Skip to content
ByLineage 01 · Private preview

Run it dry first.

Drylab (drylab.app) runs experiments on living cells in software: gene knockout and drug screens, from yeast to human cell lines. It triages the wet lab; it does not replace it.

Join the waitlist (opens in a new tab)Read the docs (opens in a new tab)Sign in (opens in a new tab)
Fig. 1A budding-yeast colony growing, with a fluorescent reporter and traced cell outlines.

Section 01 · Inside the reference cell

S. cerevisiae S288C · yeast-GEM v9.1.1[1]

1,788

Cytoplasm

reactions: most of metabolism happens here

Section 02 · From one cell to a screen

Fig. 2A budding-yeast colony growing, with a fluorescent reporter and traced cell outlines.

01 · The subject

  1. 01

    The subject

    Saccharomyces cerevisiae S288C: baker's yeast, the reference strain. A eukaryote, so nucleus, mitochondria, vacuole and peroxisomes are in the model from the start.

  2. 02

    The model

    yeast-GEM v9.1.1: 4,105 reactions, 2,748 metabolites and 1,143 genes across 14 compartments, pinned by checksum so every run uses exactly the same model.

  3. 03

    Perturb

    Delete a gene, or every gene. TP1 ranks a whole genome in seconds; TP2 runs all 1,143 single deletions through flux balance in about a minute.

  4. 04

    Over time

    TP3 adds time. Feed the culture glucose and watch it switch to ethanol when the sugar runs out: the diauxic shift, which nobody programs in; it emerges.

  5. 05

    One by one

    TP4 follows single cells and the variation between them. TP5 puts them in space: colonies on agar, and 3D spheroids with starved cores.

Section 03 · Gene and drug screens

Gene knockouts

Yeast gene essentiality: MCC 0.56 on 1,481 held-out genes.

Delete one gene or all of them, and see which deletions stop the cell from growing.

Checked for yeast gene essentiality against genes the model never saw.

Drug screens

Drug screens have not yet been checked against real data.

Dose 3D spheroids of HepG2, HCT116, MCF7 and K562 and watch where the cells die.

Theory for now: labeled that way on every result until real data says otherwise.

A whole genome, ranked.

Real output from Drylab’s production model: the chance that deleting each yeast gene is lethal. The card shows ten genes taken evenly down all 5,061.

TP1 · Glance · S. cerevisiae S288C

Which deletions are lethal?[2]

TP1 yeast gene essentiality matched held-out data: MCC 0.56 on 1,481 genes the model never saw.
  1. #1 YEL026W1.00
  2. #563 YDR299W0.85
  3. #1125 YDR177W0.60
  4. #1688 YPL059W0.38
  5. #2250 YIL042C0.24
  6. #2812 YLR207W0.17
  7. #3374 YHR053C0.11
  8. #3937 YBR298C0.07
  9. #4499 YOL065C0.03
  10. #5061 YLR240W0.00
Predicted chance each single-gene deletion is lethal. Ten of 5,061 genes, taken evenly down the full ranking, so you see the whole range.

Five tiers, one interface[3]

TP1 · Glance

Which of these genes is likely essential? Ranks a whole genome.

Method
Learned surrogate
Runtime
Seconds

Section 04 · Price before you run

Every run is quoted first.

Runtime, dollar cost and checked-or-theory status, before anything runs. Nothing is deducted from your allowance until you confirm, and a workspace budget is a hard stop: a run that would cross it does not start.

Quote · TP1 · Glance

S. cerevisiae S288C · All single deletions[4]

Matched real data the model never saw.
Estimated time
24 s
Estimated cost
$0.00086

Actual: 1.7 s, $0.000031.

Confirm and run (opens in a new tab)

Section 05 · Reproducible by default

Same inputs, same bytes.

Original · production

run
02fe2195
model sha256
34767ee0…80a95
seed
0
engine
vcl-surrogate
numpy · sklearn
2.5.4 · 1.9.1
python
3.12.15
machine
Linux x86-64 (Cloudflare)
result sha256
287ec280…b93f1c

Rerun · from the record

run
02fe2195
model sha256
34767ee0…80a95
seed
0
engine
vcl-surrogate
numpy · sklearn
2.5.4 · 1.9.1
python
3.12.13
machine
macOS arm64 (a laptop)
result sha256
287ec280…b93f1c

0 bytes differ[5]

Different machine, different operating system, different Python. Same model checksum, same seed: 511,148 identical bytes.

Section 06 · Validation

0.56

MCC · yeast gene essentiality · 1,481 held-out genes[6]

What MCC is. The Matthews correlation coefficient scores yes-or-no predictions, here “is this gene essential?”, on a scale from −1 to 1. Zero is a coin flip; 1 is perfect. It stays honest when one answer is much rarer than the other, which essential genes are.

Why held-out matters. The 1,481 genes were never used to train or tune the model. That is what makes this result checked rather than theory.

Section 07 · Drylab's principles

  1. 01

    Same inputs, same bytes.

    A result you cannot rerun is a rumour. Every run is pinned to its model, seed and engine versions.

  2. 02

    Price before you run.

    You see the runtime, the cost and the trust status before anything happens, and budgets are hard stops.

  3. 03

    The bench has the last word.

    Drylab ranks which experiments to run first. The wet lab decides what is true.

Join the waitlist (opens in a new tab)Read the docs (opens in a new tab)Sign in (opens in a new tab)

A screening and triage tool for research. It does not replace wet-lab experiments. Not for clinical use.

Section — · Sources

Sources

  1. [1]yeast-GEM v9.1.1, SysBioChalmers, the genome-scale model Drylab pins by checksum. Source ↗
  2. [2]Drylab TP1 (Glance) output for S. cerevisiae S288C, model vcl-tp1-go-essentiality v1: all 5,061 single deletions ranked (production run 02fe2195). Source ↗
  3. [3]Drylab capability matrix: the five compute tiers and the cells each serves. Source ↗
  4. [4]The quote and the run record of that same production run: estimated 24 s and $0.00086; actual 1.7 s and $0.000031 used. Source ↗
  5. [5]Rerun of production run 02fe2195 (TP1, S. cerevisiae S288C, all single deletions) from its record alone, on a different machine, 26 October 2026: identical SHA-256, 511,148 bytes. Source ↗
  6. [6]Drylab validation report: yeast gene essentiality, MCC 0.56 on 1,481 held-out genes. Source ↗