The Drug Discovery Landscape: Economics, Failure Rates, and Paradigm Evolution

The single most underappreciated fact in pharmaceutical strategy is that the average capitalized cost of developing a new molecular entity has grown faster, in real terms, than the gross domestic product of most nations. The Tufts Center for the Study of Drug Development places the figure at approximately \$2.6 billion as of the early 2020s, a number that, even after adjusting for the well-documented methodological controversies surrounding its derivation, still dwarfs the cost of any other industrial R&D endeavor per unit of output. Each year the global pharmaceutical industry expends roughly \$280 billion on research and development, yet produces fewer than 50 new molecular entities approved by the U.S. Food and Drug Administration. The implied cost per approved drug has therefore risen by an order of magnitude since the 1970s, even as the underlying sciences — structural biology, genomics, high-throughput screening, combinatorial chemistry, cloud computing, and now deep learning — have advanced at compounding rates. This paradox, captured in Jack Scannell's "Eroom's Law" (Moore's Law spelled backward), is the defining economic fact of drug discovery. Any rigorous program of inquiry into how drugs are actually discovered must begin by quantifying the paradox precisely, identifying its proximate and distal causes, and understanding why each successive wave of technological optimism — recombinant DNA in the 1980s, combinatorial chemistry and high-throughput screening in the 1990s, target-based genomics in the 2000s, and AI-augmented discovery in the 2020s — has produced at best a transient flattening of the curve rather than a true reversal.

1.1 The Productivity Paradox: From Eroom's Law to the Cost-per-Launch Frontier

To grasp the gravity of the productivity problem, one must distinguish between three quantities that industry discourse frequently conflates: out-of-pocket R&D expenditure per approved drug, capitalized R&D expenditure per approved drug, and the marginal cost of any individual program conditional on its therapeutic area, modality, and stage. The most cited headline figure, \$2.6 billion capitalized, is derived by Tufts CSDD from a survey of 10 companies that contributed data on 165 investigational compounds first tested in human subjects between 2005 and 2013. The methodology multiplies out-of-pocket spending by a "cost of capital" factor (approximately 1.5× to 1.7× depending on assumptions) to account for the opportunity cost of capital tied up over a 10–15 year development cycle. Critics — most prominently Avorn and colleagues in *BMJ* — have argued that the figure is inflated by tax deductions, that it overstates true private-sector cost, and that a more conservative estimate would be closer to \$1.4 billion. The truth is that both figures are defensible under different assumptions, and the appropriate interpretation depends on whether one is asking the question from a society's perspective or from a portfolio manager's perspective. For internal portfolio decisions, what matters is not the precise magnitude of the average but the distribution of program outcomes, the shape of which has been remarkably stable across decades.

The shape of that distribution is the central object of study. It is a heavy-tailed distribution in which the modal program loses the entire capital allocated to it, a small minority produce blockbusters with peak sales above \$5 billion annually, and a long middle of "modest successes" generate enough revenue to keep the pipeline funded but rarely to retire the cost of capital on their own. This skewed distribution implies that the median NME is unprofitable in risk-adjusted terms, and that the entire industry depends on the right tail of its own output to remain solvent. The corollary, first articulated explicitly by Bernard Munos and analyzed extensively by Scannell and others, is that the industry's apparent cost disease is not primarily caused by rising input prices (chemist salaries, instrument costs, clinical trial overhead) but by declining output per unit of input — that is, declining research productivity in its technical-economic sense.

Three explanations dominate the literature. The first is the "low-hanging fruit" hypothesis: the easy therapeutic targets (histamine receptors, proton pumps, HMG-CoA reductase, cyclooxygenases) were largely exhausted by the 1990s, and the targets that remain — central nervous system receptors, tumor suppressors, protein-protein interactions, undruggable oncogenes — are intrinsically harder to drug. The second is the "regulatory burden" hypothesis: the bar for clinical and regulatory approval has risen, with greater emphasis on comparative effectiveness, safety surveillance, and outcomes-based reimbursement, all of which inflate the cost of late-stage trials. The third is the "throw-more-money-at-the-problem" hypothesis, which holds that as companies became larger and more diversified, the political economy of internal R&D began to reward risk aversion, me-too programs, and incremental innovation over genuinely transformative targets. There is empirical support for each, and the actual productivity decline is probably best modeled as a multiplicative product of all three.

What the framing makes clear is that the central strategic question for any drug hunter is not "how do I do better chemistry" but "how do I allocate finite exploratory capital across programs with radically asymmetric, fat-tailed outcome distributions, in such a way that the expected return on the portfolio exceeds the cost of capital." The rest of this chapter builds the analytical machinery required to answer that question.

1.2 Anatomy of Modern Development: Costs, Timelines, and Probability of Technical Success

The drug development pipeline is conventionally partitioned into four human-testing phases separated from preclinical discovery by an IND (Investigational New Drug) or equivalent regulatory filing. The phase definitions are themselves a useful taxonomy but conceal important heterogeneity: a Phase I program in oncology with a cytotoxic agent is structurally different from a Phase I program in cardiometabolic disease with a small molecule intended for chronic dosing. The probability of technical success (PTS) at each phase — the fraction of programs entering a phase that ultimately reach the next milestone — is therefore reported both as a global average and as a stratified distribution by therapeutic area.

The most comprehensive public dataset on PTS is the BIO/Biomedtracker/Pharmaprojects collaboration, which has tracked more than 9,000 development programs since 2011 and is updated annually. The most recent published summary produces the following headline phase-by-phase success rates for all therapeutic areas and modalities combined: Phase I to Phase II ≈ 53%, Phase II to Phase III ≈ 35%, Phase III to filing ≈ 58%, filing to approval ≈ 85%. The cumulative probability from first-in-human to first launch is therefore approximately:

$$P_{\text{launch}} = 0.53 \times 0.35 \times 0.58 \times 0.85 \approx 0.092$$

That is, fewer than 1 in 10 programs that enter the clinic will reach the market. The figure is markedly lower in some therapeutic areas and notably higher in others. Oncology, despite having the largest pipeline by program count, exhibits one of the lowest cumulative PTS, around 5–6%, dragged down by the long Phase II bottleneck where most attrition occurs. By contrast, vaccines against infectious diseases and rare-disease gene therapies have historically exhibited cumulative PTS in the 15–25% range, although the gene therapy figure is biased by small samples and short follow-up.

The diagram below captures the structure of the pipeline numerically, making the attrition explicit. The branching structure should be read as a simulation of 100 programs entering Phase I under average BIO/Biomedtracker phase probabilities.

```mermaid flowchart LR A["Preclinical / IND"]