More data, same odds | infinitusbio

In briefThe likelihood that a drug entering Phase 1 reaches approval fell from 9.6% (2006–2015) to 6.7% (2025), even as diligence data multiplied. AI-discovered molecules pass Phase 1 far more often (about 85% vs. 40–65% historically) but succeed in Phase 2 at the same ~40% as everything else. Phase 2 tests efficacy, and efficacy failures usually come from the wrong indication, patients or partner drug. Models trained on past trials inherit that history's biases. Training-free mechanistic simulation, which predicts drug response by reproducing cell biology, addresses the efficacy question directly.

The numbers have moved the wrong way

Roughly nine in ten drugs that enter clinical trials never reach approval. The likelihood of approval from Phase 1 was 9.6% for 2006–2015 (BIO, Biomedtracker, Amplion), 7.9% for 2011–2020 (BIO, Informa, QLS) and 6.7% in 2025 (Citeline). The average cost of one approved drug, was about $2.67B per new drug at the top 20 biopharma companies in 2025 (Deloitte).

Over the same period, the volume of data available for diligence grew enormously: genomics, real-world data, trial registries, expert networks. More of it has not changed the odds.

Better molecules are not better drugs

The first wave of AI in drug development focused on discovery: designing molecules faster. It works for what it targets. A 2024 BCG analysis in Drug Discovery Today (Jayatunga et al.) found AI-discovered molecules passed Phase 1 about 85% of the time, against a historical 40–65%. But in Phase 2, where a drug must show it works in patients, their success rate was about 40%, the same as the historical average.

More than 173 AI-discovered programs were reported in clinical trials by early 2026, roughly doubling every 18 months. By our estimate, about 68% of 2024's $3.8B of AI drug R&D equity funding went to discovery engines. The industry is producing more molecules, not better answers about whether they will work.

It is rarely the molecule

An analysis of clinical trial data from 2010–2017 (Sun et al., Acta Pharmaceutica Sinica B, 2022) attributes 40–50% of clinical failures to lack of efficacy, 30% to unmanageable toxicity, 10–15% to poor drug-like properties and about 10% to commercial and strategic reasons. A drug that lacks efficacy in one indication often has a mechanism that fits another disease, a subgroup defined by a biomarker, or a partner drug that changes the outcome. The failure gets blamed on the molecule when the indication, the patients or the partner were wrong.

Why training on history inherits the problem

Most predictive tools in this space learn from past trials. That brings three biases with it:

Survivorship bias. Programs that were stopped quietly leave little data behind.
Sponsor and publication bias. Positive results are reported more fully than negative ones.
Novelty blindness. A model can only map patterns it has already seen. A new mechanism, a rare disease or an untested combination falls outside its training data, which is exactly where the decisions matter most.

What a mechanistic, training-free approach does differently

infinitusbio does not learn from past outcomes. It initializes a whole-cell simulation from a patient's mRNA, applies a drug's on- and off-targets, and lets more than 200,000 agents simulate billions of reactions. The treated clone is compared with its untreated clone, and the result is read by biomarker response.

Three properties follow. Every prediction is out-of-sample by construction, because nothing was trained. The same input always gives the same answer, because the simulation is deterministic. And every call can be traced to the molecules and pathways that produced it.

In validation, this approach ranked 17 small-cell lung cancer cell lines by drug response with a Spearman correlation of 0.82 against the wet-lab ranking (p = 0.00005), with no training. Trained machine-learning models are not a like-for-like benchmark: their training data carries its biases into every prediction. Across seven tests in cancer and autoimmune biology it agreed with wet-lab outcomes 88–100% of the time. The full results are on the evidence page.

Where it matters most

Rare diseases. About 7,000 diseases, roughly 95% without approved treatment, and too few patients to train any model.
Combinations. About 500,000 two-drug pairs from 1,000 approved drugs, per disease. Trials test a handful.
New mechanisms. First-in-class drugs have no history to learn from.
Hidden indications. PD-1 drugs were first approved in melanoma; most of their sales now come from later indications.

What it does not do

Simulation predicts efficacy; it is not clinical proof. It does not replace clinical trials or regulatory review. Retrospective validations are encouraging because the system never saw those cases, but prospective results remain the stronger test, which is why every market call infinitusbio makes is sealed and timestamped before the event it predicts.

Sources

BIO, Biomedtracker, Amplion. Clinical Development Success Rates 2006–2015 (2016).
BIO, Informa Pharma Intelligence, QLS. Clinical Development Success Rates 2011–2020 (2021).
Citeline (2025, 2026) pipeline and success-rate reports.
Deloitte, Measuring the return from pharmaceutical innovation (2026): average R&D cost per asset, top 20 biopharma companies, 2025.
Jayatunga M.K.P. et al. (BCG), Drug Discovery Today (2024). Historical Phase 1 shown as midpoint of 40–65%.
humAI.blog (2026), third-party tracker of AI-discovered clinical programs.
Sun D. et al., Acta Pharmaceutica Sinica B (2022). Why 90% of clinical drug development fails.
CB Insights and GMI (2024 AI R&D funding); 68% discovery share is an infinitusbio estimate.