Rising costs and stagnating productivity define the drug development problem that Eroom’s Law describes. The inflation-adjusted cost of bringing a new medicine to market has roughly doubled every 9 years since the 1980s, despite advances in screening and molecular biology. The result is a strategic imperative to change how targets are found, molecules are designed, and evidence is generated. AI in drug discovery is now a practical toolkit rather than a slogan. Generative models propose novel chemotypes and protein sequences. Predictive models estimate absorption, distribution, metabolism, excretion and toxicity. Natural language processing extracts signals from unstructured literature, patents and health records. This feature assesses the state of play in late 2025, using peer-reviewed studies, regulatory documents and company disclosures to evaluate what is working, where the limits sit and how organisations are restructuring to integrate these tools.
Real Fact: The FDA reports a marked rise in submissions that use AI across the drug product life cycle, reflecting growing use of AI in nonclinical, clinical, postmarketing, and manufacturing phases.
The new R&D toolkit decoding AI in the pharmaceutical context
AI modalities in discovery fall into three broad classes with distinct points of application.
Generative AI. Diffusion models, graph neural networks, and reinforcement learning generate molecules optimised for target binding, selectivity and developability, and design antibodies or other biologics with specified constraints. Case evidence now includes a generative AI that discovered a small molecule inhibitor of TNIK for idiopathic pulmonary fibrosis that achieved randomised phase 2a results in 2025, supporting the feasibility of data-driven design through to human proof of mechanism.
Predictive analytics and machine learning. Supervised and self-supervised models predict activity, physicochemical properties and ADMET liabilities, prioritise analogues and guide synthesis. Platform companies report cycle time reductions from hit finding to IND-enabling studies, and external analyses describe faster progression of candidates into the clinic. However, definitive industry-wide baselines remain under review.
Natural language processing. Modern NLP and knowledge graphs integrate signals across publications, patents and omics repositories to identify or validate targets and to propose repurposing candidates. The often cited example is the identification of baricitinib for COVID-19 through AI-assisted literature mining and reasoning, followed by prospective clinical evaluation.
Regulatory posture. Agencies have moved from observation to structured guidance. The FDA’s 2025 draft guidance sets out a credibility framework for AI models used to support regulatory decision-making in drug and biologic submissions. The EMA’s final reflection paper describes a risk-based approach for AI use across the medicinal product lifecycle. In the UK, NICE has set positions on AI in evidence generation for technology evaluations. Meanwhile, the MHRA continues to advance its Software and AI as a Medical Device change programme and AI Airlock sandbox. Although device-focused, these initiatives shape expectations for validation, human oversight, and post-market learning in adjacent use cases.
Key takeaway: machine learning pharma R&D now spans generative design, predictive modelling and NLP, with regulators emphasising risk proportionate validation and traceability rather than novelty.
Case studies: AI discovered molecules in the clinic.
Insilico Medicine’s TNIK inhibitor for pulmonary fibrosis. Rentosertib (formerly INS018_055, ISM001-055) progressed from design to a double blind randomised phase 2a trial in idiopathic pulmonary fibrosis, with results published in Nature Medicine in August 2025. The study reported acceptable safety and pharmacokinetics at 12 weeks, with secondary endpoints including lung function and patient-reported outcomes. This programme is frequently cited as a generative pipeline that moved beyond target ID to human testing in a fibrotic disease with limited options.
Timelines reported by the developer indicate an accelerated path from algorithmic hypothesis to first-in-human, followed by a controlled study. Earlier communications in 2024 noted positive topline findings from a phase 2a cohort. Registration entries corroborate the progression of the asset through clinical stages. While independent meta-analyses of overall timeline compression remain scarce, this is one of the clearest demonstrations that an AI-generated small molecule can achieve mid-stage evaluation.
Recursion’s clinical portfolio. Recursion reports multiple clinical programmes discovered or advanced using its Recursion OS platform, including oncology assets such as a CDK7 inhibitor in phase 1/2. Coverage in 2024 and 2025 highlights both progress and mixed readouts, underlining that AI enables but does not circumvent clinical risk. Speed claims include moving a candidate to the clinic in about 18 months in one programme, compared to historical figures of about 42 months. However, independent verification is still limited, and outcomes vary by indication.
Exscientia collaborations and near-term entries. Exscientia announced milestone progress with Sanofi programmes and planned 2025 initiations for AI-designed small molecules targeting LSD1 and MALT1. As of late 2025, Exscientia’s focus remains on moving these designs from IND or CTA enabling into combined phase 1/2 studies. This illustrates a different integration model where a platform company codesigns assets with a large partner to scale medicinal chemistry decisions and biomarker strategy.
Context. Journal and media reviews agree that no AI-designed drug has yet achieved regulatory approval, and that phase 3-ready assets remain uncommon. The evidence base is strongest at the preclinical and early clinical stages, with isolated phase 2a results beginning to appear.
Quantifying impact from targets to candidates
Time and success rates. A 2024 analysis of AI-discovered molecules reported phase 1 success rates of 80-90%, which are substantially higher than historic averages, while phase 2 success was around 40% on limited sample sizes. This suggests more selective early pipelines, while acknowledging that attrition later in development remains challenging. Industry reporting and consultancy work also indicate faster cycle times for certain document-heavy or analytics steps, such as clinical study report drafting or trial design iteration. Estimated reductions of 40% are projected for selected workflows, with significant gains anticipated for agentic AI in clinical development operations. These improvements are task-specific rather than blanket guarantees.
Cost and resource use. News and market analyses describe expectations of 35-45% gains in clinical productivity when AI agents support protocol design, site selection, and monitoring. They also present case examples where discovery programmes moved from hit to clinic faster than traditional averages. These data points are promising but heterogeneous, and they require confirmation across therapeutic areas and company sizes.
Bottom line: predictive models in pharmacology and generative AI drug design can materially cut specific steps. Whole programme timelines still hinge on experimental validation, manufacturing and clinical effect sizes.
The big pharma pivot from adoption to integration
Partnerships to pipelines. AstraZeneca’s multi-year collaboration with BenevolentAI has produced multiple novel targets selected into discovery portfolios across chronic kidney disease, idiopathic pulmonary fibrosis, heart failure and systemic lupus erythematosus. In October 2025, AstraZeneca also signed a milestone-based AI and CRISPR discovery deal with Algen Biotechnologies in the field of immunology. These announcements demonstrate continuing appetite for AI-enabled target discovery, even as some partners restructure.
Restructuring and capability building. Companies such as GSK and Novartis describe internal teams hiring machine learning engineers, computational biologists and data scientists to link in silico and wet lab cycles. Job postings and corporate communications indicate targeted recruitment into phenomics, human genetics and translational modelling, consistent with a move from exploratory pilots to embedded workflows. Computational drug discovery is now visible in organisational charts and career pathways, not just in vendor decks.
Mergers and alliances. Sanofi’s collaboration with Exscientia covers up to 15 small molecules across oncology and immunology, with milestone payments reported in 2024. Novartis partnered with Generate: Biomedicines to apply generative AI to protein therapeutics. These alliances spread risk and accelerate access to tooling, while raising the bar for prospective evidence that the platforms deliver better starting points.
The pipeline view where AI makes a measurable difference
Target identification and validation. Knowledge graph-based NLP-supported identification of baricitinib for COVID-19 continues to inform hypothesis generation in inflammatory and fibrotic disease partnerships. The operational lesson is that literature scale synthesis can propose tractable mechanisms quickly, but wet lab validation and clinical trials remain the arbiters of truth.
Lead design and optimisation. Generative models reduce the number of synthetic iterations by proposing series with favourable multi-parameter profiles. The Nature Medicine TNIK study provides the most visible human data point that a de novo small molecule discovered with generative AI can reach a randomised study. External commentaries caution that this is one success in a small field and should not be overgeneralised.
Translational modelling and trial operations. Regulators and consultancies describe workflow acceleration from protocol drafting to clinical study reports and data review. Gains are most significant in information processing tasks where models can draft, summarise, or cross-check content that humans then verify. The practical benefit is fewer cycles and more precise documentation, not an automatic shortcut through dose finding or endpoint selection. The data problem is garbage in, garbage out
Data quality and bias. AI performance depends on clean, representative datasets with clear provenance. The EMA reflection paper and NICE’s position on AI in evidence generation emphasise transparency, reproducibility and human oversight. FAIR principles are necessary for interoperability across discovery and development systems. Without curation, models will learn literature biases or overfit to narrow chemistries, inflating early metrics that do not translate.
Regulatory expectations. The FDA’s 2025 draft guidance proposes a risk-based credibility framework for AI used in submissions, focusing on a defined context of use, data lineage, model verification and performance monitoring. In the UK, the MHRA’s AI Airlock and its SaMD change programme push suppliers to articulate safety cases and post-market learning plans. These frameworks demand explicit model documentation and guardrails rather than black box outputs.
Security and IP. As companies link automated labs to cloud platforms, data governance and cybersecurity become core to value protection. Agency guidance on good machine learning practice and manufacturers’ obligations under device regulations, while device-centric, sets expectations for lifecycle monitoring and change control that discovery teams increasingly mirror.
Balanced evidence from independent analyses shows
Early clinical signals. The 2024 review of AI discovered pipelines reported high phase 1 success rates and phase 2 performance comparable to historical averages on small samples. Independent coverage in 2025 continues to state that no AI-designed drug has yet been approved, and that late-stage trials remain the decisive filter. AI-driven clinical trials may become faster and more precise, but attrition will persist when biology is weak or heterogeneity is high.
Operational gains. McKinsey analyses show 40% reductions in specific clinical document timelines and project significant productivity gains as agentic systems mature. News reporting links these gains to reduced reliance on animal testing and increased use of computational and human cell models, although complete replacement is not imminent. Press releases and vendor blogs can overstate generalisability. Peer-reviewed results, regulatory filings, and independent audits carry more weight. Corporate restructurings in the sector, including layoffs and programme pivots, reflect capital discipline and the difficulty of sustaining broad pipelines until clinical signals emerge. Stakeholders should demand pre-registered study plans, blinded validation and auditable data flows.
Integration into pharma workflows, people, platforms and wet labs
New roles and interfaces. R&D functions now include machine learning engineers embedded with medicinal chemists, computational biologists working alongside in vivo teams, and data product managers to standardise experiment capture. Hiring patterns at large companies confirm these shifts. Collaboration is practical rather than ceremonial when assay data, design proposals and synthesis queues live in the same system and decisions are logged.
Automation and feedback loops. End-to-end platforms aim to shorten design-make-test-analyse cycles. Robotics integrated with model-based design can increase the number of hypothesis cycles per week, but throughput without hypothesis quality wastes materials. Several companies now frame their value not as isolated algorithms but as closed-loop systems that get better with each experiment.
Governance. Companies are building internal model registries, validation playbooks and change logs to satisfy quality and regulatory needs. The direction of travel is towards traceability: what data trained a model, which version made a given decision, and how performance drift is detected.
Where is the following practical guidance for teams
Focus on questions, not tools. Start with a bottleneck that impedes a programme: target tractability, off-target risks, or human evidence generation. Select methods that address the question with measurable endpoints.
Measure like a scientist. For each AI-assisted step, define baselines and counterfactuals, track cycle time per design-make-test iteration, hit rates and false positives. Publish enough detail that peers can understand the denominator, not just the numerator.
Treat models as hypotheses. Predictions guide experiments. They do not replace them. Invest in assay quality and human oversight to improve the feedback loop in the next round.
Align with regulators early. Use the FDA credibility framework and EMA reflections to plan documentation and performance monitoring. For UK programmes, anticipate MHRA expectations on transparency and NICE needs when AI-supported evidence appears in submissions.
Build durable teams. Hire people who understand both chemistry and biology, and statistics and software engineering. Create shared incentives so wet and dry labs own the exact deliverables.
Conclusion: a new era of rational design
Eroom’s Law is not immutable. The past two years have delivered concrete progress from AI-assisted pipelines into early clinical testing, with a notable phase 2a trial of a generative AI-designed small molecule in pulmonary fibrosis and multiple platform companies advancing first-in-human assets. At the same time, late-stage attrition remains the barrier to therapeutic and commercial impact. The most credible path is a hybrid model where AI in drug discovery accelerates selection and optimisation, predictive models in pharmacology de-risk ADMET, and NLP supports evidence generation. At the same time, laboratories and clinics retain the final say.
For organisations in 2025, the strategic choice is clear. Integrate AI tools where they measurably compress cycle times or sharpen decision quality, and document models to regulatory standards. Invest in data quality and people who can bridge computation and experiment. The signal from the best case studies is that timelines can be compressed and early success rates can improve. The caution from independent reviews is that approvals will depend on biology, trial design and reproducible effect sizes. The industry is entering a data-driven, rational design phase that rewards disciplined use of generative AI drug design and rigorous validation over slogans.
