Despite shedding its “undruggable” label, KRAS in non-small cell lung cancer (NSCLC) remains a clinical dilemma. KRAS inhibitors have changed the therapeutic landscape, but they have not yet reproduced the clarity of the driver-to-drug paradigm seen in treatment-naive EGFR-mutant disease [1–5]. Consequently, many patients with KRAS-mutant NSCLC remain treated within broader chemotherapy, immunotherapy, or chemoimmunotherapy frameworks [6].

The clinical dilemma of KRAS may be better understood as a problem of context: the biological and therapeutic meaning of a mutation is shaped by the tumor state in which it is embedded. This is the sense in which KRAS raises a question of tumor semantics: whether the same mutation carries the same biological and therapeutic meaning across different tumor states.

The same mutation, situated within different constellations of co-mutation, immune ecology, transcriptional state, and treatment history, may indicate markedly different therapeutic dependencies [7]. The contrast with EGFR is instructive. In treatment-naive EGFR-mutant NSCLC, an activating EGFR mutation often identifies the oncogenic lesion, the dominant survival dependency, and the therapeutic target simultaneously; the clinical syntax is relatively shallow, because driver, dependency, and drug are closely aligned, and inhibition of the aberrantly activated kinase frequently leads to predictable tumor regression. KRAS behaves differently. It may mark an early driving event [8], but it does not by itself define the tumor’s current dependency structure. Its meaning is filtered through deeper layers of co-mutation, immune ecology, transcriptional state, lineage plasticity, metabolic adaptation, and treatment history [9]. For example, STK11/LKB1 alterations have been associated with primary resistance to PD-1 inhibition in KRAS-mutant lung adenocarcinoma [10], whereas co-alterations involving KEAP1, SMARCA4, CDKN2A, and related genes have been linked to poor monotherapy efficacy or early progression with KRAS G12C inhibitors [11].

How, then, should KRAS be read in context?

If KRAS is imagined as a word, the tumor is the sentence—and this sentence is crowded with modifiers, qualifiers, and implicit grammatical relations like KEAP1 or STK11. Some of these relations have been recognized; others remain unresolved or latent. It is at this point that the idea of a KRAS context naturally evokes the Transformer architecture used in large language models. The foundation of that architecture is “Attention Is All You Need” [12]. Its key insight is not simply that a model remembers more words, but that each token is represented as a function of its context. The meaning of a word is no longer fixed in isolation by a dictionary; it is dynamically shaped by its relations with surrounding tokens.

KRAS may therefore be viewed as a context-dependent token whose biological meaning is generated within the broader tumor sentence. One could imagine, at least conceptually, a future KRAS-context Transformer whose purpose would not be to assign a static hierarchy of importance to KRAS, STK11, KEAP1, CD8, NRF2, and other markers. Its more ambitious task would be to learn, within a specific tumor context, which variables explain others, which reshape others, and which dependency currently governs therapeutic response. When KRAS G12C is accompanied by high MAPK activity, RTK feedback, and dependence on SHP2/SOS1, KRAS signaling may remain the dominant actionable dependency [13], and the therapeutic hypothesis naturally points to vertical KRAS inhibition combined with blockade of feedback loops. When it is accompanied by STK11/KEAP1 inactivation, high NRF2 activity, low CD8 infiltration, myeloid enrichment, and a YAP/EMT state, KRAS may still be an early driver, but it is less likely to be the dominant current dependency. Research attention may then need to shift toward interventions targeting immune exclusion, metabolic stress, or lineage plasticity. Although the introduction of a large-language-model architecture remains, for now, a conceptual analogy, advances in multimodal AI pathology and spatial transcriptomics have begun to create the technical conditions for using cellular spatial distances and molecular abundance as features from which machine-learning models can learn contextual weights or attention [14].

If the essential task of lung cancer precision medicine over the past two decades was to find the driver gene, the next task may be to understand the meaning of that driver gene within the current tumor context. The lesson from KRAS may therefore extend beyond one oncogene. It points to a conceptual shift from driver identification to tumor semantics. In this framework, lung cancer may require a data-driven dynamic coordinate system capable of decoding mutational meaning, anticipating state transitions, and guiding system-level intervention. A mutation can name an entry point; it cannot, on its own, define a context-dependent tumor state.