Computational biotech consulting
Most computational drug-discovery results are not wrong because the science is hard. They are wrong because the model was never calibrated, the test set leaked, the “novel” chemotypes were pan-assay artifacts, or nobody ran a negative control. We find that out before it costs you a wet-lab budget or a term sheet.
Thirty minutes, no obligation. If you do not need us, we will tell you on the call.
What we do
Fixed scope, fixed fee, and a written deliverable you can hand to an investor or re-run yourself.
For teams about to bet on a model’s output
For teams bringing screening in-house
For teams who need the iron, not a cloud bill
For teams with a campaign and no cluster
Why trust us with this
Every number below is one we found in our own pipeline and retracted in public. A firm that audits its own claims this hard is the firm you want auditing someone else’s.
Our own CYP inhibition AUROC, once it was tested on chemistry the model had not effectively seen. We retracted the headline rather than keep selling it.
A pooled knockout result that passed — until we found the panel was contaminated. Retracted, and the labelling rule rewritten so nothing is ever inferred from a compound’s name.
Our own targets that currently clear our own adversarial gates. Not one. We could change that this afternoon by loosening them — but the gates are the product.
We took PARP1 from a raw PDB structure to a calibrated, decision-grade target card: discrimination AUC 0.9886 against an ECFP4 baseline of 0.8685, and 12 of 12 held-out actives clearing the calibrated threshold in a blinded test.
The part worth reading is the part that failed. Our pocket-knockout control came back inconclusive because it was underpowered, and we reported it that way instead of rounding it into a win. We also found no defensible novel lead in that campaign — so we did not claim one.
How an engagement runs
Thirty minutes. You describe the decision; we tell you whether we can help and roughly what it takes.
Fixed fee, fixed deliverable, a date. If the scope is wrong, we would rather find out here than halfway through.
Run against pre-registered criteria, so the verdict is written before the numbers exist and cannot drift to suit them.
Versioned engine, inputs and outputs. Your team — or your investor’s — can reproduce every number in it.
That is usually the one worth thirty minutes.