AI Will Turn Pharma into an Evidence Industry

When plausible drug candidates become easier to generate, the scarce asset will be the ability to prove which ones deserve to survive.

For the first phase of artificial intelligence in pharmaceutical R&D, the industry asked how quickly algorithms could identify targets, design molecules and optimize compounds.

The next question is more consequential:

What happens when AI expands the supply of plausible drug candidates faster than the industry can validate them in humans?

The answer may reshape much more than drug discovery. It could change how pharmaceutical companies structure R&D, allocate capital, manage pipelines, evaluate acquisitions and define competitive advantage.

AI will not transform Pharma into a software industry.

It may transform Pharma into an evidence industry.

AI Is Making Drug Candidates More Abundant — Not Medicines

A computationally promising molecule is not yet a medicine.

It may bind to its intended target. It may appear selective, synthesizable and pharmacologically attractive. It may perform well in cell-based assays and animal models.

None of this proves that the underlying biological mechanism is causal in human disease, that the molecule will produce a clinically meaningful effect or that its benefits will outweigh its risks.

Clinical attrition remains the defining economic reality of drug development.

A recent analysis covering more than 20,000 clinical development programs found substantial variation in reported clinical success rates, generally ranging from approximately 7% to 20%, depending on the methodology, therapeutic area, modality and period analyzed. (Nature Communications)

The first randomized Phase IIa trial of a drug and target combination discovered using generative AI represents an important milestone. Rentosertib showed acceptable short-term safety and a signal of potential benefit in idiopathic pulmonary fibrosis.

However, the study included only 71 participants, lasted 12 weeks and explicitly concluded that larger and longer trials were required. Its authors also noted that very few AI-discovered drugs had reached clinical trials, that Phase II failures remained comparable with conventional programs and that no AI-discovered drug had yet completed Phase III development. (Nature Medicine)

AI can substantially expand the universe of testable hypotheses.

It has not eliminated the biological uncertainty separating a promising hypothesis from a successful medicine.

The abundance of candidates does not create an abundance of clinical truth.

Human Evidence Is Becoming the New R&D Bottleneck

When molecular generation becomes faster, the strategic bottleneck moves downstream.

The scarce assets become:

• causal evidence that a target is relevant to human disease;
• translational models that reliably predict human response;
• biological samples and longitudinal patient data;
• biomarkers that identify who may benefit;
• patients who meet increasingly precise eligibility criteria;
• clinical sites capable of executing complex protocols;
• endpoints that demonstrate meaningful benefit;
• regulatory-grade evidence that can withstand independent review.

This is why human genetics has become strategically important.

A large study published in Nature estimated that drug mechanisms supported by human genetic evidence had a probability of clinical success approximately 2.6 times greater than mechanisms without such support. The association became stronger when confidence in the causal gene was higher. (Nature)

The implication is not that genetics can eliminate uncertainty.

It is that the quality of the biological evidence entering the pipeline can materially influence the probability of what eventually emerges from it.

Companies such as GSK are already presenting human genetics, functional genomics, artificial intelligence and machine learning as interconnected components of their R&D strategy rather than separate technologies.

This approach reflects an important strategic shift: using technology not merely to generate more molecules, but to improve the understanding of patients, disease mechanisms and target biology before committing substantial capital to development. (GSK Annual Report 2025)

The more efficiently AI generates drug candidates, the more valuable high-quality human evidence becomes.

Pipeline Obesity: When More Drug Candidates Create More Risk

For decades, an empty or insufficient pipeline was treated as one of the pharmaceutical industry’s greatest strategic threats.

AI could create the opposite problem: more targets, compounds and programs than an organization can responsibly finance, validate and develop.

This is pipeline obesity.

It may appear positive in investor presentations. More targets are identified. More candidates enter preclinical development. More assets can be described as AI-enabled.

But an expanding pipeline can conceal deteriorating capital discipline.

The risks include:

• multiple programs directed at the same biologically fashionable targets;
• correlated errors because competing models rely on similar public datasets;
• false positives that survive longer because their computational rationale appears sophisticated;
• internal competition for laboratory, clinical and manufacturing resources;
• emotional or political attachment to programs;
• greater pressure to move weak assets into expensive clinical phases;
• increasing competition among sponsors for the same investigators and patient populations.

AI can industrialize good hypothesis generation.

It can also industrialize convincing mistakes.

A model trained on historical evidence may reproduce the same weaknesses that already exist in biomedical research: publication bias, incomplete negative results, underrepresentation, overreliance on animal models and concentration around well-studied targets.

Different companies may produce apparently independent conclusions while relying on substantially similar data, assumptions and scientific literature.

An organization that measures innovation primarily by the number of programs entering the pipeline may therefore become less productive as its computational capabilities improve.

A larger pipeline is not necessarily a healthier pipeline.

AI Portfolio Governance and the Rise of the Kill Engine

If candidates become more abundant, the ability to create them becomes less differentiating.

The ability to terminate the wrong programs becomes more valuable.

The pharmaceutical company of the future will need not only a discovery engine, but a kill engine: a disciplined system for identifying weak hypotheses before they become expensive clinical failures.

Such a system would require:

• pre-established evidence thresholds;
• experiments designed to discriminate between competing hypotheses;
• explicit criteria for target and candidate validation;
• continuous probability updates;
• independent scientific challenge;
• comparison of programs across therapeutic areas;
• incentives that reward high-quality termination decisions;
• systematic learning from failed programs.

This is not simply an AI application. It is a governance challenge.

AI can integrate evidence, identify inconsistencies and estimate probabilities. It can surface external data that contradict internal assumptions. It can model alternative indications, patient populations or therapeutic combinations.

But the decision to terminate a program also involves judgment, accountability, organizational incentives and a willingness to accept sunk costs.

One of the greatest risks is that AI becomes another instrument used to defend programs rather than challenge them.

In an evidence industry, the purpose of AI should not be to make every asset appear more credible.

It should be to make uncertainty more visible.

When candidates are abundant, stopping weak programs becomes as important as creating new ones.

From Drug Pipelines to AI-Enabled Evidence Engines

Traditional pharmaceutical R&D is organized as a sequence of gates.

A target is selected. A molecule is developed. Preclinical studies are conducted. The program passes to clinical development. Regulatory and commercial teams become progressively involved.

Information frequently moves forward, but learning does not always travel back.

A failed clinical trial may reveal that a target was relevant only in a specific biological subgroup. A manufacturing limitation may emerge after candidate selection. A post-market safety signal may identify a mechanism that should have influenced an earlier discovery decision.

The next R&D model will need to convert those events into reusable institutional knowledge.

McKinsey has described this transition as a move from linear stage gates toward connected learning loops linking disease biology, target validation, candidate design, clinical trials and patient impact. In this model, each decision generates data that improves both subsequent and earlier decisions. (McKinsey)

An effective evidence engine would connect seven layers.

Disease Evidence

What drives disease progression, heterogeneity and treatment response?

Target Evidence

Is the target causal, modifiable and relevant in humans?

Candidate Evidence

Does the asset offer sufficient efficacy, safety, manufacturability and differentiation?

Clinical Evidence

Which patients should be treated, at what dose and against which comparator?

Regulatory Evidence

Are the data, models and analyses reproducible, traceable and appropriate for their intended use?

Economic Evidence

Does the treatment provide sufficient incremental value to support access and reimbursement?

Real-World Evidence

Does the benefit persist in broader populations and routine clinical practice?

In this architecture, the pipeline is not merely a list of assets.

It is a continuously updated portfolio of claims, uncertainties and supporting evidence.

How AI Will Change Pharmaceutical M&A

Pharmaceutical M&A will continue to be driven by familiar pressures:

• patent expirations;
• revenue replacement;
• access to new therapeutic modalities;
• entry into strategic disease areas;
• acquisition of clinical-stage assets;
• portfolio consolidation.

Those pressures are intensifying.

Reuters reported that biopharma transactions totaled approximately $84 billion in the first quarter of 2026, compared with $44.4 billion one year earlier. If that pace continued, total 2026 deal value could exceed $250 billion. (Reuters)

AI will not replace these drivers.

It will change what buyers value and how they manage uncertainty.

From AI Platform Promises to Validated Performance

During the first wave of AI investment, companies could attract capital based on:

• proprietary algorithms;
• the size of virtual libraries;
• computational throughput;
• the speed of target or molecule generation;
• access to specialized AI talent.

The next phase will demand harder questions:

• How many predictions were experimentally validated?
• How many candidates progressed into humans?
• What was the false-positive rate?
• Did the platform materially change a development decision?
• Are the data proprietary and legally reusable?
• Can the results be reproduced independently?
• Does the platform offer a durable advantage over increasingly accessible models?

A company that produces many candidates but little high-quality evidence may be less valuable than a company that generates fewer assets with substantially stronger biological validation.

More Milestone-Based and Option-Based Transactions

The structure of recent AI agreements already reflects this uncertainty.

Lilly’s expanded agreement with Insilico Medicine was announced with a potential total value of approximately $2.75 billion, but only $115 million was committed upfront. Most of the potential consideration depends on development, regulatory and commercial milestones. (Reuters)

Alnylam’s collaboration with Inceptive was valued at up to $2 billion but included upfront consideration of $30 million, with additional payments tied to preclinical, regulatory and commercial achievements.

The agreement combines Inceptive’s models with more than two decades of proprietary RNA-interference data from Alnylam. (Alnylam Investor Relations)

These structures illustrate a likely direction:

• smaller upfront payments;
• technical-validation milestones;
• rights that expand as evidence develops;
• options by indication or geography;
• acquisition rights triggered by proof of concept;
• royalties and milestones that transfer part of the scientific risk to the platform developer.

The headline value of a deal may become increasingly disconnected from the amount of capital placed at risk before human evidence exists.

Proprietary Data Becomes Part of the Acquired Asset

The Alnylam–Inceptive collaboration is significant because the strategic asset is not merely an algorithm.

It is the combination of:

• a validated therapeutic modality;
• proprietary experimental data;
• domain expertise;
• foundation models;
• development capabilities.

GSK’s collaboration with Relation Therapeutics follows a similar logic. Relation is expected to generate human cellular datasets and use them to develop AI models for target identification.

The potential value lies in the interaction between proprietary human data, experimental biology and computation. (Reuters)

The M&A target of the future may therefore be difficult to describe as simply a biotech company, data company or software company.

It may be an integrated evidence platform.

AI Creates a New Layer of M&A Due Diligence

AI-related transactions require diligence beyond conventional pipeline evaluation.

Buyers must examine:

• data provenance and consent;
• intellectual-property rights over source data and outputs;
• model validation and version control;
• dependence on third-party foundation models;
• reproducibility;
• cybersecurity;
• bias and population representation;
• regulatory acceptability;
• performance degradation as data or standards change;
• whether historical outputs can be independently reconstructed.

A sophisticated model built on unusable data may have little strategic value.

Impressive retrospective performance may disappear when a model is deployed in another population, laboratory or clinical workflow.

New AI Opportunities in Pharmaceutical R&D

The greatest opportunities may not be concentrated in molecule generation.

They are likely to emerge where the industry still faces scarcity.

Human Data and Translational Biology

Companies able to connect genomics, proteomics, imaging, pathology, electronic health records, treatment exposure and longitudinal outcomes will be better positioned to answer questions that public datasets cannot.

The advantage will come not from data volume alone, but from the ability to link biological characteristics to meaningful human outcomes.

Biomarkers and Companion Diagnostics

As pipelines become more crowded, identifying the correct patient may create more value than marginally improving the molecule.

Predictive biomarkers can:

• increase the observed treatment effect;
• reduce unnecessary patient exposure;
• improve clinical trial design;
• support differentiated labeling;
• provide evidence for reimbursement;
• guide treatment combinations and sequencing.

Automated Laboratories and Closed-Loop Experimentation

The most defensible AI platforms may be those directly connected to laboratories.

A closed loop in which models design experiments, robotic systems execute them and results update the next decision can create a cumulative data advantage.

The algorithm may eventually be replicated.

The integrated experimental-learning system is much harder to reproduce.

AI-Enabled Pharmaceutical Portfolio Intelligence

Abundant candidates increase demand for systems that integrate:

• biological probability;
• clinical feasibility;
• competitive intensity;
• manufacturing risk;
• intellectual property;
• regulatory strategy;
• patient access;
• pricing and reimbursement;
• risk-adjusted economic value.

The purpose will not be to predict winners with certainty.

It will be to make portfolio decisions more consistent, transparent and responsive to new evidence.

Regulatory-Grade AI Validation and Assurance

Any model used to support consequential scientific or regulatory decisions will need documented credibility.

This creates opportunities for companies specializing in:

• validation;
• auditability;
• model monitoring;
• data governance;
• independent performance assessment;
• regulated software lifecycle management;
• reproducibility.

Pharma AI Investment and Capital Allocation Are Already Changing

The industry is moving beyond isolated pilots.

Lilly launched TuneLab to provide selected biotechnology companies with access to AI models developed from more than $1 billion in research investment and decades of proprietary data. (Eli Lilly Investor Relations)

Lilly and Nvidia subsequently announced a co-innovation laboratory with planned investment of up to $1 billion over five years, combining computing infrastructure, robotics, scientific expertise and startup participation. (Eli Lilly Investor Relations)

Roche announced an on-premises AI infrastructure with 2,176 GPUs across the United States and Europe, intended to support therapeutics and diagnostics throughout its value chain. (Roche)

These investments do not prove that AI will improve clinical success rates.

They demonstrate that leading companies are treating AI, proprietary data and integrated experimentation as strategic infrastructure rather than peripheral technology.

Novartis is also telling investors that pipeline quality, disciplined execution and targeted acquisitions are central to its strategy.

Its 2025 annual report described more than 30 potential high-value medicines, 15 submission-enabling readouts expected over two years and a series of targeted bolt-on transactions intended to strengthen specific therapeutic and technology capabilities. (Novartis Annual Report 2025)

The common theme is not simply “more AI.”

It is the construction of systems that connect scientific evidence, capital allocation and development decisions.

Which Pharma and Biotech Companies Are Best Positioned?

Potential winners include:

• pharmaceutical companies with integrated, reusable human data;
• organizations with disciplined portfolio governance;
• companies that connect computational models to experimental validation;
• biomarker and diagnostic businesses;
• platforms capable of demonstrating clinical and regulatory value;
• organizations that learn systematically from failed programs.

Potential losers include:

• undifferentiated molecule-generation platforms;
• companies dependent primarily on public datasets;
• organizations that measure innovation by pipeline size;
• biotechs unable to finance validation;
• companies whose models cannot be independently reproduced;
• pharmaceutical organizations that implement AI without redesigning incentives and decision rights.

Pharma Will Compete on Evidence

AI may make it possible to generate more targets, compounds and development options than ever before.

That does not solve the central problem of pharmaceutical R&D.

It changes it.

The competitive question will no longer be only:

Can we generate a promising candidate?

It will increasingly become:

Can we produce enough high-quality evidence to know whether this candidate should survive?

The pharmaceutical company of the future will not be defined by the number of hypotheses it can create.

It will be defined by the speed and reliability with which it can convert uncertainty into evidence — and evidence into better decisions.

AI will not make successful medicines abundant. It will make the ability to create trustworthy evidence more valuable.

Sources and Further Reading

• Minikel et al., Refining the Impact of Genetic Evidence on Clinical Success, Nature, 2024. Read the study
• Zhou et al., Dynamic Clinical Trial Success Rates for Drugs in the 21st Century, Nature Communications, 2025. Read the study
• Xu et al., A Generative AI-Discovered TNIK Inhibitor for Idiopathic Pulmonary Fibrosis, Nature Medicine, 2025. Read the study
• IQVIA Institute, Global R&D Trends 2026. Read the report
• McKinsey, From Linear Gates to Learning Loops: Rewiring Biopharma R&D with AI. Read the analysis
• Novartis, Annual Report 2025. Read the report
• GSK, Annual Report 2025. Read the report
• Reuters, Big Pharma M&A Set for Mega Year as Patent Expiries Drive Deal Urgency. Read the report
• Alnylam, Strategic AI Collaboration with Inceptive. Read the announcement
• Reuters, Eli Lilly Extends Partnership with Insilico Medicine. Read the report
• Eli Lilly, TuneLab Platform Announcement. Read the announcement
• Eli Lilly and Nvidia, Co-Innovation AI Laboratory. Read the announcement
• Roche, AI Infrastructure for Therapeutics and Diagnostics. Read the announcement

---

Originally published on LinkedIn on August 9, 2026.