Fail the wrong molecule while failing is still cheap
Visium builds the decision layer around discovery. Triage, prediction and evidence assembly that shorten the list, expose liabilities earlier, and leave a rationale someone can follow twelve months later.
The discovery decision problem
Discovery rarely fails for lack of candidates. It fails on the order in which you look at them.Every one of the four pressure points below is a sequencing problem. The information that would have changed the decision usually existed. It arrived in the wrong month, in the wrong format, or without anyone free to weigh it.
Triage volume
Assay and screening output now exceeds the hours available to read it. Prioritisation falls back on potency because potency is the number everyone has, and the rest of the profile waits for a later stage
Developability found late
Expression, formulation and scale-up problems surface after selection. At that point the fix is a new campaign rather than a different ranking, and the cost is measured in quarters.
Toxicity and ADMET attrition
Failures land after the most expensive work is finished. The early signal was often present, in a form that no one had the time or the tooling to weigh against activity.
Signal buried in noise
Target and biomarker evidence sits across omics data, internal reports, clinical notes and published literature. No single team reads all of it, so the strongest evidence is often the most recently seen.
And it has to hold up later
Each of these decisions is revisited under pressure, sometimes years afterwards, by people who were not in the room. A ranking nobody can reconstruct is a ranking nobody can defend, which is why traceable rationale is a design requirement here rather than a feature.
Nine decisions. We start most engagements with three of them.
The three below carry the clearest cost of being wrong and the shortest path to a measurable result. They are also where a decision layer beats a bigger model.
Compound screening triage
Rank hits on the full profile rather than the single strongest number. The list that reaches your chemists is shorter, and every candidate that was cut carries the reason it was cut.
A ranked shortlist with the criteria, weights and evidence behind each position.
Developability & manufacturability
Score candidates against expression, formulation and scale-up constraints during selection. CMC risk becomes a selection criterion instead of something development discovers on your behalf.
A developability view per candidate, with the constraint that drives the score named.
Preclinical study design
Bring the evidence from prior studies into design choices before protocol lock, including the studies that did not work and the reasons they did not.
Design options with precedent attached, and a record of why each choice was made.
Target identification & validation
Assemble and rank target evidence across omics, internal reports and literature, with provenance held per claim.
Generative molecule design
Propose structures against a target profile and your own constraints, with the filters that shaped each suggestion visible.
Protein structure & binding prediction
Structure and binding estimates used to decide which hypotheses earn wet-lab time, with confidence stated per prediction.
Biomarker discovery
Detect and rank candidate biomarkers in evidence that was never structured for the question being asked.
Toxicity & ADMET prediction
Surface liabilities while a change of direction still costs weeks, alongside activity rather than after it.
Protein engineering
Prioritise variants against activity, stability and expression targets, and keep a usable record of what was tried.
What we bring to a discovery programme
Discovery teams do not need another general model. They need scientific judgement encoded in something reviewable, running on their own data, inside their own environment.
Domain models trained on your data
Our peptide work is the pattern. A sequence-to-sequence model learned to read free-form customer order language and convert it into a manufacturer's proprietary peptide code, trained on that company's own historical orders and the conventions its specialists actually use. Scientific notation is a language, and it is learnable per organisation.
Evidence pulled out of unstructured text
With Idorsia we built DiSMOL, which reads physicians' free-text clinical notes and quantifies signals of daytime impairment that diagnosis codes never capture. The same problem shape appears across discovery: the evidence exists, in prose, unusable at scale.
Rationale you can reconstruct
Every ranking, score and recommendation links back to the data and the criteria that produced it. A scientist can see why a candidate moved down the list, and can say so in a review months later without rebuilding the analysis.
Getting past the pilot
Most discovery AI stalls after a promising proof of concept. In one pharma engagement we helped build the operating model behind more than 265 use cases, cutting proof-of-concept execution time by 30% and increasing the share of solutions reaching production by more than 50%.
What we can show you
Discovery buyers ask for validated performance, published methods and named collaborations. Here is what we have on record.
%20(1).webp)
%20(1).webp)
.webp)