Computational Modeling vs. the High Cost and Failure Rate of Drug Development

Drug development is one of the most expensive and risky enterprises in modern science. Bringing a single new medicine to market typically takes over a decade and costs in the range of one to several billion dollars. Yet around 90% of candidate drugs that enter clinical trials still fail, most often because they do not work well enough in patients or reveal unacceptable toxicity once tested at scale. Analyses of late-stage trials show that lack of efficacy has overtaken poor pharmacokinetics and toxicity as the leading cause of failure, turning many Phase II and III programs into very costly demonstrations that a target is not clinically relevant.
This combination of long timelines, high costs, and late surprises exposes a structural weakness in the traditional pipeline. Preclinical models—cell lines, animal studies, and small early-phase trials—often cannot predict how a drug will behave across a diverse human population. Species differences, under-powered studies, and limited characterization of dose–response relationships mean that key questions about exposure, target engagement, and variability are pushed into the most expensive part of development: large, randomized clinical trials. When a program fails at that point, years of work and hundreds of millions of dollars are lost, and patients have been exposed to an intervention that was never likely to succeed.
Over the past decade, computational modeling has emerged as a practical way to move some of this risk and exploration out of patients and into mathematically transparent environments. At the discovery stage, structure-based virtual screening and molecular docking allow researchers to sift through millions to billions of virtual compounds and prioritize only a small subset for synthesis and in-vitro testing. By enriching for high-probability binders and discarding weak or non-specific chemotypes early, these tools can substantially reduce the cost and time associated with high-throughput wet-lab screening. AI-accelerated screening is now pushing this frontier further, enabling efficient exploration of ultra-large chemical libraries.
Once a viable mechanism and early lead compounds exist, pharmacokinetic/pharmacodynamic (PK/PD) modeling and Model-Informed Drug Development (MIDD) begin to play a central role. These models integrate preclinical and early clinical data to describe how a drug is absorbed, distributed, metabolized, and eliminated, and how concentration translates into effect (or toxicity) over time. Regulators have started to formalize this approach: recent ICH and FDA guidance on MIDD explicitly encourage the use of quantitative models to inform dose selection, trial design, and labeling. In practice, good PK/PD models allow teams to simulate different dosing regimens, anticipate non-linearities, and choose doses that are more likely to be both effective and safe in subsequent studies—thereby reducing the risk of “dose-finding by trial-and-error” in expensive late-phase trials.
Quantitative Systems Pharmacology (QSP) extends this paradigm by combining PK/PD with mechanistic representations of biological pathways, disease processes, and biomarker dynamics. QSP models can be used to explore questions that are hard to answer in the clinic: Which pathway nodes are most sensitive to intervention? How might combinations behave? What variability do we expect between patients with different genetics, co-morbidities, or immune states? Recent work has applied QSP to areas as diverse as immuno-oncology and CRISPR-based gene editing, using the same model framework to translate between mouse, non-human primate, and human responses. By simulating heterogeneous “virtual patients,” QSP helps identify subgroups likely to respond, flags scenarios where efficacy is inherently limited, and supports rational selection of endpoints and biomarkers before launching large trials.
The next level of integration is in silico clinical trials and digital-twin–style models. Here, mechanistic models and clinical data are combined to emulate not just drug action, but entire clinical studies in a virtual population. Knowledge-based programs such as SIRIUS, for example, simulate the long-term cardiovascular effects of lipid-lowering therapies like inclisiran before outcome trials are complete.  Such emulations can be used to test alternative trial designs, inclusion criteria, and endpoints, and to stress-test safety or adherence scenarios that would be difficult or unethical to examine empirically. They do not eliminate the need for real trials, but they can filter out weak designs, refine hypotheses, and focus resources on the most promising strategies.
Taken together, these methods change where and how uncertainty is handled. Instead of discovering fundamental problems only after hundreds or thousands of patients have been exposed, computational models allow teams to confront many of those issues earlier—on a laptop, not in a ward. Virtual screening cuts down the number of low-value compounds entering the pipeline. PK/PD and QSP modeling sharpen dose and regimen decisions and bring population variability into view before Phase II/III. In silico trials and digital twins enable systematic exploration of “what-if” scenarios and provide a quantitative backbone for adaptive, efficient clinical development. Regulatory agencies are increasingly receptive to these approaches, recognizing their potential to de-risk programs and support more evidence-based decisions.
The core message is simple: the high cost and failure rate of drug development is not just a problem of biology—it is also a problem of insufficient prediction. Computational modeling is not a silver bullet, but it is one of the few tools that directly addresses this predictive gap. By moving part of the exploration and optimization phase into mathematically explicit, reproducible environments, we can protect patients, conserve resources, and increase the odds that the drugs reaching Phase III truly deserve to be there. For an industry facing relentless pressure to be faster, more precise, and more ethical, embracing model-informed strategies is less an option than a necessity.

References

• Waring MJ et al. An analysis of the attrition of drug candidates from four major pharmaceutical companies. Br J Pharmacol. 2015. ResearchGate
• Hingorani AD et al. Improving the odds of drug development success through human genomics and pharmacogenomics. Sci Rep. 2019. Nature
• Sun D. Why 90% of clinical drug development fails and how to improve it? 2022. PMC+1
• Lin X, Li X, Lin X. A Review on Applications of Computational Methods in Drug Discovery and Development. Front Pharmacol. 2020. PMC+2MDPI+2
• Madabushi R et al. Role of Model-Informed Drug Development in the Clinical Pharmacology Review. Clin Pharmacol Ther. 2022. PMC
• FDA & ICH. M15: General Principles for Model-Informed Drug Development (Draft Guidance). 2024–2025. U.S. Food and Drug Administration+2U.S. Food and Drug Administration+2
• Verma M et al. Quantitative systems modeling approaches towards enhancing drug development decisions. Front Syst Biol. 2023. pure-oai.bham.ac.uk+1
• Desai DA et al. A QSP platform for translation of CRISPR therapies from preclinical species to humans. Front Pharmacol. 2024. Frontiers
• Angoulvant D et al. In-silico trial emulation to predict cardiovascular protection of inclisiran. Eur J Prev Cardiol. 2024. OUP Academic+1
• Alasmari MS et al. Model-informed drug discovery and development: a transformative approach. 2024. ScienceDirect+1

Originally published at https://www.linkedin.com/pulse/computational-modeling-vs-high-cost-failure-rate-drug-ramyar-azar-wxmhf