Use case
Target discovery
Predict structure and conformational states for candidate targets, including the ones with no close homologue in the PDB.
Applications · In development
Four programmes, one of them the priority.
We sell compute today. These are the models we are building on top of it, in order of how much of our own time they take.
Status. Every programme on this page is in development. None of them is a released product, and we do not have a model in production for any of the four. If you need a vendor with a shipped model today, we are not that vendor yet.
01 — Priority programme
Genome sequencing analysis and precision medicine.
Read a whole genome, find the variants that matter, and say which of them a clinician can act on. This is the hardest of the four and the one our roadmap is built around.
What the model has to do
Variant interpretation against clinical evidence; screening for pathogenic sites; matching a patient's variant profile to targeted therapies; and predicting resistance before it shows up in a treatment response.
The output has to be a ranked, sourced statement a lab can put in front of a clinician, with the evidence attached.
Where the data comes from
Public reference resources — 1000 Genomes, ClinVar, COSMIC — give coverage of known variation but not of outcome. The part that matters clinically is the link between a variant, a disease course and a drug response, and that has to come from de-identified clinical data under agreement with the institutions that hold it. Building that corpus is the slow part of this programme.
Why it needs a GPU cluster
Compute profile
- Per sample
- Whole-genome inputs run to TB scale before variant calling
- Pipelines
- GPU-accelerated alignment and variant detection, hours rather than days
- Modelling
- Long sequence context, learned across cohorts rather than one genome at a time
- Inference
- Batch scoring across cohorts, then per-sample service for labs
- Serving
- Diagnostic labs, oncology groups, precision medicine programmes
- Status
- In development — foundation models training, no released product
02 — Programme
Protein structure prediction.
Structure prediction is no longer the difficult part on its own. What teams still need is a pipeline that keeps up with a screening campaign.
Use case
Binding site analysis
Epitope and pocket characterisation, mutation effect on fold and binding, and candidate ranking before anything is expressed in a lab.
Use case
Binder design support
Structure prediction as the filter in a design loop: generate, predict, score, discard. Throughput decides how many designs you can afford to consider.
The compute argument is straightforward: prediction is cheap per structure and expensive per campaign. Model training and high-throughput inference both belong on the same cluster, which is the reason this programme sits under our compute business rather than beside it.
03 — Programme
Small molecule generation.
Plenty of models can generate molecules. Far fewer produce candidates that survive developability, selectivity and synthesis.
Design loop
Concept
- Generate
- Sample chemical space conditioned on the target and the constraints
- Score
- Activity, selectivity, ADMET, toxicity, synthetic accessibility
- Filter
- Multi-objective ranking rather than one score with weights nobody agrees on
- Verify
- Hand the shortlist to a synthesis and assay partner
- Feedback
- Assay results retrain the scoring models
Why we think this is worth doing
The bottleneck in early discovery is not ideas, it is the number of candidates a team can afford to evaluate. A generation loop that runs continuously on owned compute changes the economics of that evaluation — the marginal candidate stops costing a week of someone's attention.
What we are not claiming
No molecule has been designed, synthesised and validated through this programme yet. What is described above is the design of the system we are building.
04 — Programme
Medical imaging analysis.
The one most likely to be useful first, because the models are better understood and the buyers are easier to reach.
Detection and segmentation
Lesion detection, organ and tumour segmentation, and volumetric quantification across pathology, CT and MRI.
Combined with genomic data
Imaging features linked to variant profiles, for studies that need both a picture and a genotype — the reason this programme belongs in the same portfolio as the first one.
Where the compute goes
High-resolution volumes make training memory-bound, and clinical use makes inference concurrent. Both fit the same cluster shape we already run.
05 — Working with us
Three ways to get involved before anything ships.
Research groups and biotech teams do not have to wait for a product to work with us. These are the arrangements we are running now.
01
Compute only
You bring the model and the data; we provide the cluster and the plumbing. No involvement in your science, no claim on your results.
02
Joint development
A defined piece of model work delivered against a milestone, on your data, with the intellectual property terms agreed before the first commit.
03
Data partnership
For institutions holding de-identified cohort data and clinical follow-up. Terms are set by your ethics and legal review, not ours.
Have data and no compute, or compute and no model?
Both are the same conversation. Tell us what you have and what you are trying to answer.