Preclinical pipeline decisions: prioritising candidates before the expensive phases begin
Every compound pipeline is a chain of expensive, sequential decisions. We build models that bring all available evidence into those decisions – with probabilities and uncertainty intervals rather than a point estimate nobody can interpret.
The situation
Hundreds of candidates enter a pipeline, a fraction reaches the end, and the cost of failure rises at every stage. A substantial share of late discontinuations traces back to ADME properties and toxicity signals that could in principle have been spotted earlier.
The problem is rarely missing knowledge but scattered knowledge. Assay results, ADME profiles, toxicology and structure-activity data sit in separate systems and in the heads of individual experts. Decisions at the gates are therefore often made on whichever data points arrived last – not on the evidence the organisation already holds.
Typical data basis
Assay and screening results, ADME profiles, toxicological findings, structure-activity data and historical project records.
The part most often missing and most important: the discontinued candidates. A model that learns only from successful projects cannot predict what fails. Documented no-go decisions from the past are the real training basis.
Why such initiatives fail
Survivorship bias in the training data. Where discontinued candidates are undocumented or unfindable, the model learns what successful projects look like – not what others failed on. Such models say "proceed" to almost every new candidate.
Temporal data leakage. Models are often validated against data that did not exist at the actual point of decision: later assay results, retrospective annotations. That produces excellent metrics and a model that is useless in operation. Validation has to reconstruct the state of knowledge at each gate.
The model answers the wrong question. A model predicts efficacy while the gate decides on developability. Scientifically sound, operationally irrelevant – and the reason we start with the decision point rather than with the data.
Approach
The decision point first. Which decision should improve, who makes it, and how will "better" be measured? Only then does it become clear which data sources matter. This prevents models that are statistically convincing and operationally useless.
Features that stay interpretable. Feature engineering together with your chemistry and biology functions. Results have to hold up in scientific review boards, and for that it must be traceable what a model weights and how.
Predictions with stated uncertainty. Probabilities with intervals per candidate, plus a breakdown of the driving features. A point estimate without a confidence measure is worthless for a go/no-go decision.
Validate against historical decisions. The most convincing test is not a validation set but the question: would the model have made the last few years' decisions better than the process that was actually used?
What you get
A prioritised candidate list with confidence measures, a traceable justification per candidate, and an assessment of how the model would have performed against historical gate decisions. Embedded in your existing decision process, not beside it.
Where our experience comes from
We have quantified decision-making in compound pipelines for international life science groups – in crop protection, where the structure is identical: many candidates, sequential gates, expensive trial series, decisions under uncertainty.
Last updated: 4 August 2026
