Big Picture Bio’s £2.2M World Model for Cancer Combinations: 86% ASCO Call Rate

Cancer is not one disease driven by one target. Resistance pathways re-route. Combinations are how the field already tries to stay ahead — roughly one in three FDA solid-tumour approvals between 2011 and 2023 were combination regimens. The combinatorial space for even a modest library of 100 drugs at 100 doses is tens of millions of experiments. No wet lab can brute-force it.

London-based Big Picture Bio emerged from stealth in mid-September 2026 with a combined £2.2 million (£1.5 million pre-seed co-led by Kadmos Capital and Exceptional Ventures, plus £700k non-dilutive from Innovate UK’s Investor Partnerships Programme). Founders Dr Kerstin Papenfuss (CEO) and Dr Mark Hammond (CTO) spent seven years at Deep Science Ventures building therapeutics companies and agentic discovery systems. Their bet is a generative “world model” that treats tumour, immune cells and microenvironment as a coupled dynamical system rather than a static target list.

What the model actually claims

The platform simulates interactions across disease sites and patient subgroups, then proposes combinations and sequences that are more likely to hold residual disease near zero instead of delivering a temporary response followed by regrowth. Predictions are returned with readable causal chains, not opaque scores. The company reports that all predictions above a 65% confidence threshold have been correct so far.

One concrete call: before Regeneron’s fianlimab melanoma readout, the model predicted both tested doses would miss the primary endpoint and gave a progression-free survival hazard-ratio range of 0.83–0.90 for the high-dose arm. The reported result was 0.845.

Ahead of ASCO 2026 the team published PASS/FAIL and confidence for fourteen phase-3 readouts on the public @bigpicturebio account. Twelve of the fourteen matched the eventual data (company-reported; not yet independently published).

From prediction to wet-lab validation

The near-term use of the capital is to move the first AI-designed combinations — largely built from approved or clinical-stage agents — into laboratory testing. Initial focus is solid tumours where dense single-cell datasets already exist. The longer thesis is that the same modelling stack can eventually design novel agents rather than only recombine existing ones.

Papenfuss has stated the team can already see routes to 50–200% overall-survival improvements in some hard cases if the right existing drugs are sequenced correctly. That claim will be tested by the wet-lab results, not by retrospective trial calls.

Why the approach matters for the broader stack

Most current AI drug-discovery tools optimise a single objective (affinity, ADMET, trial success probability). A world model that explicitly reasons about resistance rewiring and microenvironmental feedback is a different layer. If the causal-chain explanations hold up under wet-lab scrutiny, the platform becomes a filter that shrinks the experimental search space rather than another black-box ranker.

The competitive landscape is crowded at the single-target and generative-chemistry level (Enveda, Insilico, Recursion and others). Big Picture Bio is positioning upstream of that stack — as a systems-level design layer that tells partners which combinations are worth making and in what order.