Hasn’t this been tried before, like the interest in identifying libraries of single nucleotide polymorphous systems around 2000, or agreements with 23 and me that didn’t go anywhere, were ideas for using high throughput combinatorial chemistry to identify drug targets. It seems like we’ve seen this dance before, just to a different tune.///. Tempus AI, Inc. (NASDAQ: TEM), a technology company leading the adoption of AI to advance precision medicine, and Merck, known as MSD outside of the United States and Canada, today announced an expanded, multi-year collaboration aimed at accelerating the discovery and development of precision medicine biomarkers and supporting Merck’s oncology and potentially broader therapeutic portfolios. “This collaboration builds on our existing relationship and reflects our shared commitment to harnessing the power of multimodal datasets with AI to deliver better options for patients,” said Ryan Fukushima, CEO, Data & Apps at Tempus. “We’ve spent years configuring our Lens Platform to seamlessly leverage our library of de-identified multimodal data with the necessary AI computing power to train and fine-tune specific models for healthcare. Working with the great scientists at Merck, we have exciting opportunities to translate the insights from AI models into precision medicine strategies and improve patient outcomes across oncology and beyond.” Under the terms of the agreement, Merck will use Tempus’ de-identified data along with Tempus’ Lens Platform and Workspaces environment, which offers an advanced computational configuration powered by one of the industry’s largest GPU infrastructures, which enables researchers to efficiently conduct complex analyses on training-ready multimodal datasets, generating novel insights to accelerate the development and optimization of candidate therapies at scale. “The combination of new AI technologies and large curated multimodal data sets are transforming the way we conduct discovery research,” said George Addona, Senior Vice President, Discovery, Preclinical Development and Translational Medicine, Merck Research Laboratories. “This collaboration with Tempus positions Merck to advance our precision oncology strategy through the application of the latestAI/ML capabilities to discover novel precision biomarkers, identify mechanisms of cancer cell resistance, and inform rational combinations for drugs in our early pipeline.”
Yes—your instinct is quite accurate. The Tempus–Merck announcement is part of a recurring cycle in biomedical R&D where a new technology promises to unlock drug targets from large biological datasets. Variants of this idea have appeared several times in the past 30 years. The tune changes (SNPs, combinatorial chemistry, genomics, AI), but the underlying dream—systematically mining biology for drug targets—remains the same.
Here are a few of the earlier “dances” that resemble what Tempus and Merck are describing.
1. SNP-mapping and genomic target discovery (late 1990s–early 2000s)
Around 2000, there was enormous enthusiasm for creating large libraries of single nucleotide polymorphisms (SNPs) across the human genome.
Companies like:
-
Celera
-
Millennium Pharmaceuticals
-
deCODE Genetics
argued that linking SNP patterns to disease risk would reveal entirely new drug targets.
The logic was:
-
Map SNPs across populations
-
Associate SNP haplotypes with disease
-
Identify causal genes
-
Build drugs against those pathways
Pharma companies signed multi-year data access deals very similar in tone to today’s Tempus collaboration.
The outcome:
-
Some scientific insights
-
A few targets
-
Far fewer drugs than hoped
The main problem was that complex diseases rarely reduce to simple genetic targets.
2. High-throughput combinatorial chemistry (mid-1990s)
Before genomics, pharma believed chemistry itself could generate drugs industrially.
The idea:
-
Create millions of compounds using automated synthesis
-
Screen them against biological targets
-
Let scale solve the discovery problem
Companies built enormous compound libraries and robotic screening facilities.
What happened:
-
Screening scale increased dramatically
-
But hit quality remained low
-
Biology, not chemistry, was the bottleneck
By the early 2000s, pharma quietly pivoted toward biology-driven discovery.
3. Consumer genetics data (23andMe partnerships)
Another echo of today’s strategy appeared in the 2015–2020 era.
Several pharma companies partnered with 23andMe to use their massive genetic database.
For example:
-
GlaxoSmithKline invested $300M in 23andMe
-
The promise: mine millions of genomes to discover drug targets
The logic was almost identical to the Tempus story:
-
Large dataset
-
AI / statistical analysis
-
New therapeutic targets
Results so far:
-
Some promising early programs
-
But few major clinical successes yet
Again, the bottleneck is translating correlations into actionable biology.
4. Systems biology / omics platforms (2000–2015)
There were also waves involving:
-
Gene expression arrays
-
Proteomics platforms
-
Network biology
-
“Systems medicine”
Each promised that integrated biological data would reveal hidden pathways.
These efforts produced valuable science, but drug discovery remained slow and uncertain.
5. What is actually new in the Tempus model
The Tempus–Merck collaboration is not entirely the same as those earlier efforts. A few things really are different:
1. Multimodal clinical data
Tempus combines:
-
Genomics
-
Pathology images
-
Clinical outcomes
-
Treatment histories
Earlier datasets were usually single-modality.
2. Real-world oncology data
Tempus built its database from clinical oncology practice, not just research cohorts.
This potentially allows discovery of:
-
drug resistance mechanisms
-
response predictors
-
combination strategies
3. Modern machine learning
AI models can analyze:
-
pathology images
-
sequencing data
-
clinical records simultaneously
That kind of integrated modeling was not technically feasible in 2000.
6. The deeper recurring problem
Despite the new tools, the core difficulty remains:
Biology is extremely messy.
Even with massive datasets:
-
correlations are easy
-
causal mechanisms are hard
Drug discovery still requires:
-
experimental validation
-
animal models
-
clinical trials
No dataset can eliminate those steps.
7. The realistic interpretation
The Tempus–Merck deal should probably be viewed as:
a discovery-enabling data partnership, not a drug discovery revolution.
Best case:
-
identify biomarkers
-
stratify patients better
-
discover a few useful targets
Worst case:
-
lots of correlations
-
little therapeutic payoff
Historically, most waves land somewhere in the middle.
A slightly cynical industry rule
Veterans of pharma R&D often say:
Every decade we invent a new technology that promises to industrialize drug discovery.
Examples:
-
combinatorial chemistry
-
genomics
-
proteomics
-
systems biology
-
AI
Each improves tools somewhat, but drug discovery stubbornly remains artisanal.
If you’d like, I can also explain why Tempus specifically is attractive to pharma despite this history, and why companies like Merck are still willing to pay for these collaborations. It actually has more to do with clinical trial strategy and biomarker enrichment than with discovering entirely new drugs.
The announcement of a multi-year collaboration between Tempus AI and Merck to use large multimodal datasets and artificial intelligence to discover precision medicine biomarkers has a familiar ring to anyone who has watched biomedical innovation cycles for several decades. The language—large datasets, computational analysis, discovery of novel drug targets—echoes earlier waves of enthusiasm that promised to industrialize the discovery of new therapies. The technology evolves, but the core ambition remains remarkably constant: use large-scale biological data to systematically identify therapeutic targets.
Around the year 2000, the most prominent version of this idea centered on single nucleotide polymorphisms (SNPs). After the Human Genome Project, biotechnology companies and academic groups began cataloging large libraries of SNP variants across human populations. The hope was straightforward. If particular SNP patterns correlated with diseases such as diabetes, Alzheimer’s disease, or cancer, then the associated genes might reveal the underlying biological pathways. Once those pathways were identified, drug developers could design therapies that modulate them. Companies such as Celera, Millennium Pharmaceuticals, and deCODE Genetics built business models around precisely this concept. Pharmaceutical companies signed data-sharing agreements that sounded strikingly similar to today’s AI-driven collaborations.
The scientific insights were real. The genetic architecture of many diseases became clearer, and certain targets did emerge. Yet the overall expectation—that genome-wide association studies would rapidly produce a pipeline of new drugs—proved overly optimistic. Most complex diseases are polygenic and mechanistically complicated, and genetic correlations rarely translate directly into tractable therapeutic targets.
An earlier wave, in the mid-1990s, focused not on biology but on chemistry. At that time, pharmaceutical companies invested heavily in high-throughput combinatorial chemistry, an approach that sought to generate millions of chemical compounds automatically and screen them against biological targets. The belief was that scale alone could solve drug discovery: if enough molecules were synthesized and tested, useful drugs would inevitably appear. Massive robotic screening facilities and compound libraries were constructed across the industry.
The results were instructive. Screening throughput increased dramatically, but the number of genuinely useful drug leads did not increase proportionally. Researchers eventually realized that the bottleneck was not the number of molecules but the quality of biological targets and the complexity of disease biology. The industry gradually shifted toward more biology-driven discovery programs.
More recently, another variant of the same idea appeared in partnerships involving consumer genetic databases, particularly those built by companies such as 23andMe. Pharmaceutical firms recognized that millions of genotyped individuals might reveal disease-associated genetic variants at an unprecedented scale. The collaboration between GlaxoSmithKline and 23andMe, for example, promised to mine population genetic data for drug targets using advanced statistical and machine-learning methods. Once again, the language emphasized large datasets, computational analysis, and systematic discovery of new therapies.
These collaborations have produced some promising programs and interesting biological insights, but the transformative wave of new drugs that many expected has not yet materialized. As before, identifying statistical associations in data has proven much easier than translating them into actionable biological mechanisms and safe, effective therapeutics.
Seen in this historical context, the Tempus–Merck collaboration appears less like a revolution and more like the latest technological iteration of a recurring strategy. What has changed is the nature and scale of the data. Tempus has assembled a large database that integrates several types of information simultaneously—genomic sequencing, clinical records, treatment outcomes, and pathology images. In addition, modern machine-learning systems can analyze such heterogeneous datasets in ways that were technically infeasible twenty years ago.
The distinctive features of the current approach are therefore:
-
Multimodal data integration: combining genomic, imaging, and clinical information rather than analyzing each type separately.
-
Advanced machine learning: models capable of identifying patterns across diverse datasets.
These advances genuinely expand the analytical toolkit available to researchers. Yet the fundamental challenge remains the same as it was during earlier waves of genomic or combinatorial enthusiasm: biological causation is extraordinarily complex. Statistical correlations emerging from large datasets must still be validated experimentally, interpreted biologically, and ultimately tested in clinical trials. None of these steps can be replaced by computation alone.
For that reason, many experienced observers view such collaborations in a more pragmatic light. The greatest near-term value of platforms like Tempus may not lie in discovering entirely new drugs but rather in improving biomarker discovery and patient stratification. In oncology, identifying which patients respond to a particular therapy, or why resistance develops, can be just as important as discovering a new molecular target. Large clinical datasets linked to treatment outcomes are particularly powerful for that purpose.
Indeed, the pharmaceutical industry has seen enough cycles of technological enthusiasm to recognize a recurring pattern. Roughly every decade, a new technology promises to transform drug discovery into a more systematic and scalable process. Among the successive candidates have been combinatorial chemistry, genomics, proteomics, systems biology, and now artificial intelligence. Each wave contributes useful tools and insights, but none has eliminated the inherent uncertainty of translating biological knowledge into safe and effective medicines.
In that sense, the Tempus–Merck collaboration represents both progress and continuity. The analytical tools are more sophisticated and the datasets vastly larger than those available in earlier eras. Yet the underlying aspiration—to mine large biological datasets in search of therapeutic breakthroughs—remains a familiar theme in the history of biomedical innovation. The dance continues, even if the music has changed.
