CPTAC-PROTSTRUCT: Proteomics in Oncology

• CPTAC-PROTSTRUCT is a proteomics instruction-tuning dataset that uses individualized tumor profiles to enable patient-specific molecular reasoning in oncology.
• It employs a two-stage process—schema alignment and clinical reasoning—to teach models to interpret and respond to proteomic measurements with natural language.
• By integrating protein-protein interaction networks via the KRONOS framework, the approach enhances predictive accuracy for outcomes like mortality and survival.

CPTAC-PROTSTRUCT is a patient-level instruction-tuning dataset for molecular understanding in oncology, introduced together with KRONOS, a graph-LLM framework for patient-specific clinical reasoning from proteomics (Adam et al., 26 Sep 2025). It is presented as the first instruction-tuning dataset for proteomic understanding in oncology, derived from individualized proteomic profiles curated from the Clinical Proteomic Tumor Analysis Consortium (CPTAC), and organized to teach a model both how to interpret proteomics as a modality and how to answer clinically meaningful natural-language questions grounded in that modality. The associated framework treats a patient not as a flat abundance vector but as a proteomics-informed molecular network, combining protein-protein interaction topology with patient-specific abundance measurements for structured reasoning about cancer type, histologic subtype, pathologic stage, recurrence risk, treatment response, and survival-related outcomes.

1. Definition, motivation, and conceptual scope

CPTAC-PROTSTRUCT was introduced to address two limitations identified in oncology-oriented multimodal language modeling: the absence of instruction-tuning data for patient-level proteomic interpretation, and the absence of architectures tailored to biomolecular interaction structure (Adam et al., 26 Sep 2025). The dataset is therefore not a conventional classification table. It is an instruction-response corpus designed to teach a model how to answer open-ended questions about individualized proteomic states.

The clinical motivation is that proteomics reflects the real-time manifestation of disease pathogenesis more directly than genomics alone. In cancer, this includes active cellular response, tumor progression, treatment adaptation or resistance, inter-patient heterogeneity, and pathway-level disruption. The paper argues that individualized tumor behavior, therapeutic resistance, recurrence, prognosis, and disease heterogeneity often arise from coordinated molecular network behavior rather than isolated protein abundance values. This motivates a network-aware representation in which interacting proteins, complexes, and pathways are treated as part of the relevant clinical signal rather than as background biological context.

The dataset is assigned an explicitly educational role with two stages. The first is schema alignment, in which the model learns to navigate the proteomics modality and associate molecular values with language. The second is clinical reasoning, in which the model uses the same molecular information for higher-level oncologic interpretation. This suggests that CPTAC-PROTSTRUCT was designed not only as supervision for downstream prediction, but also as a modality-alignment substrate linking proteomic measurements to natural-language reasoning.

Formally, the instruction-tuning corpus is defined as

$\mathcal{D} = \{(P_i, q_i, a_i)\}_N,$

where $P_i \in \mathbb{R}^{M \times d}$, $q_i$ is the instruction, and $a_i$ is the answer (Adam et al., 26 Sep 2025). The formalism makes clear that the central unit is a patient-specific proteomic representation paired with a natural-language question and a gold response.

2. Source data, cohort composition, and curation pipeline

CPTAC-PROTSTRUCT is built from CPTAC proteomics, described in the paper as the largest U.S. proteomics cancer study, using the extraction pipeline from Payne et al., “Simplified and Unified Access to Cancer Proteogenomic Data” (Adam et al., 26 Sep 2025). Each patient is represented by at least one sample, with at most one tumor sample and at most one normal sample per patient.

After preprocessing, the dataset contains 1,658 samples, comprising 1,086 tumor samples and 572 normal samples, with 4,824 proteins spanning 10 cancer types (Adam et al., 26 Sep 2025). The abstract states that the instruction dataset is derived from more than 1000 patients, but the exact per-cancer counts and the names of the 10 cancer types are not listed in the provided paper text. The generation pipeline also uses protein-protein interaction structure from STRING, clinical metadata associated with CPTAC patients, and curated biological knowledge from Reactome, OncoKB, and COSMIC. The paper is explicit, however, that the foundation model inputs are primarily the proteomic abundance profiles themselves.

The curation pipeline applies several thresholded preprocessing steps before instruction generation. Patients with more than $35\%$ missing protein data were removed, producing a $5\%$ dropout rate. A core protein set was then defined by selecting proteins present in more than $70\%$ of both tumor and normal samples, reducing the feature space from 16,869 proteins to 5,078 proteins. Missing values in this core set were imputed using Multiple Imputation by Chained Equations (MICE) within each cancer type, in order to preserve tissue-specific patterns. The bottom $5\%$ lowest-variance proteins per dataset were removed as uninformative, yielding a final feature set of 4,824 proteins (Adam et al., 26 Sep 2025).

Instruction generation does not query every measured protein. Instead, the authors curate a queryable protein subset using three criteria: high-impact proteins, high-variance proteins, and cancer-associated proteins. High-impact proteins are identified by PPI network degree centrality using a top $10\%$ rank threshold, together with Reactome pathway priority for proteins involved in cell cycle regulation, DNA damage response, metabolism, and established drug targets. High-variance proteins are selected using a $10\%$ threshold. Cancer-associated proteins are taken from OncoKB and COSMIC. This makes the dataset clinically motivated and computationally tractable rather than exhaustive.

3. Instruction-tuning construction and task structure

CPTAC-PROTSTRUCT contains two subsets corresponding to two training phases: a schema alignment dataset and a clinical reasoning dataset (Adam et al., 26 Sep 2025). The overall corpus size, using the explicit counts in the methods section, is 380,969 instruction-response examples, although the paper also reports “over 370k” in the abstract and “over 380,000” in the introduction. The explicit subset totals are 354,812 schema alignment questions and approximately 26,157 clinical reasoning QA pairs.

Subset	Examples	Purpose
Schema alignment	354,812	Modality understanding
Clinical reasoning	$P_i \in \mathbb{R}^{M \times d}$0	Higher-level oncologic reasoning
Total	380,969	Full instruction corpus

The schema alignment subset teaches direct interpretation of proteomic measurements. It includes five question categories: direct protein abundance queries, abundance threshold queries, ranking and ordering queries, comparative abundance queries, and interaction network-based abundance queries. Example prompts include “Find the proteins whose measurements exceed two standard deviations from the mean value,” “Could you tell me the concentration of $P_i \in \mathbb{R}^{M \times d}$1 in this patient?,” and “Which proteins belong to the uppermost 90\% when ranked by their abundance?” Example gold responses are short natural-language statements such as “The abundance level of SMARCA5 in this patient is 0.101.” Questions were template-generated and then paraphrased using DeepSeek-R1-Distill-Qwen-32B to increase linguistic diversity while preserving semantics.

The clinical reasoning subset targets treatment response prediction, tumor code prediction, overall survival days, histologic grade, pathological stage, recurrence risk, and histologic subtype. Example prompts include “What does the molecular network predict for treatment response?,” “Predict overall survival days based on the molecular profile,” and “Determine histologic grade and pathological stage from the molecular network.” Example gold responses include “The patient is predicted to experience recurrence” and “Based on the molecular signature, the histologic type is Glioblastoma.” These QA pairs were generated using DeepSeek-R1-Distill-Qwen-32B with manually created few-shot examples and associated clinical metadata as contextual input, with the stated goal of resembling questions an oncologist might ask when interpreting proteomic profiles (Adam et al., 26 Sep 2025).

Responses are open-ended natural language rather than class identifiers, although many tasks collapse onto discrete clinical labels in phrased responses. The paper does not provide a full formal ontology of label spaces for all QA types, and it does not describe a formal train-validation-test split for CPTAC-PROTSTRUCT itself. What is described instead is downstream evaluation using 5-fold nested cross-validation.

4. KRONOS: patient-specific graph-LLM integration

KRONOS, short for Knowledge Representation of patient Omics Networks in Oncology via Structured tuning, is the modeling framework built to use CPTAC-PROTSTRUCT for patient-specific molecular reasoning (Adam et al., 26 Sep 2025). Its core procedure is to begin from a STRING PPI network, inject the patient’s proteomic abundance measurements as node features, encode the resulting patient-specific graph with a GNN, project the graph embedding into the LLM token embedding space, prepend it as a special token to the text instruction, and generate a response conditioned on that structured molecular state.

Each patient graph is written as

$P_i \in \mathbb{R}^{M \times d}$2

with

$P_i \in \mathbb{R}^{M \times d}$3

where topology is supplied by STRING and node features are patient-specific proteomics-informed values (Adam et al., 26 Sep 2025). The graph encoder is denoted $P_i \in \mathbb{R}^{M \times d}$4, with layerwise message passing

$P_i \in \mathbb{R}^{M \times d}$5

initialized by $P_i \in \mathbb{R}^{M \times d}$6. After $P_i \in \mathbb{R}^{M \times d}$7 layers, graph-level pooling yields a patient representation $P_i \in \mathbb{R}^{M \times d}$8. Pooling is treated as a tunable design choice; the paper mentions max pooling, global mean pooling, and a search over $P_i \in \mathbb{R}^{M \times d}$9.

This graph embedding is then mapped into the LLM embedding space by a trainable dense connector and prepended to the text token sequence. The LLM backbone used in experiments is Vicuna7bv1.5. For graph encoding, the paper considers GAT, GraphSAGE, and GIN, with the best-performing KRONOS variant using GAT (Adam et al., 26 Sep 2025).

Training follows a two-stage curriculum inspired by LLaVA. In Stage 1, using the 354,812 schema alignment QA pairs, the LLM backbone is frozen and only the connector and graph encoder are updated, with the goal of aligning molecular graph representations to the LLM semantic space. In Stage 2, using the 26,157 clinical reasoning QA pairs, the LLM, connector, and GNN encoder are all updated, with the goal of enabling advanced molecular reasoning and prognostic inference. This tightly couples CPTAC-PROTSTRUCT and KRONOS: the dataset provides the supervision, and the model provides the structured mechanism for proteomics-aware language reasoning.

5. Evaluation, ablations, and biological interpretation

KRONOS is evaluated on four prognostic or predictive tasks using CPTAC or TCGA outcomes: mortality prediction, cancer type classification, overall survival prediction, and stage classification (Adam et al., 26 Sep 2025). These benchmarks are not identical to the generative instruction-following tasks. Instead, the model is further adapted with a supervised linear probe on top of KRONOS representations. Evaluation uses 5-fold nested cross-validation with identical grid-search-style hyperparameter tuning across baselines.

Task	Metrics	KRONOS
Mortality prediction	AUC, F1	0.857 ± 0.025, 0.673 ± 0.031
Cancer type classification	AUC, Macro-F1	0.849 ± 0.011, 0.742 ± 0.018
Overall survival prediction	C-Index, t-AUC 1-yr	0.664 ± 0.058, 0.628 ± 0.067
Stage classification	AUC, Macro-F1	0.823 ± 0.014, 0.618 ± 0.029

The comparison set includes linear models, deep MLPs, patient similarity network GNNs, PPI graph classification methods, and biomolecular instruction-tuned LLM variants. Reported strong baselines include EMOGI+GAT, MOGONET+GAT, Vicuna7bv1.5+NODE, Elastic Net, and SVC. Across the four tasks, KRONOS attains the best reported performance in the paper. A plausible implication is that the benefit derives not only from instruction tuning, but also from the explicit use of patient-specific PPI graphs rather than patient similarity networks or raw feature encoders.

The ablation study directly tests that hypothesis. Graph encoders outperform node encoders across all tasks; among graph encoders, GAT is best, followed by GIN and then GraphSAGE. The paper highlights gains of $q_i$0 AUC in mortality prediction, $q_i$1 AUC in cancer type classification, and $q_i$2 C-Index in survival prediction for GAT-based graph encoders relative to GAT-based node encoders (Adam et al., 26 Sep 2025). This reinforces the paper’s claim that patient-specific pathogenesis is better represented as a proteomics-informed molecular network than as a flat feature vector.

The biological interpretation advanced by the paper is correspondingly network-centric. Proteomic signatures become node features; STRING edges provide mechanistic context; centrality and pathway curation prioritize cell cycle, DNA damage response, metabolism, and drug targets; and OncoKB/COSMIC enrichment focuses the instruction space on cancer-associated proteins. The framework is stated to support diagnosis, prognosis, and treatment stratification. The benchmark tasks directly support diagnosis and prognosis claims, whereas treatment stratification is more prospective and is represented in prompts such as treatment response prediction. The paper also notes that, despite references in the abstract to temporal trajectory modeling, no temporal trajectory benchmark or temporal equations are actually reported; the reported longitudinally oriented task is survival prediction rather than explicit temporal trajectory modeling.

6. Limitations, governance, and distinctions from other CPTAC resources

Several important boundaries are explicit in the paper. It does not provide the exact names and counts of the 10 cancer types, the exact number of unique patients beyond “more than 1000,” a formal train-validation-test split for the instruction dataset itself, a formal loss function for instruction tuning, temporal modeling equations, explicit interpretability visualizations or pathway attributions, or a comprehensive label ontology for all reasoning QA tasks (Adam et al., 26 Sep 2025). The authors also identify two major limitations: distribution shift sensitivity in graph learning architectures and computational cost from graph construction and joint tuning of the LLM and encoder. Additional limitations implied by the text include dependence on LLM-generated QA pairs, curated rather than exhaustive protein coverage, and CPTAC-derived cohort representativeness.

The data governance statement is unusually clear. Data come from publicly available CPTAC resources via the NCI; original CPTAC collection had IRB approval and patient consent; data are de-identified under HIPAA guidelines; the secondary computational analysis did not require additional IRB approval; and the work adheres to NCI data-use agreements and access policies (Adam et al., 26 Sep 2025).

CPTAC-PROTSTRUCT should also be distinguished from several other CPTAC-linked resources. Image-based CPTAC benchmarks such as STAS_CPTAC in lung adenocarcinoma are digital pathology datasets and do not perform proteomic or protein-structure-linked analysis (Pan et al., 18 Mar 2025). Likewise, held-out CPTAC whole-slide image benchmarks for cancer-type classification do not define or mention a resource called CPTAC-PROTSTRUCT (Weihrauch et al., 10 Jun 2026). External CPTAC validation cohorts for slide-level representation learning are pathology benchmarking resources rather than proteomics instruction-tuning datasets, and they do not mention CPTAC-PROTSTRUCT specifically (Lenz et al., 2024). By contrast, network-guided penalized regression on CPTAC-HNSCC proteomics is relevant as a network-based statistical framework for hub-aware modeling, but it contains no explicit protein structure analysis and serves primarily as a network-guided prioritization method rather than an instruction-tuned graph-LLM resource (Ahn et al., 29 May 2025).

Taken together, these distinctions clarify the identity of CPTAC-PROTSTRUCT. It is not a pathology benchmark, not a whole-slide image cohort, and not a general CPTAC multi-omics label release. It is a proteomics-centered instruction-tuning dataset, paired with a patient-specific graph-LLM architecture, intended to align molecular network states with natural-language clinical reasoning in oncology (Adam et al., 26 Sep 2025).