hash` | string | wcEcoli commit that ran the simulation | | `parca_cache` | string | Parca (fitted-parameter) cache the run used, by name; the cache itself is not distributed | | `kb_sha256` | string | sha256 of the fitted parameter set (simData.cPickle) used; the set itself is not distributed | | `sim_options_hash` | string | sha256 of the run's simulation options minus per-run keys and the keys carried as columns (git_hash, timeline, multi-knockout indices); key into `provenance/sim_options/` | | `_file`, `_row_offset`, `_n_rows` | string, int64, int32 | locator of this trajectory in each config: file path in the repo, first row, row count (null when the trajectory is not in that config) | ## Product schema Rows of one trajectory are contiguous and in time order within one file. Units are also in each field's parquet metadata (`unit`). | column | type | unit | meaning | |---|---|---|---| | `sample_id` | dictionary | | joins to `index.parquet` | | `t` | int32 | s | seconds since this cell's birth; row 0 is the birth state | | `t_abs` | int32 | s | seconds since the lineage began (continues across generations) | | `cell_mass` | float | fg | total cell mass | | `dry_mass` | float | fg | dry mass | | `protein_mass` | float | fg | protein mass | | `rna_mass` | float | fg | RNA mass | | `dna_mass` | float | fg | DNA mass | | `small_molecule_mass` | float | fg | small-molecule mass | | `cell_volume` | float | fL | cell volume | | `growth_rate` | float | 1/s | instantaneous growth rate (null at row 0) | | `ppgpp_conc` | float | uM | ppGpp concentration | | `aa_supply_total` | float | counts/s | total amino-acid supply rate | | `aa_synthesis_total` | float | counts/s | total amino-acid synthesis rate | | `mrna_counts` | int32 | molecules | total mRNA molecules | | `n_oric` | int32 | count | origins of replication | | `fba_objective` | float | mmol/gDCW/h | FBA objective value | | `exchange_flux_total` | float | mmol/gDCW/h | sum of exchange fluxes | | `reaction_flux_total` | float | mmol/gDCW/h | sum of reaction fluxes | | `ribosome_elongation_rate` | float | aa/s | ribosome elongation rate | | `ribosome_actual_elongations` | float | aa | amino acids elongated in the step | | `mrna_counts_matrix` | fixed_size_list[2773] | molecules | mRNA counts per transcription unit (ids are mRNA TU ids); column k is the row of `provenance/reference/mrna_counts.parquet` with `product_column == k` | | `protein_counts_matrix` | fixed_size_list[4174] | molecules | protein counts per monomer; column k is the row of `provenance/reference/protein_counts.parquet` with `product_column == k` | | `reaction_flux_matrix` | fixed_size_list[3974] | mmol/gDCW/h | flux per metabolic reaction; column k is the row of `provenance/reference/reaction_flux.parquet` with `product_column == k` | | `exchange_flux_matrix` | fixed_size_list[82] | mmol/gDCW/h | flux per exchange reaction; column k is the row of `provenance/reference/exchange_flux.parquet` with `product_column == k` | ## Loading ```python from datasets import load_dataset index = load_dataset("cSynBioSys/well4cell", "index", split="train") core = load_dataset("cSynBioSys/well4cell", "core", split="single_knockout", streaming=True) expression = load_dataset("cSynBioSys/well4cell", "expression", split="wildtype_media") ``` DuckDB, one trajectory: ```sql SELECT t, cell_mass, growth_rate FROM 'hf://datasets/cSynBioSys/well4cell/data/core/*/*.parquet' WHERE sample_id = 'T1/WT/acetate/s0/g0' ORDER BY t; ``` One trajectory by its locator (reads only the row groups it spans): ```python import pyarrow.parquet as pq from huggingface_hub import hf_hub_download row = index.filter(lambda r: r["sample_id"] == "T1/WT/acetate/s0/g0")[0] path = hf_hub_download("cSynBioSys/well4cell", row["expression_file"], repo_type="dataset") table = pq.read_table(path, filters=[("sample_id", "=", row["sample_id"])]) assert table.num_rows == row["expression_n_rows"] ``` ## Ids - `cell_id` is the campaign experiment name, slugified, with the tier as a plain `T1`..`T5` prefix: `[T2_CORE] argG_KO/basal` becomes `T2/argG_KO/basal`. The raw name is in `experiment_name`. - Regulation (T4) transcription-factor cells are the exception: the campaign named them from the sorted TF *symbols* while the simulator indexes the sorted TF *ids*, so their experiment names name the wrong TF (`[T4] araC/active` ran trpR). Their `cell_id`, `tf_target` and `tf_target_id` use the TF that actually ran (`T4/trpR/active`); `experiment_name` keeps the campaign's name. - A knockout removes a whole transcription unit: the model sets the expression of one TU to zero, so other genes of that TU are silenced too. `knockout_genes` lists the targeted genes, `knockout_tu_ids` the TUs actually knocked out. - `lineage_id = cell_id + "/s{seed}"`; `sample_id = lineage_id + "/g{generation}"`, e.g. `T1/WT/acetate/s0/g0`. - `parent_id` is the previous generation's `sample_id` (null at generation 0). ## Time axis - Steps are 1 s. `t` restarts at 0 for every generation; `t_abs` continues across the generations of a lineage, so generation g+1 starts where generation g ended. - Row 0 is the birth state: the daughter at half the mother's mass. `growth_rate` is null at row 0 (no step has been taken yet). - Only one daughter is followed at each division, so a lineage is a chain, not a tree. - A generation stopped by the per-generation time limit (`TimeLimitReached`) has 10,801 steps (0..10,800 s). ## Outcomes `divided` is true when the cell divided. `terminated` marks the generation in which the lineage ended, and `termination_class` says why: | termination_class | meaning | trajectories | |---|---|---:| | `none` | divided normally | 112,545 | | `time_limit` | reached the time limit without dividing: the only clean non-division (slow or arrested growth) | 1,973 | | `negative_counts` | a molecule count went negative (numerical failure) | 2,647 | | `fba_infeasible` | the metabolism FBA problem had no solution | 3 | | `equilibrium_unstable` | the equilibrium solver failed | 6 | | `other` | any other simulator exception | 0 | `time_limit` is the only clean non-division; the other classes are solver failures, and the rows before the failure are kept as recorded. The same outcomes in the `qc` vocabulary of the Cellarium corpus: | well4cell `termination_class` | Cellarium `qc` | meaning | |---|---|---| | `none` | `ok` | the cell divided | | `time_limit` | `no_division` | the generation ran to the time limit without dividing; the model's "did not divide" | | `negative_counts` | `crashed` | simulator error (a molecule count went negative) | | `fba_infeasible` | `fba_infeasible` | metabolism solver failure | | `equilibrium_unstable` | `crashed` | equilibrium solver failure | | `other` | `crashed` | any other simulator error | Only `time_limit` is a prediction of the model about the cell; the other non-`none` classes are numerical failures of the simulation, and a viability task should treat them as such. Cellarium's `implausible_channel` and `over_replicated` have no counterpart here because well4cell does not run those checks; its vocabulary is defined in the Cellarium repository's `docs/QC_STATUSES.md` (https://github.com/evanniko1/cellarium). A generation can divide while its ribosomes have effectively stopped, because chromosome replication does not require ongoing translation. The Cellarium corpus flags this as `translation_collapse` when the mean ribosome elongation rate over the generation is below 1 aa/s, an order of magnitude below the slowest elongation rate measured in *E. coli* (Dai et al. 2016). Here the flag is the index column `translation_collapse`, the same predicate as `ribosome_elongation_rate_mean < 1` in the `summary` config (null for a trajectory without a summary row). It flags 6,273 trajectories in this release, 4,188 of them with `termination_class` `none` (divided, but with translation collapsed). Treat them as inviable in any viability task. ## Suggested tasks 1. **Growth-rate regression from the perturbation**: predict `growth_rate` (index) from the perturbation columns (`knockout_genes`, media, `tf_target`, ...) and optionally the first rows of the trajectory. 2. **Viability classification**: predict whether a generation is *not* terminated by the time limit (`termination_class != 'time_limit'`), i.e. whether the perturbed cell still divides. 3. **Forecasting**: given the trajectory up to time t, predict the channels at t + h, with horizons h stated in seconds (e.g. 10, 60, 600 s) so 1 s and 10 s products compare directly. Use `split_by_lineage` (train 93,703 / val 11,732 / test 11,739) to hold out whole replicate lineages of each design, or `split_by_tu` (train 90,837 / val 12,039 / test 14,298) to hold out whole transcription units; both are fixed columns of the index (seed 20261008) and both keep bit-identical twins on one side. For viability use `viable`, not `divided`; to count each experiment once keep `canonical`; leave out `superseded` and `withdrawn` rows (see Limitations). ## Provenance - Simulator commits (`index.git_hash`): `18a908d28303c2feeba161e962c049e4f712454e`, `5f5fa5f4fe2302af12385528d176f4d73c4bf3a6`, `c41efd7561d40316e5e83b20e3fb27316845b0bc`, `c462ae99c47d24c327a0311edd785f664d8038bc`, `c47683e19614b17e9c022762a238a1b36aa12dd2`, `ddc88250ad5c3ab07872c46888cd18f0fd3d32e6`, `e319d09f3749168155d246b103b7d37ba492ec63`. - Simulation options hashes (`index.sim_options_hash`): `2ba80aae18ba5904625753010432220257822ee442fe87c162cb651d35f5cc0f`. - Fitted parameters (`index.parca_cache`, with the sha256 of its `simData.cPickle` in `index.kb_sha256`): `5558948af4ad806f21d7c80c` (sha256 `8ff93a58576f484edfd6fea641eeeb4fc554b52b97da49f4d8b275ca74496667`), `fa6a9ebd4b1c988a36ef9fa1` (sha256 `b1750cf04409c37ebce9d4d3009daeae0779dd66508812c2807744d83baf1ab7`). The Parca caches themselves are not included in this repo (see Licence). - Every trajectory ran with operons on and the steady-state tRNA charging model (`trna_charging` true, no kinetic charging), from one of two fitted parameter sets identified by `kb_sha256` and `parca_cache`; do not pool rows across the two parameter sets without checking `kb_sha256`. - `provenance/scan_runs.parquet`: every run directory read (32,072 runs), with its generation count and reference-id hash. - `provenance/failed_jobs.parquet`: 971 campaign jobs that failed, with the error message and the generations they left behind. - `provenance/export_qc.parquet`: the export checks of every trajectory in every config (585,870 trajectory x config writes checked, 0 failed): rows landed in the part file, steps are contiguous, counts are valid. A trajectory that fails is left out of that config. These checks say the rows were **written** correctly, not that the cell was **healthy**: the cells' outcomes are in `termination_class`, `translation_collapse` and `viable` (`manifest.json`: `export_checks` versus `cell_health`). - `stats/.parquet`: per product, channel and matrix column: mean, population std, min, max and row count. - `manifest.json`: row and byte totals per config, counts per family. Degradation rates. Not every mRNA half-life in the fitted parameters is a fit to that transcript. The fitted rates are bounded below by the slowest measured mRNA decay (the longest measured half-life), and a unit whose rate sits on that bound is on the floor. Rates faster than the fastest measured decay (the shortest measured half-life) are clipped to it, and those units are on the ceiling. A unit none of whose genes has a measured half-life is imputed: it carries the mean measured half-life. Any claim about the stability of these transcripts, or that depends on it, rests on a bound or a population mean, not on a measurement of that transcript. Which units are in each class is derived from the fitted parameters, so like them it is kept with the campaign and not distributed (see Licence). This repo has the counts only. - `5558948af4ad806f21d7c80c`: of 3,133 mRNA transcription units, 2,293 have a fitted rate and 840, carrying 12.2% of basal mRNA expression, do not: 246 on the floor, 7 on the ceiling and 587 imputed. - `fa6a9ebd4b1c988a36ef9fa1`: of 3,133 mRNA transcription units, 2,293 have a fitted rate and 840, carrying 12.2% of basal mRNA expression, do not: 246 on the floor, 7 on the ceiling and 587 imputed. Never recorded: `aa_pool_size`, `ntp_pool_size`, `trna_charged_fraction` are in the exporter's channel list but were not written by any simulation, so they are absent from every product. ## Licence CC BY 4.0 for everything in this repo. The wcEcoli model is by the Covert lab (Stanford) and is licensed separately under the Stanford Academic Software License; it is not included here. The fitted parameter sets (Parca caches) are not distributed either and are identified only by `kb_sha256` / `parca_cache`, and the same applies to anything derived from them, such as the per-unit degradation-rate classification, of which only the counts are published (see Provenance); reproducing these simulations requires obtaining wcEcoli under its own licence. ## Citation ```bibtex @misc{well4cell, title = {The Well for The Cell: whole-cell E. coli simulation trajectories}, author = {Yermakov, Alexey}, year = {2026}, url = {https://huggingface.co/datasets/cSynBioSys/well4cell} } ``` TODO(authors): the author list above is incomplete; it is completed before 1.0.0.