The $2.6 Billion Lesson: What Pharma’s Failed Programs Are Trying to Tell Us

Every terminated drug program leaves behind a fingerprint of why it failed. AI can now read those fingerprints — and turn them into the blueprint for the next success

Pharmaceutical R&D generates failure at an extraordinary rate — and buries it. >90% of discovery programs never reach an IND. ~70% of Phase II trials fail – the clinical valley of death. $2.6B, the average cost per approved drug (DiMasi et al., 2016). The most important data the industry produces almost never gets systematically analyzed. That is the problem AI can now solve.

These numbers represent an industry that has, historically, treated failure as a cost to be absorbed rather than a signal to be learned from. The data generated by every terminated target program, every abandoned lead series, every failed clinical trial disappears into lab notebooks, LIMS databases, archived clinical study reports, and institutional memory that walks out the door when scientists move on.

The core insight driving the next era of pharmaceutical AI is this: failure data is not noise to be discarded. It is the highest-signal training data in existence — because it defines precisely where the boundaries of biological and chemical possibility lie.

Two Pipelines, Two Failure Landscapes

Drug discovery and drug development are distinct scientific challenges with distinct failure modes — and therefore require distinct AI approaches.

In drug discovery — from target identification through lead optimization — failure is molecular and largely invisible. It lives in the SAR series that hit a toxicity cliff at a specific molecular weight, the target program terminated because the protein was undruggable, the HTS campaign that generated 50,000 apparent hits and zero tractable leads. This data almost never gets published. It sits in proprietary archives, siloed by program, never analyzed across the portfolio.

In drug development — from Phase I through regulatory approval — failure is clinical and more structured. It lives in clinical study reports, safety narratives, adverse event databases, and the biomarker analyses from Phase II trials where the drug worked in some patients but not the enrolled population.

The shared problem: In both cases, the most valuable failure data — the internal records of what was tried, why it failed, and what was learned — is never systematically mined. AI can now change that. Not by predicting the next failure, but by reading every past failure and extracting the pattern that makes the next program smarter.

What AI Reads in a Failed Program

Applied retrospectively to a pharmaceutical organization’s internal data, AI performs three distinct functions that traditional program review cannot:

• Failure diagnosis at source: Using retrieval-augmented generation (RAG) over ingested clinical study reports, lab notebooks, and assay records, AI can answer questions like “why did compound X fail — was it target engagement, ADMET, off-target toxicity, or patient selection?” with specific evidence citations rather than institutional recollection.
• Cross-program pattern recognition: Patterns invisible at the program level become statistically robust when analyzed across a full portfolio. Compounds with a specific pharmacophore consistently fail hERG screening at a particular potency threshold. Targets in a specific kinase family reliably exhibit selectivity failure. Trials in an indication without a validated biomarker show 3x higher Phase II attrition. These patterns transform program design — but only an AI system analyzing the full historical portfolio can surface them.
• Repurposing from the failure archive: A compound that failed in oncology because it lacked CNS penetration may be the ideal candidate for a neurological indication where that penetration is required. A target program terminated due to on-target toxicity in one tissue may be valid in a disease confined to a different tissue. The mechanism of action data, off-target binding profiles, and responder subgroup analyses from failed programs are the raw material for repurposing — and AI can score those systematically against the full landscape of disease biology.

The Evidence from AI-Native Discovery

This is not theoretical. Organizations that have built failure data into their AI infrastructure are producing measurable results:

• Insilico Medicine used PandaOmics — which explicitly mines failed target programs — to identify and progress ISM001-055 for idiopathic pulmonary fibrosis from target identification to Phase IIa in approximately 18 months, versus an industry average of 3-6 years. Results were published in Nature Medicine; a $2.75B Eli Lilly partnership followed.
• ClinicalReTrial (Xing et al., 2025), a multi-agent AI system applied retrospectively to 101,145 clinical trial records, improved 83.3% of failed trial protocols with a mean success-probability gain of 5.7% — at a cost of $0.12 per trial.

The Solix EAI Pharma Approach: Internal Data Sovereignty

The critical differentiator in pharmaceutical AI is not access to the best published literature — every competitor has access to the same published literature. The differentiator is the ability to learn from your own failure history, at depth, while keeping those insights proprietary.

Solix EAI Pharma is built for exactly this. The platform ingests and governs the full spectrum of pharmaceutical dark data — electronic lab notebooks, LIMS databases, SAR archives, clinical study reports, regulatory filings, email communications around program termination decisions — and applies a governed AI layer that makes that data queryable, analyzable, and actionable.

The result is three compounding capabilities:

• Failure post-mortem on demand: Scientists can ask, in natural language, why a specific compound or program failed — and receive evidence-grounded answers sourced from the organization’s own historical records.
• Portfolio-level failure intelligence: Leadership can see, for the first time, the systematic failure patterns across the organization’s full discovery and development history — and use those patterns to set better stage-gate criteria for active programs.
• Repurposing dossiers from the failure archive: The AI scores every terminated asset against the current landscape of unmet medical need, generating ranked repurposing hypotheses with mechanistic rationale and regulatory pathway assessment.

Every terminated program, rather than representing a sunk cost, becomes a strategic asset that makes every future program smarter. The organization’s failure history is its most defensible competitive advantage — if it can be read. Solix EAI Pharma makes it readable.

References:

• DiMasi, J.A. et al. (2016). ‘Innovation in the pharmaceutical industry: New estimates of R&D costs.’ Journal of Health Economics. DOI: 10.1016/j.jhealeco.2016.01.012 — Foundational $2.6B cost estimate.
• Xing et al. (2025). ‘ClinicalReTrial: Clinical Trial Redesign with Self-Evolving Agents.’ arxiv.org/pdf/2601.00290 — Multi-agent system improving 83.3% of failed trial protocols.