GSK: The AI-Driven Science Factory

Artificial Intelligence & Machine Learning , The Future of AI & Cybersecurity GSK: The AI-Driven Science Factory AI Rebuilds $50B Pharma Giant's Thinking, Plan Could Help Every Data-Driven Firm Rahul Neel Mani (@rneelmani) • May 18, 2026     Get Permission Source: Shutterstock Earlier this year, The Economist sent a correspondent to GSK's London base in King's Cross, an industrial rail district reborn as one of the city's liveliest neighborhoods. What the correspondent found there were no lab benches, no equipment, no white coats. The researchers were busy building a software tool trained to read genomes and generate hypotheses about disease from genetic data, entirely inside a computer. See Also: AI Impersonation Is the New Arms Race-Is Your Workforce Ready? For many, it may sound unexpected. But for a leading biopharma firm like GSK, it is part of the company's strategic direction. Over the past six years, GSK has quietly been building something far more consequential than merely deploying stray artificial intelligence tools. The company is on course to rebuild its entire research and development operating model around AI, from data pipes in the basement to agents writing scientific hypotheses. The results, though still early signals, are more than encouraging. Not a Data Engineering Company As McKinsey reported in 2019, GSK formally introduced machine learning as a key pillar of its corporate strategy. But AI isn't a departmental experiment or an experimental project. Kim Branson, senior vice president and global head of AI/ML for GSK, was recruited to turn that AI into an architectural asset. His first - and most counterintuitive - move was organizational. "People forget that a machine learning organization is not meant to do data engineering," Branson said. To address this, GSK created, Onyx, a suite of next-generation platforms to process and scale data and use it to discover new medicines. While most other enterprises either outsource this work or assume their ML teams can absorb it, GSK did neither. "Some companies have partnerships, but we needed to build this muscle internally," Branson told McKinsey. GSK's edge comes from treating operational data and model-training data as fundamentally different assets. Most large companies possess the former by default. The latter demands deliberate engineering, specialized teams and sustained investment. Onyx feeds a proprietary data ecosystem that integrates unpublished experimental results, clinical trial data, real-world population data and licensed external databases. It forms the foundation on which everything rests. "Data with real-world outcomes is gold. If you haven't got the outcome data from patients and can't build that linkage model, it's very difficult to calibrate," Branson said. The AI Scientist Writes Its Own Code Sitting atop the Onyx infrastructure is GSK's proprietary AI scientist, Cogito Forge. By equipping large language models and autonomous agents with scientific literature, datasets, analytical tools and computational resources, Cogito Forge enables its researchers to iterate more rapidly across the questions, analyses and decisions that underpin new medicine development. It autonomously formulates a strategy, retrieves data from internal and external sources, writes its own analytical code, synthesizes the findings, and produces a presentation with conclusions and testable hypotheses. Branson demonstrated it live. When prompted with a biology question, the system wrote code, gathered datasets, assembled them, generated charts and proposed a hypothesis before conducting a three-agent literature review. This is a classic example of a multi-agent AI doing science, with the human researcher retaining full oversight throughout. Cogito Forge is already in use across target discovery, causal mechanism analysis, biomarker identification and patient population assessment. Beyond usage, the scale is significant. What once required multiple teams of analysts working for weeks can now be initiated by a single researcher in minutes. The system can also explore hundreds of research directions in parallel, whereas manual workflows could manage only dozens. Research at Machine Speed Here's something even more fascinating. Cogito Forge doesn't work in isolation. Its research evidence layer is powered by Undermind, an AI system that goes beyond standard retrieval to exhaustively follow citation trails. It mirrors how a human researcher synthesizes what is known and how the evidence has evolved. More than 1,000 GSK scientists use it daily, both independently and as part of longer agentic workflows. GSK surveyed more than 130 of its scientists about the tool. Undermind achieved an NPS of 63, well above published benchmarks for enterprise software, placing it in the 95th percentile. The feedback was unambiguous. "Work, which would take weeks with an agency, took a few hours," a researcher reported. Another noted, "I can find something new that was never explored." Seeing Differently Through AI A parallel transformation is underway in the GSK lab. Its AI imaging platform integrates three custom-built tools, namely Neural Phenotypic Fingerprint, Image2Omics and Phenobind. This combined analysis of cellular responses identifies potential drugs at a scale and resolution previously unattainable. High-content imaging is at least 50 times more efficient than standard omics profiling. NPF addresses this with deep learning, detecting roughly four times as many phenotypes as conventional tools. Image2Omics goes even further. It predicts molecular data from imaging data, computationally replacing expensive laboratory analysis. In one colorectal cancer study, it predicted approximately 30% of the transcriptome from imaging alone and, across three experimental cycles, recaptured 98% of key findings while conducting only 30% of the usual laboratory experiments. The platform has reduced the need for omics data generation by at least threefold. Image2Omics pipeline: From raw cell images to the final output predictions for transcript and protein abundances Patents and the AI Bet None of this is purely an ambition. GSK, like any other pharmaceutical company, faces patent challenges as these rare, revenue-generating drugs lose exclusivity. At the 2026 JP Morgan Healthcare conference, the company's chief scientific officer, Tony Wood, outlined how AI-driven R&D sits at the center of its commercial strategy, not as a long-range bet, but as a near-term efficiency multiplier with specific targets. In a single day in January 2026, GSK signed two AI-focused deals, including a multi-year partnership with genomics firm Helix for access to its GenoSphere cohort data, and a five-year, $50 million agreement with AI-native biotech Noetik for virtual cell models in non-small cell lung cancer and colorectal cancer. Wood minced no words in calling out Phase II trial attrition as the exact problem AI must solve. The broader market is moving in line with projections. According to BCC Research, annual investment in AI-driven drug discovery is projected to grow from $3.8 billion in 2025 to $15.2 billion by 2030. AI-designed drugs are now clearing preclinical phases in 12 to 18 months, down from the historical three to five years, with extremely high success rates in safety trials. "The explosion of collecting data at scale is going to be the most interesting thing. It will be phenomenal to witness the integration of information from disciplines we have never combined before," Branson said. Lessons for CIOs What makes GSK's program stand out among other pharmaceutical companies is its architecture. It is a complete taxonomy of what structural enterprise AI requires, and the three principles run through every layer of GSK's AI strategy. GSK has put data infrastructure before models. It built Onyx before scaling AI. Most enterprises tend to skip this and blame the models for underperformance. GSK paid enough attention to generate data with intent. The most noticeable comment made by Branson is "data collected to run operations is different from data designed to train models." Finally, GSK knew governance can be a spoiler, so "govern from day one." Its AI Governance Council sets enterprisewide standards, monitors U.S. and EU regulatory shifts, and runs risk-based assessments on every deployed AI system, treating governance not as mere compliance but as a strategic asset in an environment of diverging global regulation. The lab in King's Cross has something more enduring than usual: An organization that has done the unglamorous work of embedding AI into its foundations and is slowly reaping the returns.