We build and run prediction models for biotechs and AI drug-discovery startups. Compressed fingerprints, ADMET and safety endpoints, screening pipelines. Fixed scope, fixed price, results in weeks.
Each is a project with a defined input, a defined deliverable, and a price agreed before anything starts.
You have assay data and want predictions for compounds you have not made yet. We build the model on your data, benchmark it against public baselines, and hand over the trained model, the training code and the validation report. If the data will not support a useful model, you hear that in week one rather than week eight.
You get: model, code, validation reportSend a compound list, get back predicted absorption, metabolism, permeability and safety-liability scores, each with a confidence estimate and a short written read on what to deprioritise. Used to cut a synthesis list before anyone orders reagents.
You get: ranked scores, written readSimilarity and model-based screening across large commercial or in-house libraries, with the molecular representation tuned to your chemistry rather than taken off the shelf. Output is a ranked, de-duplicated shortlist with the reasoning behind the ranking.
You get: shortlist, selection rationaleThe thing your team keeps meaning to build: reproducible featurisation, data versioning, model training and retraining, and scaling to HPC, Ray or Databricks when compound counts get large. Delivered as a repository your team owns and can run without us.
You get: a repository you ownEvery small-molecule programme has to clear hERG. Wet-lab panels cost thousands per compound and weeks of calendar time. Score the list first, then pay only for the assays worth running.
For the first few teams, send compounds whose hERG results you already hold. We score them without seeing your results, then you compare. Free, or heavily discounted for a large set.
If the predictions do not match your data, you have lost nothing and we have learned something worth knowing. If they do, you have a number you can act on and we have a case study. That is the whole reason for running it blind.
hERG has more public data behind it than almost any other safety endpoint, and every programme is obliged to check it. That makes it the honest place to start: we can show our numbers against free public tools, on data anyone can inspect, rather than asking you to take a private benchmark on trust.
From $2,000 per batch, or included in a monthly retainer. Larger and repeat sets are quoted per project.
We publish benchmark numbers against free public tools as soon as they exist. If a free tool does the job for your chemistry, we will say so, and you should use it.
Ask about a pilotThe rest of this page describes what we can build. This part describes what has not been demonstrated, because you would find it out anyway and it is better said here.
The library has been exercised on materials and quantum-chemistry datasets: redox-flow battery candidates, non-fullerene acceptors, QM9. A drug-discovery benchmark is in progress and will be published as a preprint with the code to reproduce it. Until then, treat accuracy on pharma data as unproven.
This is a new practice. There are no logos to show, and none have been invented for this page. The open pilot above is how the first case study gets made, and the first teams get the better deal for taking that risk.
Against bit-packed fingerprints the storage gain is parity, not a saving. What is guaranteed is the feature count; the speed and accuracy effects that follow from it depend on your model, data and hardware. The documentation says so at length rather than burying it.
Projects are delivered by the author of the library, not routed to a junior. That is the strength and also the constraint: capacity is finite, and we would rather decline work than quote a timeline we cannot hold.
Ordered by what a programme stands to lose if the endpoint is missed. How much public data exists decides whether a model can be built honestly at all, so it is stated for each one.
Where failure is expensive and arrives late. These carry the screening work on their own.
| Endpoint | Why it matters | Public data | Status |
|---|---|---|---|
| hERG blockade | Cardiac risk, checked before IND. Wet-lab panels run into thousands per compound. | Strong | Validating now |
| Mutagenicity (Ames) | Regulatory requirement; an in-silico call is already accepted under ICH M7. | Strong | Next |
| CYP450 inhibition | Drug–drug interaction risk across 3A4, 2D6, 2C9, 2C19 and 1A2. | Strong | Next |
| Other cardiac channels | Nav1.5 and Cav1.2, needed for a full CiPA-style cardiac read. | Moderate | On request |
| Hepatotoxicity | A leading cause of late-stage attrition and post-approval withdrawal. | Moderate, noisy | On request |
| Off-target panels | Kinase and GPCR selectivity. Panel screening alone runs to tens of thousands per compound. | Good for kinases | Research |
Less dramatic, asked for constantly, and cheap to bundle into a screening batch.
| Endpoint | Why it matters | Public data |
|---|---|---|
| Aqueous solubility | Blocks formulation and assay work before either can start. | Strong |
| P-glycoprotein | Efflux, brain exposure and interaction risk; substrate and inhibitor. | Good |
| Metabolic stability | Microsomal clearance decides whether dosing is feasible at all. | Moderate |
| Permeability | Caco-2 and MDCK, for oral absorption. | Moderate |
| Plasma protein binding | Free fraction is what drives efficacy. | Moderate |
| Blood–brain barrier | Make-or-break for a CNS programme. | Moderate |
logP and logD, pKa, topological polar surface area, molecular weight, rotatable bonds, Lipinski and Veber compliance, QED drug-likeness, synthetic accessibility, PAINS and structural alerts.
Short and bounded by design. You can stop after the second step and keep what you paid for.
Thirty minutes. You describe the chemistry and the decision you are trying to make. We say whether modelling helps.
One week, fixed fee. A written memo on what is achievable, what is not, and what it would cost. Yours to keep.
Fixed price against a fixed deliverable, agreed in writing before any work begins.
Weekly updates with results as they arrive, not silence until the end.
Code, trained models, documentation and validation. An optional retainer if you want us on call afterwards.
Published so you can decide whether to call. Each figure is a starting point, quoted properly once we understand the problem.
| Package | What you get | Typical duration | From |
|---|---|---|---|
| Feasibility review | A written memo on what a model could realistically achieve with your data. No model built. | 1 week | $1,500 |
| Screening batch | Up to 5,000 compounds scored on agreed endpoints, ranked shortlist and written read. | 1–2 weeks | $3,000 |
| Custom model build | One endpoint on your data: trained and validated model, training code, validation report. | 4–6 weeks | $12,000 |
| Pipeline build | Reproducible featurisation, training and scaling, handed over as a repository you own. | 6–10 weeks | $20,000 |
| Retained support | Agreed days per month for modelling, screening and advice. | 3-month minimum | $2,500/mo |
Invoiced from the EU and VIES-registered, so cross-border invoices within the EU are zero-rated. Compound structures stay confidential, NDAs are signed as standard, and work can run on your own infrastructure if structures cannot leave it.
FPembed is MIT-licensed and public. Read the code, run it on your own molecules, and form your own view on whether the people behind it know what they are doing.
# pip install fpembed
from fpembed import EmbeddedFingerprintGenerator
gen = EmbeddedFingerprintGenerator(
fp_type="ecfp", fp_size=2048,
compression=16, fp_params={"radius": 2},
)
emb = gen.GetFingerprintFromSmiles("CCO")
emb.shape # (128,)
# Everything needed to reproduce this result later:
gen.representation_manifest()Ten fingerprint types through one class. Six compression methods, from block-wise weighting to Hadamard and random projection. SMILES, SELFIES and RDKit Mol objects as input. Drop-in with any scikit-learn model.
Source on GitHub · Package on PyPI · Archived release with DOI
That last one matters more than it looks. If you have ever had to explain to a reviewer or a regulator exactly which representation produced a result, you already know why a folder of feature vectors is not enough on its own.
Author of FPembed. PhD in engineering, MSc in chemistry. He writes the models, runs the projects and answers the calls.
[Mikhail’s own paragraph goes here: why he built FPembed, the problem that kept recurring in his own work, and why the existing tools did not solve it.]
A different conversation from the one above, and we are actively looking for both.
We can join as the named technical partner for a computational work package: molecular representation, property and safety prediction, pipeline engineering. There is an existing open codebase and a public release history behind it, which makes us low-risk to write into a proposal.
Horizon Europe and Innovate UK both fund work on alternatives to animal testing and on computational safety assessment. Those calls need someone who can own the in-silico deliverable.
Send us the call reference and the deadline.
We are looking for a university group or CRO with hERG or off-target panel data, in exchange for modelling work at no cost. You get models built on your data and a co-authored publication; we get the validation that no amount of public data can substitute for.
This is a real offer rather than a line on a page. It is the fastest route to the benchmark the section above says we do not yet have.
Tell us what your dataset covers.
Thirty minutes with the person who would do the work. Bring the chemistry and the decision you are trying to make. If modelling is the wrong tool for it, we will tell you on the call and you will have lost half an hour.