Throughout history, groundbreaking scientific discoveries have often emerged from unexpected connections. Optogenetics revolutionized neuroscience by combining microbial photoreceptors with neuromodulation, while CRISPR-Cas9 transformed biotechnology by reimagining bacterial immune systems as gene-editing tools. These paradigm shifts arose not from incremental improvements, but from creative synthesis across seemingly unrelated domains. From an engineer’s viewpoint, this observation begs the question:

Can creative breakthroughs be systematically reproduced—or even better, be automated?

We propose Spacer as an answer to this challenge. Through a unique methodology we call 'deliberate decontextualization,' this system demonstrates that AI can derive novel scientific concepts absent from existing knowledge bases. We stand at a critical juncture where AI may evolve from serving merely as a sophisticated research instrument to becoming an independent architect of scientific breakthroughs.

Current Limitations in AI for Science

Today's AI systems deliver impressive results across various scientific domains. AlphaFold predicts protein structures, machine learning models screen drug candidates, and analyze genomic data. However, all these systems operate within predefined search spaces established by humans.

Large Language Models (LLMs) face similar constraints. While they excel at recombining patterns from training data, they approximate probabilistic reconstructions of existing knowledge rather than generating genuinely novel concepts. When prompted to propose biological research ideas, LLMs repeatedly suggest variations of "new therapies using CRISPR" precisely because CRISPR has been overrepresented as the canonical example of innovation throughout their training data.

This limitation stems from the fundamental characteristics of the transformer architecture. While the structure excels at generating stable text, it has sacrificed its ability to envision paradigm-shifting thoughts by design—optimizing for coherence and suppressing outputs that deviate from established patterns.

Decontextualization: A Paradoxical Path to Creativity

Spacer approaches this problem through counterintuitive means. By decomposing information into atomic units called "keywords" and then exploring unexpected connections between them, the system creates new contexts by removing existing ones. It is akin to dismantling LEGO blocks entirely in order to construct an entirely new structure.

The system's core, the Nuri engine, navigates a keyword graph constructed from 180,000 biological papers. Each keyword becomes a node, with the edge weights being determined by the co-occurrence frequency and the Field-Weighted Citation Impact (FWCI). Nuri's role involves identifying keyword combinations that remain unconnected yet possess high potential.

Remarkably, Nuri's evaluation function classifies high-impact papers with an AUROC of 0.737. This suggests that connection patterns between keywords genuinely correlate with scientific impact.

Schematic of Spacer’s approach to engineered scientific inspiration.

From Thesis to Statement: Multi-stage Manifesting Pipeline

Spacer employs a sophisticated three-stage pipeline to transform high-potential keyword combinations extracted by Nuri into robust hypotheses, leveraging LLM knowledge and reasoning capabilities to explore coherent connections between keywords.

The first stage, the Revealing Framework, uncovers hidden conceptual links between keywords to explore how they can be integrated into a coherent research thesis.

The second stage, the Scaffolding Framework, transforms the initial thesis into logic graphs to ensure structural completeness. Each scientific claim is represented as a graph node, with causal relationships and logical connections depicted as directed edges. This graph structure is crucial in detecting and rectifying loopholes or fallacies. Notably, this process automatically searches and analyzes 100–200 peer-reviewed papers, basing each logical connection on actual scientific evidence. Each rationale is linked to a specific DOI for later tracing and verification.

The final Assessment Framework evaluates the scientific plausibility and feasibility of constructed hypotheses. This stage assesses several qualities of the proposed hypotheses, such as non-contradiction with existing scientific knowledge, verifiability within current technology, and potential impact. This evaluation system successfully identifies scientifically problematic hypotheses with an 88.2% recall rate.

Overall architecture of Spacer.

A Detailed Case: Restoring Calcium Oscillations in Hepatocellular Carcinoma

For perspective on the hypotheses that Spacer derives, let us take a closer look at an actual case. The system focused on the differences in calcium homeostasis between hepatocellular carcinoma cells and normal hepatocytes. Previous research has shown that calcium ions in normal cells exhibit regular oscillatory patterns synchronized with the cell cycle, whereas cancer cells display irregular or absent oscillations.

Spacer built on this knowledge by introducing a concept from physics—stochastic resonance, the principle that appropriate noise levels can actually enhance signal coherence in nonlinear systems. By connecting these two domains, the system derived the hypothesis that externally controlled calcium concentration changes could restore disrupted calcium oscillations in cancer cells.

We developed comprehensive experimental protocols using Grok 4 to verify the practical feasibility of the hypothesis. The results proved remarkably detailed and immediately actionable.

Spacer’s output and corresponding experimental protocol of Restoring Calcium Oscillations in Hepatocellular Carcinoma. See full-text from Supplements.

Confirming the Approach

Spacer interprets scientific discovery as a network phenomenon, reducing it to a graph-theoretic problem. This enables systematic exploration of new concepts that can emerge from current knowledge levels.

Notably, Spacer successfully reconstructed over 85% of core concepts from the latest cutting-edge publications in top-tier journals, such as Nature and Science, using just a few keywords alone. This suggests that any novel scientific concept can indeed be expressed as a connection of keywords.

Furthermore, an embedding space analysis revealed that Spacer's outputs were semantically closer to published papers than those from state-of-the-art LLMs, such as GPT-5 and Gemini 2.5 Pro. This demonstrates that the decontextualization approach can effectively overcome existing limitations of LLMs.

Spacer's Limitations and the Journey Toward Scientific Superintelligence

Currently, it would be hasty to call Spacer a scientific superintelligence. While the system can reveal original scientific concepts, it cannot independently verify or develop them. However, technological progress may overcome these limitations. As robotics and automated laboratory technologies advance, end-to-end scientific discovery cycles that automatically verify AI-proposed hypotheses are coming into reality.

When these systems are left to conduct autonomous research iteratively—formulating hypotheses, designing experiments, analyzing failures, and integrating learnings—they may ultimately achieve cognitive capacities worthy of the title “scientific superintelligence.”

Accelerating Scientific Discovery

Spacer presents an intriguing future. If such systems operate at scale, scientific discoveries that would have taken decades might occur within months. In the near future, paradigm shifts that once relied on individual intuition and serendipity would occur systematically and affordably. Furthermore, researchers would gain access to interdisciplinary insights that had previously been available only to a few with expertise across all related fields.

Of course, this change does not mean the complete replacement of human scientists' roles. Rather, humans will assume higher-dimensional roles, selecting valuable possibilities from AI's numerous proposals and applying them to real-world problem solving.

Despite its current limitations, Spacer validates a crucial premise: machines can derive genuine originality. What once belonged purely in the realm of speculations—the notion of scientific superintelligence—has now materialized into an achievable objective supported by technological foundations.

Observing the evolution of these systems and their impact on humanity's scientific and technological advancements will be one of the most fascinating subjects of our era.

Minhyeong Lee

CEO and Founder of Asteromorph