2.1. Study Design and Participants

The Transcriptomic Profiles of Cancer-Related Fatigue Cross-Sectional (TRIXIE) study was conducted at the University of Maryland Marlene and Stewart Greenebaum Comprehensive Cancer Center in Baltimore, Maryland, between October 2022 and December 2023. The primary aim was to investigate the transcriptomic profile of patients with cancer experiencing high vs. low levels of fatigue during chemotherapy treatment.

Participants were eligible if they had a diagnosis of breast, gastrointestinal, or gynecological cancer, stage 0, 1, 2, or 3; were assigned female sex at birth; had at least 4 weeks and 2 cycles of chemotherapy and at least one scheduled session remaining; could communicate in English; and were at least 18 years old. This study was approved by the University of Maryland Institutional Review Board (HP-00100268). All participants provided written informed consent.

2.2. Demographics, Fatigue Measures, and Functional Assessments

Demographics were obtained from a structured questionnaire and clinical characteristics from medical records.

Questionnaires were completed either the same day as the blood draw and functional assessments or within the previous 5 days. Patient-reported fatigue was assessed using the Functional Assessment of Chronic Illness Therapy-Fatigue (FACIT-F) [28], Brief Fatigue Inventory (BFI) [29], and a Symptom Inventory. The FACIT-F is a 40-item, validated measure that comprises five subscales: physical wellbeing, social wellbeing, emotional wellbeing, functional wellbeing, and a fatigue subscale [28]. Participants responded with how true various statements were over the last 7 days such as “I feel fatigued” and “I have to limit my social activities because I am tired” with five response choices ranging from 0, “Not at all,” to 4, “Very much.” Scoring yields five subscale scores, a total score, a general quality of life score (Functional Assessment of Cancer Therapy-General [FACT-G]; physical + social + emotional + functional wellbeing), and a trial outcome index (TOI; physical wellbeing + functional wellbeing + fatigue subscale). A higher score indicates higher wellbeing/quality of life or less fatigue. The BFI is a 10-item questionnaire that assesses fatigue now, usual, and worst fatigue in the last 24 h and how much fatigue has interfered with general activity, mood, etc. [29]. The item scales range from 0, “No fatigue” or “Does not interfere,” to 10, “As bad as you can imagine” or “Completely interferes.” The average of all the items yields a global fatigue score with a higher score indicating worse fatigue. Cronbach alpha reliability ranges from 0.82 to 0.97 [29]. Lastly, there were four questions related to fatigue as part of a Symptom Inventory—fatigue, weakness, drowsiness, and sleep problems. Participants rated the symptoms at their worst in the past seven days from 0, “Not present,” to 10, “As bad as you can imagine.”

Functional assessments were performed the same day as the blood draw after a snack was offered. Grip strength and muscle fatigability were assessed using handgrip dynamometry (handheld Hydraulic Hand Dynamometer, Model 5030J1, Jamar Technologies, Hatfield, PA, USA). Maximal voluntary isometric contraction (MVIC) was assessed on both the left and right arms using the protocol described in Mustian et al. [30]. In brief, the participant stood with their elbow joint at a constant 90° angle and the instrument was positioned so the palm fit comfortably in the rear of the instrument and the fingers curled around the adjustable grip. The participants were asked to squeeze as hard as possible and hold for 5 s. Several practice squeezes were performed, after which three trials were performed with each hand, alternating, with at least 30 s between trials. The measurement from the peak-hold needle was recorded. The average of the three trials was used in the analysis.

At least 1 min after the MVIC test, muscle fatigability was assessed via the static fatigue index using the same dynamometer [31]. For this trial, the participant squeezed the dynamometer maximally and held it for 30 s. Maximal force was recorded, as well as force at 5, 10, 15, 20, 25, and 30 s. The area under the curve was calculated (equal to AUC2), as well as the maximum possible AUC based on the maximal force (AUC_max). Static fatigue index was calculated as 100% × (1 − AUC2/AUC_max) (Supporting Figure 1). A greater score indicates less fatigability.

The 30-s chair stand test is a measure of leg endurance and muscular endurance [32]. We measured the number of times the participant could rise from a sitting position in a standard chair in 30 s without assistance, holding onto the chair or other objects, or using their arms.

2.3. Blood Collection and Complete Blood Count (CBC)

A fasted blood draw was performed in conjunction with clinical labs approximately 1-2 h before the participant’s chemotherapy appointment. The clinical labs included a CBC of hemoglobin, red blood cells (RBCs), white blood cells (WBCs), platelets, neutrophils, lymphocytes, and monocytes; these values were obtained from the medical record. For RNA-Seq, venous blood (3 mL) was collected into a Tempus tube (ThemoFisher Scientific, Waltham, MA, USA) containing 6 mL of stabilizing reagent, which inactivates cellular RNases and selectively precipitates RNA. The tubes were placed at −80°C within 2 h and stored there until processing.

2.4. RNA-Seq

RNA was extracted and sequenced at the Genomics Core Facility, University of Maryland School of Medicine (UMSOM), using the Tempus Spin RNA Isolation Reagent Kit (ThermoFisher). Extracted RNA was run on Bioanalyzer gels to obtain the RNA integrity number (RIN; all were ≥ 7.4 on the quality scale).

At Maryland Genomics, part of the Institute of Genome Sciences, libraries were prepared from 25 ng of RNA using the TruSeq RNA Sample Prep kit (Illumina, San Diego, CA, USA) according to the manufacturer’s instructions, with an additional PCR cycle. Samples were sequenced on an Illumina NovaSeq6000 with a 150 base pair (bp) paired-end read configuration. Sequences were checked for overall quality using FastQC [33], and low-quality reads, duplicates, and adapter surplus were trimmed using Trimmomatic [34]. Filtered reads were aligned to the human reference Ensembl Homo sapiens.GRCm38.96 using HiSat (Version HISAT2-2.0.4) [35], and the number of reads by gene was determined using HTSeq [36]. Overall, 29/30 samples met quality control criteria and were used in further analyses.

2.5. Statistical Analysis

Demographics, clinical characteristics, fatigue, and functional measures were compared between the two groups using a t-test allowing unequal variances (Microsoft Excel, Redmond, WA, USA, or JMP, SAS Institute, Cary, NC, USA).

For DGE analysis, gene counts were normalized to the log₂ scale, and genes with low mean normalized counts were removed from the analysis. Normalized and filtered gene counts were utilized in the principal component analysis (PCA), which enabled the identification of any outliers and clusters of samples by demographic and/or clinical data. The variables that showed associations with the first two principal components (PCs) were used as covariates in the downstream DGE analysis. To detect genes associated with fatigue, we used a dual approach; first, following the criteria applied by Eek et al. [37], we set a fatigue cutoff of 30 on the FACIT-F fatigue subscale to dichotomize samples into high (> 30, n = 14) and low (≤ 30, n = 15) fatigue and compared their gene expression. In parallel, we were interested in evaluating gene expression changes by changes in fatigue score units. While using fatigue as a continuous variable is less useful for clinical applications, it is statistically more powerful and avoids the establishment of arbitrary cutoffs that could alter the biological interpretation of the analysis results. Both analyses were conducted using DESeq2, which models gene expression data using the negative binomial distribution [38]. We applied the likelihood ratio test and computed p values as the difference in deviance between the full models (including cancer stage, WBC, neutrophils, lymphocytes, RBC, platelet, and fatigue score or group [high/low]) and reduced models (including cancer stage, WBC, neutrophils, lymphocytes, RBC, and platelets). We considered DEGs statistically significant if their Benjamini–Hochberg adjusted p value was equal to or lower than 0.1.

We utilized CIBERSORTx [39–41] to infer whole blood cell-type fractions from RNA-seq data. To this aim, we first normalized the gene counts as counts per million (CPM) and used a signature gene expression matrix (LM22) implemented in the software that cell-type specific markers. To confirm the accuracy of the fractions of each blood cell type estimated using the CIBERSORTx, we assessed for correlations for neutrophil and monocyte concentrations between those estimated with CIBERSORTx and those that were measured clinically. We found strong correlations (neutrophils R² = 0.85 and monocytes R² = 0.56, Supporting Figure 2).

For pathway analysis, we ranked all genes analyzed for differential expression by the product of their log₂fold-change (log₂FC) and p values. We tested them for enrichment in biological processes/pathways using Gene Ontology (GO)’s biological processes functionally related to cancer-related fatigue. As gene sets for the analysis, we used the collection gene set C5 (Version 2023) that derives from the GO resource [42] and contains biological processes, cellular components, and molecular functions in addition to gene sets derived from the Human Phenotype Ontology (HPO) [43]. The enrichment score (ES) and its significance were computed with the multilevel algorithm implemented in the package fast gene set enrichment analysis (fgsea) [44], which performs permutations to adjust for multiple testing. In addition to the p value and adjusted p value, fgsea output includes an ES whose sign (positive or negative) indicates whether the larger part of the genes in the considered gene set is up- or downregulated. In addition, fgsea highlights the leading genes, which have the largest difference in expression and give the strongest contribution to the gene set. Subsequently, we conducted sensitivity analysis using the “leave one out” method [45]. We considered the DEG finding robust when significance was retained for > 90% of the trials.

3.1. Sample Description

A total of 68 women were approached; 38 did not enroll (18 declined and 20 were not eligible [10 had stage 4 disease and/or metastases, 2 were receiving chemotherapy at a different hospital, 4 did not speak English, and 4 did not have any more chemotherapy scheduled]) and 30 women enrolled between November 2022 and November 2023 (Table 1). Participants were 53.4 ± 13.5 years old. Approximately half identified as Black and/or African American and half identified as White. Participants were on average overweight (BMI: 25.8 ± 7.2 kg/m²), consistent with the American population. Most (96.7%) had breast cancer and one had uterine cancer. Participants were on a variety of chemotherapy regimens.

3.2. Fatigue and Functional Assessments

Participants experienced a wide range of self-reported fatigue as measured by the FACIT-F, BFI, and single-item questions (Table 2). Participants were dichotomized according to fatigue level (≤ 30 for low vs. > 30 for high on the FACIT-F fatigue subscale [37]) for some subsequent analyses; there were no statistically significant differences between these two groups in regard to demographics or clinical characteristics (Table 1). Interestingly, while those in the “low fatigue” group had less fatigue and/or greater wellbeing on all scales, statistical significance was reached only for physical wellbeing, functional wellbeing, the BFI total score, and the single-item questions for fatigue, weakness, and drowsiness, and not social wellbeing, emotional wellbeing, or sleep problems. In regard to physical assessments, those with low fatigue exhibited higher grip strength and lower fatigability, though differences did not reach statistical significance.

3.3. Transcriptome Analyses

We obtained an average of > 106 million bp paired ends reads per sample, and 96.62% of them matched properly to the reference, guaranteeing an average of 10× coverage of the transcriptome (Supporting Table 1). Of the mapped reads, 93.07% aligned to exons, 5.94% to introns, and 0.99% to intergenic regions.

While we did not observe any outliers, PCA analysis enabled us to detect significant clustering of the first two components (PC1 and PC2) of expressions of transcripts by cancer stage, overall WBC counts, and singular neutrophils and lymphocyte counts. In addition, PC1 was significantly correlated with RBCs and platelet counts (Supporting Figure 3). These variables were used as covariates in the DEG analyses.

Comparison of gene expression between the high and low fatigue groups highlighted three DEGs: SH3 domain containing ring finger 1 (SH3RF1) alias POSH (Log₂FC = −1.04, FDR = 4.08e− 03), caprin family member 2 (CAPRIN2, Log₂FC = 0.42, FDR = 8.23e− 03), and baculoviral IAP repeat containing 2 (BIRC2, Log₂FC = 0.97, FDR = 2.42e− 02) (Figure 1, Supporting Table 2, Supporting Figure 4). With higher fatigue coded as a continuous score, SH3RF1 again was found to be significantly downregulated and CAPRIN2 upregulated with higher fatigue, and there was significant perturbation of eight other genes: citrate synthase (CS), family with sequence similarity 153 member A (FAM153A), mitochondrial 54S ribosomal protein MRPL51 (MRPL51), MAPK activated protein kinase 3 (MAPKAPK3), VPS18 core subunit of CORVET and HOPS complexes (VPS18), zinc finger protein 731 (ZNF731P), protein phosphatase 2 scaffold subunit alpha (PPP2R1A), and coiled-coil-helix-coiled-coil-helix domain containing 2 (CHCHD2) (Supporting Table 3). A sensitivity analysis was performed using the “leave one out” method [45]; key genes identified using linear regression were confirmed to be robust to potential outliers (i.e., SH3RF1, CAPRIN2, and BIRC2, Supporting Table 4).

All genes were then ranked in decreasing order by the product of the log₂FC and significance and tested their enrichment in GO (biological processes, cellular components, and molecular functions) and HPO gene sets using fgsea [43] (Table 3). Twenty-four biological processes were significantly enriched, related to mitochondrial function (e.g., aerobic respiration, generation of precursor metabolites, and electron transport chain), ATP synthesis, and fatty acid β-oxidation. In addition, DEGs were primarily enriched in mitochondria and ribosomal cellular components and molecular functions linked with mitochondrial protein complexes and oxidoreductase (Table 3). Interestingly, the ES was negative for all the GO and HPO terms, indicating that most genes included in the gene sets were downregulated in samples with high fatigue compared to low fatigue. Results of the enrichment analysis were similar when we analyzed fatigue as a continuous variable, but the ES increased with increasing fatigue score (Supporting Table 5).

4.1. Strengths and Limitations

This study has several strengths. First, it collected blood from fasted participants and accounted for WBC population within blood, which helps identify metabolic perturbations beyond those in the fed vs. fasted state and differences in populations of inflammatory cells. Second, it recruited a racially diverse population of women undergoing chemotherapy, increasing the generalizability of our results to populations who have been historically underrepresented in research. Third, we used state-of-the-art transcriptomic analysis techniques to identify potential biological and physiological targets to prevent, treat, monitor, and eventually subtype cancer-related fatigue.

However, this study has some limitations. It is a relatively small study (n = 29-30) and, therefore, was underpowered to see differences between high vs. fatigue groups in regard to functional assessments (ES values ranging 0.2–0.4). Also, the population was heterogenous concerning cancer subtypes, cancer stages, and chemotherapy regimens, which might have reduced our ability to resolve differences between groups; our small sample size precluded our ability to adjust for any of these variables. However, our population did not have advanced cancer, which is associated with cachexia, and the muscle wasting syndrome potentially has independent causes of energy dysregulation and fatigue [57]. We assessed and did not find differences between groups with high vs. low fatigue in regard to age, BMI, and other potential confounders, but it is possible that residual confounding exists. Indeed, obesity can be associated with more fatigue [58] and different metabolic transcriptomic profiles [59]. We were able to control for WBC type, which helps to control for inflammatory state. Also, this was a cross-sectional study; fatigue can fluctuate greatly over the course of treatment [12], and there are large differences in physiology between individuals; longitudinal studies have more statistical power to assess changes within people vs. differences across people. In our dichotomous comparisons between those with high vs. low fatigue, it is possible that there was misclassification, for example, if the day or time during which they completed the fatigue questionnaires was not representative of their global fatigue experience. In order to avoid misclassification, we administered several questionnaires that inquired about fatigue over several time periods (e.g., now, in the last 7 days) and confirmed that these values correlated. In addition, we modeled fatigue as a continuous variable, which is more nuanced in fatigue quantification. Hence, these results should be interpreted with prudence and tested for confirmation in independent studies.

4.2. Clinical and Research Implications

Fatigue is a multifactorial experience of physical and mental lack of energy with elusive underlying mechanisms. This study confirms current clinical practice guidelines that fatigue should be discussed frequently between the patient and the provider in order to recognize and manage it as best as possible [60, 61]. Further research is needed to incorporate functional and/or objective measures of fatigue into the clinic to complete patient-reported outcomes and help quantify and characterize it. While patient report is the current gold standard to diagnose and track fatigue, people experiencing moderate–severe fatigue may have reduced focus and concentration, which may reduce the accuracy of questionnaire data, especially for mental fatigue. Functional data are useful, especially in assessing changes over time, because they reflect physical aspects of fatigue that may not be noticeable and, therefore, not reportable. In addition, objective biomarkers such as blood-, saliva-, or urine-based measures should continue to be explored [62, 63]. While direct targeting of mitochondrial function has not yet led to interventions with large effect sizes (e.g., [64]), current recommendations such as exercise [61] and nutrition [54] may work through these pathways and future research may be able to optimize it.