When algorithms stop being mere tools and become “co-creation partners” in research, the boundaries of the life sciences are being redefined. In our previous article (Elixir Explores #2: The Breakthrough — From “Measurement” to “Understanding,” the Next Stop for Medical Big Data), we proposed the key to breaking through: “NMR (nuclear magnetic resonance) provides panoramic data, Transformer provides systemic intelligence.”
In this issue, we take a closer look at the Elixir™ scientific research large model — and at how this “research brain,” trained on 160,000 cases of data, can become the core engine that helps you break through R&D bottlenecks.
Part.1 More than a model: 
A “research partner” that understands life
We are used to seeing AI models as “algorithms” or “tools.”
But the foundation of Elixir™ is more than cold, impersonal code —
it is more like a specialist physician trained at a top medical school,
endowed with a deep reserve of “biological common sense.”
After pre-training on 160,000 plasma/urine NMR samples (including a cohort of more than 100,000 healthy individuals), Elixir™ has already:
▲ Built a precise “health baseline”:
A deep understanding of the normal fluctuation range of molecular signals in healthy people, enabling effective filtering of individual variation and noise interference.
▲ Learned the grammar of the “language of biomolecules”:
Mastering the underlying rules of synergy, antagonism and interaction among molecules.
▍This means what you are facing is not a “model,”
▍but a research partner that can read the language of molecules.
Part.2 The technical core: how the Transformer
“Thinks” through complex living systems
Elixir™ adopts the Transformer architecture that underpins ChatGPT, but the language it was trained on is humanity’s most complex language — life itself.
●“Attention mechanism”: capturing the “quantum entanglement” between molecules
Conventional models are like readers who can only read word by word, struggling to grasp the meaning of a whole passage.
Elixir™ works more like a highly efficient expert —
able to scan an entire NMR spectrum at once,
computing the strength of association between every signal and all the others,
and automatically focusing on the molecular combinations most relevant to a disease or physiological state.
▍Even when the change in a single molecule is tiny,
▍it can still discern the systemic signal behind it.
●“Parallel processing”: from “serial computation” to “parallel insight”
In the past, we used to say: “to compute one person, you would need fifty years.”
Now, the Transformer has turned that into a few minutes.
With its powerful parallel computing capability,
it can rapidly resolve, across vast numbers of molecules,
high-dimensional, non-linear interactions,
making the discovery ofsystem-level biomarkers possible.
▍This is how AI “thinks” about life —
▍moving from fragmented observation to systemic understanding.
Part.3 Pre-training + fine-tuning:
Letting research “stand on the shoulders of giants”
Elixir™ follows the efficient paradigm of AI —
pre-training (general education) + fine-tuning (specialized training).
● Stage 1: Pre-training — learning the universal rules of life
● Output: a powerful foundation model equipped with biological common sense
● Stage 2: Fine-tuning — rapid specialization for a specific disease
● Input: a small amount of labeled, disease-specific data (e.g., pancreatic cancer samples)
● Output: a customized, high-performance disease model
▍This means researchers no longer need to start from scratch —
▍with just a few hundred samples, they can achieve a breakthrough “standing on the shoulders of giants.”
Part.3 Moat and value:
Why Elixir™?
In summary, the moat of the Elixir™ large model comes from a closed loop of three elements:
▲ Data moat:
An exclusive, large-scale NMR database of the Chinese population — the very foundation for training an accurate model.
▲ Algorithm moat:
A Transformer architecture optimized and tuned for the language of biomolecules, providing powerful “systems thinking” capability.
▲ Paradigm moat:
The “pre-training + fine-tuning” model, which turns cutting-edge AI into standardized services that can be deployed rapidly.
This is the very foundation that enabled the LUCA™ lung cancer early screening model to be developed rapidly and reach an AUC > 0.95.
LUCA™ is only the beginning. The capabilities of Elixir™ will continue to expand into more disease models and research directions.
In the next article, we will use the LUCA™ lung cancer early screening model as an example to break down in detail how you can build on this powerful foundation to create a disease model of your own. Stay tuned.
