If the NMR spectrum is an "encrypted book of heaven" written by life itself, then the Transformer is the intelligent "decipherer" capable of extracting structure, semantics and logic from a vast flood of signals.
After the launch of the "Lighthouse Program," many partners raised one key question:Where exactly does Elixir™'s true technological moat lie?
Today we go down to the foundations and reveal, in full, how Elixir™ turns a complex NMR spectrum into life-science insight that is understandable, explainable and reproducible.
Part 1 The source: the NMR spectrum — life's "encrypted book"
When a tube of blood enters the NMR platform, what is produced is not an image but a multi-dimensional spectrum containing millions of data points.
This spectrum is the raw "full orchestral score" formed by the superposition and entanglement of every molecular signal in the body. Every tiny peak shape, chemical shift and intensity is a "note" on that score, and together they record the real-time state of the living system.
The difficulty with this spectrum lies in:
Signal overlap:
Multiple molecular peaks intertwine to form complex superpositions;Noise interference:
Weak but meaningful biological signals must be separated out from the background;Systemic correlation:
A subtle change in any single peak may involve the coordinated action of several metabolic pathways.
Conventional analytical methods are like "listening only to a few instruments playing solo," whereas what Elixir™ sets out to do is "understand the theme and emotion of the entire symphony."
Part 2 Deciphering: the "three realms" of the Transformer
The Transformer architecture adopted by Elixir™ is precisely this "top-tier decipherer," trained on massive volumes of data. Its workflow can be distilled into three progressively deeper realms.
01 Tokenization — turning the "encrypted book" into "basic units" that can be understood
This is the first step toward understanding. The high-dimensional, continuous NMR spectrum is structurally processed and discretized into hundreds or thousands of ordered tokens.
These tokens are not mechanical slices; they carry the morphological features of local spectral regions, and through pre-training they continually build associations and generate higher-level biological semantics.
In this way, an originally complex and chaotic spectrum is translated into a "molecular language system" that AI can understand and reason over.
02 Attention mechanism — understanding the contextual relationships between "basic units"
This is the very soul of the Transformer. The model dynamically computes attention weights for every token, automatically focusing on the parts most relevant to the task amid a flood of signals, while fading out irrelevant interference.
When identifying liver cancer risk, for example, the model will:"Ignore" metabolic fluctuations caused by short-term diet;"Focus" on key signal changes in pathways such as bile acid metabolism and hepatic ammonia detoxification.
This enables it to read the "precise collaboration network" formed among molecules, understanding the synergy and antagonism of signals from a global perspective rather than simply measuring their concentrations.
03 Semantic output — from "understanding" to "insight"
Building on its understanding of global context, the model is able to carry out two types of higher-order task:● Discrimination tasks:
Outputting a "high risk" or "low risk" conclusion together with the corresponding confidence level. This is what most AI models can already do.● Discovery tasks:
Identifying "new grammatical patterns" that human experts have never defined, formed by the coordination of multiple weak signals — this is the underlying support for our discovery of new disease subtypes and for multimodal data fusion.
Part 3 Elevation: from "technical deciphering" to "empowering research"
The value of Elixir™ lies not in replacing scientists, but in extending the boundaries of human understanding. We are committed to giving scientists unprecedented powers of insight:
●
Expanding the "perceptual resolution" of living systems●
Capturing high-order signals and correlation patterns the human eye cannot recognise●
Maintaining stable inference in complex scenarios full of noise and intertwined variables●
Providing testable, traceable data clues for unknown phenomena
In research settings, these capabilities take concrete shape as follows:
▶ In exploratory research:
Revealing latent data structures, population subtypes and anomalous patterns.▶ In clinical cohort studies:
Identifying disease subgroups or functional stratifications that conventional indicators cannot distinguish.▶ In multimodal integration:
Constructing a unified semantic space for metabolic, protein and clinical data.
Elixir™ therefore never plays the role of "final arbiter of science." Its mission is:Surface the phenomenon → expose the structure → point to a hypothesis → hand it to scientists for verification and interpretation.
This is the next-generation research collaboration paradigm we believe in —AI as the engine of discovery, scientists as the core of judgement and decision-making.
▍ At its core, the Elixir™ large model foundation is a platform for semantic understanding and discovery of living systems. The reason we can confidently launch the "Lighthouse Program" is precisely that this underlying technology provides a solid guarantee —
▍ It can not only answer "what is," but also explore "why," and anticipate "what may come next."
▍ This is the computing foundation for the life sciences of the future.
