The (10⁶⁰) Paradox: Is Big Pharma’s AI Just Automating Risk Aversion?
Artificial intelligence was supposed to transform drug discovery. Yet the industry's most successful AI collaborations may tell us as much about the limits of historical data as the power of machine learning. Faced with an almost incomprehensibly vast chemical universe, AI promised to help scientists move beyond traditional trial-and-error methods and uncover entirely new therapeutic possibilities. Theoretical estimates suggest that the number of drug-like organic molecules with molecular weights below 500 Da may approach 10⁶⁰, a figure so large that only an infinitesimal fraction could ever be synthesized and tested experimentally.[1]
Against this backdrop, the pharmaceutical industry's flagship AI collaboration, MELLODDY (Machine Learning Ledger Orchestration for Drug Discovery, specifically federated machine learning for QSAR prediction), represented a landmark achievement. Rather than attempting to explore the full breadth of chemical space, ten global pharmaceutical companies joined forces to train machine-learning models on their private experimental data without sharing the underlying datasets themselves.[2]
The result was one of the largest federated learning initiatives ever deployed in biopharmaceutical research. More than 2.6 billion experimental activity data points, spanning over 21 million physical small molecules and 40,000 assays, were incorporated into a privacy-preserving machine-learning framework.[2] The project demonstrated something many considered impossible only a few years earlier: competing pharmaceutical companies could collectively improve predictive models without exposing their proprietary scientific assets.
Yet beneath the technical success lies a more provocative question. Having spent much of my career working at the intersection of pharmaceutical science, biomarkers and data platforms, I've learned that the quality of a prediction often reflects the assumptions embedded within the data that produced it. If AI systems are trained primarily on historical pharmaceutical datasets, are they truly expanding the boundaries of drug discovery, or are they simply codifying decades of industry assumptions about which molecules are worth pursuing?
The SME Mirage: Who Really Built the Models?
One common criticism of industry-led AI is that it risks excluding smaller, more experimental organizations whose willingness to challenge conventional thinking often drives breakthrough innovations.
MELLODDY was not entirely a closed ecosystem. Alongside ten pharmaceutical companies, the consortium included four SMEs, two academic institutions, NVIDIA, and specialist technology contributors such as Owkin, Iktos, Kubermatic, and the Substra Foundation.[3]
This diversity matters. Smaller technology companies played critical roles in designing the infrastructure, machine-learning workflows, and federated learning framework that made the project possible.
However, technical participation is not the same as data participation.
The primary experimental evidence came from the pharmaceutical companies themselves, supplemented by public datasets.[2] Consequently, while the computational ecosystem was diverse, the underlying experimental record remained largely shaped by the historical research priorities of large pharmaceutical organisations.
This distinction does not prove that MELLODDY's models are biased against unconventional chemistry. However, it raises an important question: what happens when an AI's understanding of "drug-like" molecules is built predominantly upon compounds that established pharmaceutical research programmes previously chose to synthesise, screen and optimise?
The Anatomy of Big Pharma Bias
Whenever AI learns from historical data, it inherits characteristics of the decisions that generated those data. In drug discovery, several potential limitations emerge.
The Synthesis Selection Bias
Pharmaceutical datasets largely reflect compounds that researchers chose to synthesise, purchase or test. Those decisions are shaped by practical realities including cost, synthetic accessibility, assay compatibility and commercial priorities.
Compounds that were never made simply cannot provide experimental evidence.
The consequence is not that AI learns difficult-to-manufacture molecules are inherently bad. Rather, predictive confidence may decline when evaluating molecules that differ substantially from those represented in the training data. This aligns with a long-recognised challenge in QSAR modelling: predictions become less reliable once compounds fall outside a model's applicability domain.[4]
Put simply, AI can only confidently extrapolate so far beyond experience.
The Filter-First Mentality
Drug development operates through layers of screening, prioritisation and elimination. Over decades, researchers have continuously selected certain compounds for additional testing while abandoning others.
As a result, historical datasets capture not only molecular properties but also the decisions that determined which experiments happened at all.
Because MELLODDY was built to learn from existing experimental evidence, it naturally inherits traces of those decisions. A genuinely unconventional molecule may lie outside the domain in which the model can generate reliable predictions, not because it is intrinsically inferior, but because there is insufficient comparable evidence available for evaluation.
Whether federated models can reliably generalise into unexplored chemical territory remains an empirical question rather than a settled assumption.
The Small-Molecule Silo
MELLODDY focused specifically on datasets involving physical small molecules.[2]
This is an important qualification because drug discovery is increasingly expanding beyond traditional small-molecule chemistry. Emerging therapeutic modalities include:
Targeted protein degraders (PROTACs)
Macrocycles
RNA therapeutics
Cell and gene therapies
Novel biologic platforms
Some overlap undoubtedly exists between these areas and conventional medicinal chemistry. Nevertheless, MELLODDY's reported successes do not automatically establish predictive capability across every future therapeutic modality.
Extending AI into these areas requires appropriate data, modelling strategies and independent validation.
The Metadata Myth: Engineering for Messy Science
A persistent objection raised by R&D leaders is that experimental data generated across multiple organisations are simply too messy, inconsistent and poorly labelled to support a large-scale collaborative AI system. In my experience, this concern is far from theoretical. Anyone who has worked across research organisations knows that two highly capable teams can be generating excellent science while describing similar concepts, assays or biomarkers in entirely different ways.
Drug discovery laboratories optimise for scientific discovery, not meticulous metadata management. Experimental protocols evolve, naming conventions differ, and data structures often vary significantly between organisations.
MELLODDY's most underappreciated achievement may be how it addressed this reality.
Rather than demanding perfect human-driven data curation, the consortium engineered around the problem. Local preprocessing pipelines, including tools such as MELLODDY-TUNER, operated within each company's firewall to standardise molecular representations and prepare data for federated learning. Meanwhile, machine-learning frameworks such as SparseChem were explicitly designed to handle extremely high-dimensional and highly sparse datasets.
In other words, MELLODDY succeeded not because pharmaceutical data suddenly became clean and perfectly harmonised, but because the system was built to tolerate scientific reality.
This offers a broader lesson for organisations pursuing AI-enabled discovery. Competitive advantage rarely comes from waiting for flawless datasets. It comes from building systems resilient enough to extract useful signal from imperfect data.
The future of scientific AI may depend less on achieving perfect information and more on engineering architectures capable of thriving amid inevitable uncertainty.
The ADME Paradox: Why Automate What We Already Know?
Another intriguing question emerges from MELLODDY's reported results.
The consortium evaluated model performance across a variety of endpoints, including target activity, off-target activity and ADME-related properties. Improvements were observed across participating organisations, although results varied depending on endpoint and evaluation metric.[2]
At first glance, this seems counterintuitive.
Scientists already understand a great deal about the biological mechanisms underlying absorption, distribution, metabolism and excretion. Why invest enormous resources teaching AI to predict something researchers already study extensively?
The answer lies in the difference between understanding mechanisms and making predictions. This is a distinction I have seen repeatedly misunderstood in discussions around scientific AI. Deep biological understanding remains essential, but understanding alone does not help a medicinal chemist decide which compound to synthesise next week. Prediction and explanation solve different problems, and successful drug discovery requires both.
Knowing how metabolic enzymes function is not the same as accurately predicting the behaviour of a new molecule that has never been synthesised before. Biological knowledge provides principles; medicinal chemistry requires practical forecasts.
This is precisely where predictive modelling becomes valuable.
By learning across datasets spanning ten pharmaceutical companies, MELLODDY expanded both the amount and diversity of available training information. The resulting models achieved improvements in predictive performance and broader applicability domains across many tasks.[2]
For medicinal chemists, this means better prioritisation. More accurate predictions can help identify molecules worth synthesising, flag likely liabilities earlier and reduce unnecessary experimentation.
Importantly, however, such predictions should not be confused with forecasting clinical success.
A model that predicts an assay outcome more accurately is not automatically predicting whether a medicine will prove safe and effective in patients. Human biology remains vastly more complex than any single screening experiment.
Another advantage of MELLODDY's approach is its access to unpublished corporate data. Unlike academic literature, which may reflect publication bias toward positive findings, industrial screening databases capture extensive records of inactive compounds, failed optimisations and negative results.[2]
That broader record of scientific reality can strengthen predictive modelling. Yet even here, more data does not automatically eliminate bias. The benefits depend on the diversity, quality and representativeness of the underlying experiments.
MELLODDY's achievement was therefore not solving ADME prediction once and for all. Its accomplishment was demonstrating that predictive performance could be improved through federated learning while preserving data confidentiality.
Optimisation vs. Exploration
Ultimately, determining whether MELLODDY is "risk-averse" requires a subtle distinction.
This distinction matters because optimisation and discovery are often treated as though they are the same problem, when in practice they require very different capabilities.
The project was never designed to search all 10⁶⁰ possible molecules in chemical space.
Instead, it was built to extract greater value from existing experimental knowledge.
This highlights a crucial divide emerging within AI-driven drug discovery.
The Explorer
Generative AI systems, deep molecular design models and advanced simulation platforms seek to propose entirely new molecular structures, including candidates that may not resemble compounds already present in existing datasets.[5,6]
Their purpose is expansion.
They explore.
The Optimizer
Frameworks like MELLODDY are designed to improve predictions within and around known experimental territory.[2]
Their purpose is refinement.
They optimise.
These approaches solve fundamentally different problems.
Generative systems attempt to widen the search horizon. Federated predictive systems attempt to improve decision-making within the knowledge already available.
In my view, framing exploration and optimisation as competing approaches misses the point. Drug discovery has always required both. The challenge is not choosing between them but determining how to combine them effectively.
In fact, future breakthroughs may depend on combining both. Exploration without reliable prediction generates overwhelming uncertainty. Prediction without exploration risks reinforcing existing assumptions.
Drug discovery needs both capabilities operating together.
The Real Legacy of MELLODDY
Perhaps MELLODDY's most significant contribution is neither chemical nor biological.
It is architectural.
The project demonstrated that fierce commercial competitors could collaborate on machine learning without surrendering ownership of their proprietary data.[2,3,7,8] In doing so, it established a blueprint for privacy-preserving scientific collaboration at an unprecedented scale.
That achievement addresses a longstanding barrier in pharmaceutical R&D: valuable experimental information often remains locked within individual organisations.
MELLODDY showed that sharing insights may be possible even when sharing data is not.
Yet the deeper scientific challenge remains unresolved. From my perspective, this is where the most interesting questions in pharmaceutical AI now sit. The industry has become increasingly effective at extracting value from what it already knows. Whether AI can help us systematically discover what we do not yet know may prove far more important.
How do we combine the predictive power of historical experimental evidence with methods capable of evaluating chemistry beyond the boundaries of past experience?
How do we ensure AI learns not only from what researchers have tested, but also from what they have ignored?
These questions sit at the heart of the 10⁶⁰ paradox.
The theoretical chemical universe is vast beyond imagination. MELLODDY represents a powerful way of navigating the tiny corner humanity has already explored. Whether future AI systems can reliably push beyond those boundaries into genuinely unfamiliar territory remains one of the most important questions in modern drug discovery.
If there is one lesson I take from MELLODDY, it is that scale alone does not guarantee discovery. Larger datasets and better predictions are valuable, but scientific breakthroughs often emerge when established assumptions are challenged rather than refined.
The future belongs to systems capable of doing both.
Disclosure
Generative AI tools were used to assist with editing and language refinement. All analysis, interpretations, conclusions, fact-checking, and source verification are the author's own.
References
[1] Reymond, J.-L., van Deursen, R., Blum, L. C. & Ruddigkeit, L. (2010). Chemical space as a source for new drugs. MedChemComm, 1, 30-38. doi: 10.1039/C0MD00020E
[2] Heyndricks, W. et al. (2023). MELLODDY: Cross-pharma Federated Learning at Unprecedented Scale Unlocks Benefits in QSAR without Compromising Proprietary Information. J Chem Inf Model, 64(7), 2331–2344. doi: 10.1021/acs.jcim.3c00799
[3] MELLODDY Consortium. FAQs and project partners. Available at: IHI Innovative Health Initiative
[4] Weaver, S. & Gleeson, M. P. (2008). The importance of the domain of applicability in QSAR modeling. Journal of Molecular Graphics and Modelling, 26(8), 1315-1326. doi: 10.1016/j.jmgm.2008.01.002
[5] Tong, X. et al. (2021). Generative Models for De Novo Drug Design. Journal of Medicinal Chemistry, 64(19), 14011-14027. doi: 10.1021/acs.jmedchem.1c00927
[6] Cheng, Y. et al. (2021). Molecular design in drug discovery: a comprehensive review of deep generative models. Briefings in Bioinformatics, 22(6). doi: 10.1093/bib/bbab344
[7] MELLODDY Consortium. (2020). Project meets its Year One objective: deployment of a secure platform for multi-task federated learning in drug discovery. Available at: owkin.com
[8] guardora blog (2024): Privacy-Preserving Machine Learning in Drug Discovery: Bridging Security and Innovation. Available at: guardora