Why Drug Discovery Keeps Failing and What Deterministic Convergence Reveals About the Path Forward
By Matthew Hardy
Founder & CEO, NomosLogic Inc.
Author: Deterministic Convergence: Biological Systems, Architecture, and the Search for Hidden Order
Author: The Adaptation Paradox: How Evolution's Gifts Became Medicine's Problems (Second Edition, 2026)
The pharmaceutical industry spends $2.6 billion, on average, to bring a single drug to market. Approximately 90% of oncology drug candidates fail in clinical trials. The majority of those failures occur in Phase II and Phase III, where the costs are highest and the patients are real.
The industry's response to this failure rate has been consistent for two decades: better targets, better models, more data, bigger screens, faster AI. The underlying assumption has remained unchanged. We have not found the right target yet. Keep looking.
I believe the assumption is wrong. Not occasionally wrong. Structurally wrong, in the majority of disease systems the industry is trying to drug.
The data from 26 independent deterministic evolutionary simulations across cardiovascular, neurological, oncological, and hepatic domains supports a different conclusion. The problem is not finding the right target. The problem is the assumption that a single target is sufficient.
This article presents the evidence, the implications, and the path forward.
The Single-Target Assumption
Modern drug discovery is organized around a premise: identify the molecular target that drives a disease, develop a compound that modulates that target, and the disease will respond. This premise has produced transformative medicines. Imatinib for CML. Trastuzumab for HER2-positive breast cancer. Ivacaftor for specific CFTR mutations in cystic fibrosis.
These successes share a structural property. The diseases they treat are driven by dominant molecular drivers in systems where removing or modulating that driver collapses the disease mechanism. In the language of my research, these are reducible systems. The biology converges toward a single dominant configuration. Hit that configuration and the system falls.
The problem is that these successes have been generalized into a universal strategy. The industry applies single-target logic to every disease system, regardless of whether the system is structurally reducible or not. And the failure rate tells us that the majority of disease systems are not reducible.
What Deterministic Convergence Reveals
Deterministic Convergence is a principle I identified through computational evolutionary simulation. The principle is straightforward: when you run deterministic simulations on genomic systems under identical input conditions, stable multi-variant interaction structures emerge reproducibly across independent runs.
The key word is deterministic. Identical inputs produce identical outputs. Every time. Not probabilistically. Not with high confidence. Identically. This is not a statistical model making predictions. This is a simulation engine revealing structure that is already present in the biological data.
What that simulation reveals is that biological systems organize their variant interactions into constraint architectures. And those architectures fall into two structurally distinct categories.
Reducible systems converge toward a dominant interaction configuration. The top-ranked variant or variant set accounts for the majority of the system's fitness. Cross-run concordance is high. The HHI (Herfindahl-Hirschman Index, a standard measure of concentration) exceeds 1500. These systems have a structural center of gravity. Hit it and the system collapses. Single-target strategies are structurally appropriate for reducible systems.
Distributed systems maintain multiple stable configurations simultaneously. No single variant or variant set dominates. The HHI falls below 1500. Cross-run concordance shows variation in which specific configurations rank highest, but convergence itself is preserved. The system converges to structure, but the structure is distributed rather than concentrated.
Across 26 independent studies spanning four clinical domains, 88.5% of observed genomic systems exhibit distributed architecture.
That number should concern anyone running a single-target drug discovery pipeline.
The Perturbation Evidence
The most instructive finding is not the classification itself. It is what happens when you perturb a distributed system.
In our cardiovascular validation case, we identified the dominant variant (rs11570112) and excluded it from the simulation. This models, computationally, what a perfectly effective single-target drug would accomplish: complete elimination of the primary target.
The system did not collapse.
Twenty-nine previously unranked variants rose to top-ranked positions. Fitness retention was 99.996%. The system reconverged, and it did so 50 generations faster than the baseline simulation.
Read that again. The system did not just survive the removal of its dominant variant. It reorganized more efficiently without it.
The dominant variant was not a structural anchor. It was a concentration artifact. The constraint architecture underneath was distributed the entire time. The variant that every target-selection pipeline in the industry would have identified as the lead candidate was, architecturally, dispensable.
The system reorganized through pathways that already existed in the constraint architecture. There was nothing to acquire. There was nothing to evolve. The alternative configurations were already present, already stable, already capable of maintaining system fitness. The selective pressure of target elimination simply revealed them.
This Is Not Resistance. This Is Reorganization.
The pharmaceutical industry calls this "acquired resistance." The language implies that the disease system develops something new in response to therapy. A mutation. An adaptation. An escape mechanism that did not previously exist.
The simulation data tells a different story. The reorganization pathways are not acquired. They are pre-existing. They are encoded in the constraint architecture of the disease system. They are discoverable computationally before a single patient is dosed, before a single dollar is committed to clinical development.
The distinction between "acquired resistance" and "pre-existing reorganization" is not semantic. It changes the entire therapeutic strategy.
If resistance is acquired, the rational response is to monitor for it, detect it early, and switch therapies when it emerges. This is the current standard of care.
If reorganization is pre-existing, the rational response is to characterize it before treatment, identify the specific pathways that will carry the redistributed load, and design the therapeutic strategy to address them from the outset. This is not the current standard of care. But it could be.
The Pre-Clinical Application
The application to drug discovery is direct.
Before committing to a target, run the perturbation analysis. Exclude the candidate target from the evolutionary simulation. Observe what happens to the system.
If the system collapses: the target is a structural anchor in a reducible system. Single-target development is structurally appropriate. Proceed with confidence.
If the system reorganizes: the target is a concentration artifact in a distributed system. Single-target development will face predetermined reorganization through alternative pathways. The simulation identifies which specific pathways carry the redistributed load. Those pathways are your combination targets. Those pathways are the resistance mechanisms your Phase II trial will discover three years from now and $400 million later, unless you characterize them computationally first.
This analysis takes 130 seconds. One hundred thirty seconds of computation that could redirect hundreds of millions of dollars in clinical investment.
The Validation Data
The claims above are supported by specific, reproducible performance metrics:
Predictive accuracy: AUC 0.9386 (95% CI: 0.91 to 0.97) against 1,000-trial empirical null distributions. Cross-domain validation AUC 0.959. PR-AUC 0.9417. Brier score 0.048.
Architectural classification: 88.5% distributed (HHI < 1500) across 26 studies in four clinical domains. 11.5% reducible (HHI > 1500).
Perturbation response: Dominant variant exclusion in cardiovascular systems produced 29 novel risers, 99.996% fitness retention, and convergence 50 generations faster than baseline.
Reproducibility: All computation is deterministic. Identical inputs produce identical outputs across independent runs. This is not a statistical property. It is an architectural guarantee.
This research has been submitted to Nature Biology for peer review.
Why Probabilistic AI Has Not Solved This Problem
The pharmaceutical industry has invested significantly in probabilistic AI for drug discovery. Machine learning models that identify targets. Deep learning models that predict binding affinity. Generative models that design novel molecules.
As of 2026, no FDA-approved drug has been developed through an AI-originated mechanism of action.
This is not because the AI is poorly engineered. It is because the AI is optimizing within a framework that contains a structural flaw. If the objective function assumes single-target sufficiency, the model will confidently optimize for a target the system can reorganize around. The model does not know biology. It knows patterns in the data it was trained on. If the data encodes a flawed structural assumption, the AI accelerates the flaw.
There is a deeper problem. Probabilistic models are, by definition, non-deterministic. The output changes when the model is retrained, when the training data is updated, when the feature weights shift. A target identified by an ML model on Monday may not be identified by the same model retrained on Friday.
This is incompatible with the requirements of drug development. A mechanism of action must be stable across the development timeline. A regulatory submission must be reproducible. A patent claim must be defensible. A clinical trial design must be based on a finding that does not change between protocol design and enrollment.
Deterministic systems do not have this problem. The output is the same every time because the constraint architecture is real. It does not change with the objective function, the training data, or the calendar.
The Combination Therapy Imperative
If 88.5% of disease systems are distributed, the therapeutic implication is clear: the industry needs combination strategies, not better single-target strategies.
Current combination therapy design is largely empirical. Combine drug A with drug B. Dose-escalate. Assess. Hope for synergy. The combinations are selected based on clinical experience, mechanism-of-action logic, and sequential trial-and-error.
Deterministic Convergence offers a rational alternative. The perturbation analysis identifies which pathways carry the redistributed load when the primary target is eliminated. Those pathways are not random. They are specific, identifiable, and reproducible. The combination design follows from the architecture.
The combination is not two drugs that individually showed activity. It is two drugs that together eliminate the system's ability to reorganize. The distinction matters clinically because the first approach produces additive benefit and the second produces architectural collapse.
The Neoantigen Problem
The same architectural principle applies to cancer vaccine development. Current neoantigen selection pipelines identify candidate antigens based on HLA binding affinity and expression-level filtering. Approximately 95% of predicted neoantigens fail to elicit meaningful immune response.
The 95% failure rate has the same structural cause as the 90% drug candidate failure rate. Single-antigen selection applied to a distributed system. A tumor that can reorganize its antigenic profile around a single neoantigen target will escape the vaccine through the same pre-existing reorganization pathways that enable drug resistance.
Multi-neoantigen designs informed by the constraint architecture, selecting the set of targets the tumor cannot lose simultaneously without collapsing its fitness, are the structural answer. The simulation identifies that set. The vaccine platform delivers it. The combination is rational, not empirical.
The Clinical Trial Stratification Problem
Distributed architecture also explains a significant portion of Phase III failure attributed to "heterogeneous response."
When a trial enrolls patients without characterizing their pharmacogenomic architecture, it mixes patients who will respond with patients who will not respond for genotype-specific reasons. A CYP2D6 ultra-rapid metabolizer and a CYP2D6 poor metabolizer receiving the same dose of the same drug are not receiving equivalent therapy. One is effectively underdosed. The other is at toxicity risk. The trial averages their outcomes and concludes the drug has "modest efficacy."
The drug does not have modest efficacy. It has excellent efficacy in a pharmacogenomic subgroup and unacceptable toxicity in another. The trial design could not see the difference because it did not stratify by the variable that determines response.
TRINITY, our multi-omic fusion engine, processes 1.3 million anchored clinical rules and 362+ FDA drug-gene interactions in under three minutes. Applied to trial enrollment, it stratifies patients by the pharmacogenomic architecture that determines whether they will respond, require dose adjustment, or should be excluded for safety. This is not post-hoc subgroup analysis. This is prospective stratification that increases trial signal, reduces heterogeneity, and identifies the responder population before randomization.
The Regulatory Defensibility Advantage
Every output from a deterministic system is reproducible, traceable, and cryptographically auditable. SHA-3-512 provenance on every analytical step. Every finding anchored to source evidence: PubMed PMIDs, ClinVar accessions, FDA guidance documents, CPIC guidelines.
For companion diagnostic development, this matters. The FDA requires validated, reproducible analytical methods. A CDx built on a probabilistic model that produces different outputs after retraining meets this requirement by assertion. A CDx built on deterministic infrastructure meets it by architecture. The difference will matter at the PMA stage.
For patent claims, it matters equally. A mechanism of action discovered by a deterministic simulation is reproducible on demand, documentable to specific computational steps, and defensible against prior art challenges. A mechanism discovered by an ML model that was subsequently retrained presents a more complex IP narrative.
The Path Forward
I am not arguing that single-target drugs are always wrong. 11.5% of observed systems are reducible. For those systems, single-target strategies are structurally appropriate. The successes of the past two decades, imatinib, trastuzumab, ivacaftor, are likely examples of drugs that hit reducible systems.
I am arguing that the industry has generalized the reducible-system playbook to all disease systems without first determining whether the system being targeted is reducible or distributed. That generalization explains the failure rate. And the failure rate is not a mystery to be solved with more data or better AI. It is a predictable consequence of applying the wrong structural assumption to the majority of disease biology.
The classification is now computationally possible. It takes 130 seconds. It costs a fraction of what the industry spends on a single Phase I study. It produces a determination, reducible or distributed, that fundamentally changes the development strategy.
For reducible systems: proceed with single-target development.
For distributed systems: characterize the reorganization pathways, design combination strategies, and stratify patients by pharmacogenomic architecture.
The tools exist. The infrastructure is live and in production. The research is published and submitted for peer review. The question is not whether the industry can do this. The question is how many more $2.6 billion failures it will tolerate before it does.
Matthew Hardy is the Founder and CEO of NomosLogic Inc., a deterministic molecular medicine infrastructure company. He is the primary inventor on 14 patent pillars filed with Wilson Sonsini Goodrich & Rosati and the author of Deterministic Convergence: Biological Systems, Architecture, and the Search for Hidden Order and The Adaptation Paradox: How Evolution's Gifts Became Medicine's Problems (Second Edition, 2026). He is a member of the American College of Medical Genetics and Genomics (ACMG) and the Biotechnology Innovation Organization (BIO).
