A cfDNA fragmentomics-based machine learning model for predicting treatment response and recurrence in locally advanced rectal cancer after neoadjuvant chemoradiotherapy

Abstract

Neoadjuvant chemoradiotherapy (nCRT) is standard treatment for locally advanced rectal cancer (LARC), but responses vary substantially. We developed Fragmentia-AI Response, a machine learning model that uses cell-free DNA (cfDNA) fragment features to predict pathological response and recurrence risk without relying on somatic mutations. We analyzed 102 patients with LARC treated with nCRT followed by surgery, using 510 longitudinal plasma samples collected at five perioperative time points. The model was trained on post-nCRT (T4) cfDNA from 62 patients to predict pathological complete response (pCR), then evaluated in an internal validation cohort (N = 40) and an independent external cohort (N = 96). Associations with tumor regression grade and recurrence-free survival (RFS) were assessed. The model achieved AUCs of 0.909 (95% confidence interval [CI]: 0.812-1.000) in the training cohort, 0.903 (95% CI: 0.785-1.000) in the internal validation cohort, and 0.835 (95% CI: 0.754-0.916) in the external cohort. Model scores were higher in patients with pCR than in those with non-pCR (Wilcoxon p < 0.001) and were strongly associated with tumor regression grade (Jonckheere-Terpstra test p < 0.001). Higher post-nCRT and post-surgery scores were associated with improved RFS (log-rank p = 0.012 and p = 0.024, respectively). Fragment-level analyses revealed a selective enrichment of 50 – 100 bp cfDNA fragments and a higher proportion of subnucleosomal particles in pCR patients, suggesting a biologically grounded signal underlying model prediction. Fragmentia-AI Response is a mutation-independent cfDNA fragmentomics model that showed reproducible discrimination of pathological response and associations with recurrence risk in LARC, supporting further evaluation as a minimally invasive biomarker for treatment stratification and surveillance.

Graphical abstract

Highlights

• •A cfDNA-based machine learning model predicts pathological response after nCRT.
• •The model shows robust performance across internal and external cohorts.
• •Prediction scores stratify postoperative recurrence risk.
• •Longitudinal ctDNA dynamics reflect treatment response and prognosis.
• •Fragmentomic features provide biological interpretability of model outputs.

1. Introduction

Rectal cancer is a common malignancy of the lower gastrointestinal tract, with a substantial proportion of patients presenting with locally advanced disease at diagnosis [1,2]. Standard treatment for these patients typically involves neoadjuvant chemoradiotherapy (nCRT), which combines intensive chemotherapy regimens with dose-escalated radiotherapy, followed by total mesorectal excision (TME) for local tumor control and survival [[3], [4], [5], [6]]. However, TME carries considerable perioperative risks, including bleeding, infection, anastomotic leak, and postoperative bowel obstruction, as well as long-term functional complications, such as urinary and sexual dysfunction, bowel incontinence, and impaired quality of life (QOL) [[7], [8], [9]]. Notably, a subset of patients achieves pathological complete response (pCR) after nCRT, defined as the absence of viable tumor cells in the resected specimen, which is strongly associated with favorable outcomes [10,11]. These findings have raised the possibility that patients achieving pCR might safely avoid surgery through organ-preserving “watch-and-wait” approaches. Nevertheless, because pCR can only be confirmed postoperatively, opportunities for adaptive treatment strategies are limited. Early identification of patients likely to achieve pCR, as well as those unlikely to respond, could enable personalized management, minimizing surgical morbidity and unnecessary treatment exposure in the former, while allowing timely treatment adaptation in the latter.

Several studies have sought to predict pCR before surgery to guide individualized treatment in patients with locally advanced rectal cancer (LARC). Conventional clinical features, such as carcinoembryonic antigen (CEA) levels and lymphocyte-to-monocyte ratio, have shown some association with treatment response but remain insufficiently accurate for clinical decision-making [12]. Imaging-based radiomics have also shown promise in predicting response, potentially enabling pretreatment stratification and personalized care [[13], [14], [15]]. Tissue and molecular biomarkers, such as gene expression profiles and specific genetic or microbiota signatures, are under investigation for their predictive value, but none are yet robust enough for routine clinical use [[16], [17], [18]]. Beyond clinical and imaging features, circulating tumor DNA (ctDNA) has emerged as a promising biomarker for early cancer detection, monitoring minimal residual disease (MRD), and treatment response prediction. Several studies have shown that ctDNA clearance during or after nCRT correlates with tumor regression and recurrence risk [[19], [20], [21]]. However, these ctDNA analyses typically rely on mutation detection, which may be limited by tumor heterogeneity, sequencing depth, and the low ctDNA abundance in certain tumor types and patients. ctDNA represents the tumor-derived fraction of circulating cell-free DNA (cfDNA), whereas the broader cfDNA pool also contains fragmentation and sequence-context information reflecting chromatin organization and tissue of origin. Such mutation-independent cfDNA features may provide complementary signals when somatic mutations are sparse or tumor-derived DNA abundance is low, although contributions from non-tumor tissues can complicate their interpretation.

Large language models (LLMs) and artificial intelligence (AI) are rapidly transforming nucleotide sequence analysis by extracting complex patterns, enabling advanced genomic annotation, classification, and functional prediction [22,23]. LLMs can capture long-range dependencies and subtle context-dependent features that traditional mutation-based ctDNA analyses may miss. LLMs have been applied to learn informative representations from genomic and epigenomic sequences for tasks such as variant effect prediction, methylation state inference, and cancer type classification. Notably, transformer-based architectures such as DNABERT-2 [24], which uses byte-pair encoding to tokenize nucleotide sequences and multi-head self-attention to capture long-range dependencies, have demonstrated strong performance in modeling DNA sequence patterns. These applications suggest that LLMs could also be leveraged to analyze cfDNA fragmentomic patterns in a mutation-agnostic manner, providing a promising strategy to detect tumor-derived signals and treatment-induced molecular changes in preoperative plasma samples.

Building on this concept, we developed Fragmentia-AI Response, a Transformer-based framework that directly learns predictive representations from cfDNA sequences, independent of mutation calls. This approach captures subtle sequence-based information reflecting tumor-derived signals and treatment-induced molecular changes, offering a complementary and potentially more sensitive strategy for pCR prediction. Our findings demonstrate that Fragmentia-AI Response not only predicts nCRT response but also provides additional information on residual disease and recurrence risk, supporting its further evaluation for individualized patient management.

2. Methods

2.1. Cohort overview and samples

The discovery cohort prospectively enrolled 102 LARC patients from February 2016 to May 2017, all of whom received neoadjuvant chemoradiotherapy (concurrent capecitabine + irinotecan chemotherapy; 50 Gy/25 fractions) followed by total mesorectal excision (TME) and five cycles of adjuvant chemotherapy (capecitabine and oxaliplatin). Eligible patients had histologically confirmed rectal adenocarcinoma staged T3-4 and/or N+, tumors ≤12 cm from the anal verge, no distant metastasis, were aged 18-75 years, and had ECOG 0-1. Patients with prior chemotherapy or immunotherapy were excluded, as were those who were pregnant or breastfeeding, had prior malignancy within 5 years (except basal cell carcinoma or cervical carcinoma in situ), uncontrolled seizures or psychiatric disorders, severe cardiovascular disease, organ transplantation or immunosuppressive therapy, severe uncontrolled infection or comorbidities, or allergy to study drugs.

Pathological tumor regression grade (TRG) was assessed on surgical tumor specimens according to the 8th American Joint Committee on Cancer (AJCC) Staging Manual [25], with pathological complete response (pCR) defined as pT0N0M0 (TRG 0). Plasma samples were collected at five perioperative time points: T1 (baseline, pre-nCRT), T2 (before the 15th fraction of nCRT), T3 (before the 25th fraction of nCRT), T4 (post-nCRT, 0-1 day before surgery), and T5 (post-surgery, 1 week after surgery). An independent cohort of 96 LARC patients receiving the same treatment with available pre-surgical plasma samples was included for external validation. All study procedures were conducted in accordance with the principles of the Declaration of Helsinki and approved by the institutional review board of Nanjing Geneseeq Medical Laboratory (NSJB-MEC-2025-16). All patients provided written informed consent for participation and publication.

2.2. Sample processing, library construction, and sequencing

The plasma fraction of the peripheral blood was isolated within 2 h after specimen collection, which was first centrifuged at 1800 g for 10 min, followed by circulating cell-free DNA extraction and purification using the QIAamp Circulating Nucleic Acid Kit (Qiagen Cat. No. 55114). Extraction of genomic DNA from white blood cells in sediments for each specimen was performed using the DNeasy Blood and Tissue Kit (Qiagen Cat. No. 69504), subsequently utilized as the normal control. Genomic DNA was quantified using the dsDNA HS assay kit on a Qubit 3.0 fluorometer (Life Technologies, US), followed by NGS library construction using the KAPA Hyper Prep kit (KAPA Biosystems). Hybridization-based target enrichment was carried out using xGen lockdown probes targeting 422 cancer-related genes (Nanjing Geneseeq Technology Inc.). Library fragment size was analyzed on a Bioanalyzer 2100 using the High Sensitivity DNA Kit (Agilent Technologies, Santa Clara, CA). Sequencing was performed on the enriched libraries on the Illumina HiSeq4000 platform. The average raw sequencing depth was approximately 200× for blood control samples. Across cfDNA samples, the raw sequencing depth had a median of 3939× (IQR, 3358–5000×; range, 1681–8639×), while the deduplicated sequencing depth had a median of 1962× (IQR, 1676–2344×; range, 1076–3760×).

2.3. Mutation calling and variant allele fraction metrics

FASTQ files were subjected to quality control using Trimmomatic [26], removing low-quality bases (quality <30) and ambiguous N bases at read ends. Reads were aligned to the human reference genome GRCh37/hg19 using modified BWA-MEM (BWA-MEM, v0.7.12, https://github.com/lh3/bwa/tree/master/bwakit) with default parameters [27]. Local realignment around indels was performed using Genome Analysis Toolkit (GATK v3.4.0), followed by base quality score recalibration using GATK v4.1.4. Candidate somatic variants were identified using VarScan2 [28] and were required to meet both a minimum variant allele frequency (VAF) of 0.5% and at least 5 variant-supporting reads, together with a base quality ≥30, support from both strands, strand bias ≤10%, and a somatic p-value ≤0.1. The minimum VAF threshold of 0.5% was established using cfDNA sequencing data from 50 healthy donors, with no variants meeting the mutation-calling criteria at this threshold. The same threshold was subsequently applied to an independent set of 50 additional healthy donors, in whom no variants met these criteria. Variants detected in >10% of a panel of 1000 normal samples or present at >1% population frequency in the 1000 Genomes or ExAC databases were excluded as likely artifacts or common germline variants.

To characterize mutation abundance, we defined three VAF metrics for each cfDNA sample based on all detected somatic variants: median VAF, representing the middle value of variant allele frequencies; mean VAF, representing the arithmetic average of all VAFs; and maximum VAF (max VAF), representing the highest VAF observed in the sample. These metrics were used to assess their predictive value for pCR, either alone or in combination with Fragmentia-AI Response model outputs.

2.4. Fragmentia-AI Response model development

Fragmentia-AI Response is a cfDNA-based predictive model designed to distinguish pCR from non-pCR using post-nCRT (T4) plasma samples, implemented using a deep learning-based multiple-instance learning (MIL) strategy (see Supplementary Methods for details). In brief, BAM files from targeted NGS were preprocessed to extract high-quality paired-end reads with mapping quality ≥30. Fragment start and end coordinates were extended by 6 bp upstream and downstream to account for fragment-end variability, and cfDNA fragments ≤250 bp were retained. Corresponding reference sequences were retrieved from GRCh37 in the reference-genome orientation, irrespective of read-pair alignment strand, and no strand-specific reverse-complement transformation was applied before model input.

Within each sample, individual cfDNA fragments were treated as unlabeled instances, while the sample-level pCR outcome served as the bag-level label. An MIL approach combined with stochastic pool sampling was used to train the model on randomly sampled fragment subsets from each sample, thereby improving robustness to fragment heterogeneity. Fragment sequences were tokenized using Byte Pair Encoding (BPE) [29] and encoded by a transformer-based DNA language model pretrained on large-scale genomic sequences [30], followed by attention-based pooling and fully connected layers to generate sample-level prediction scores [31,32].

The model was trained on T4 cfDNA from the training cohort and validated on a temporally separated internal cohort, with performance further evaluated in an independent external cohort. Training was performed using binary cross-entropy loss with AdamW optimization (learning rate = 1 × 10⁻⁴; weight decay = 1 × 10⁻³). The model achieving the highest validation AUC was selected for downstream analyses.

2.5. Enrichment analysis of ctDNA and motifs

To identify cfDNA fragments most likely contributing to pCR, attention scores and MLP outputs were combined to rank fragments. The top 1% of fragments by combined score were considered contributing signals to pCR.

Motif discovery analysis was performed using HOMER (v4.11). Positive samples comprised the top 1% of fragments identified by our model from pCR cases, while negative samples consisted of randomly selected 1% of fragments from the non-pCR group. For each fragment, 20 bp sequences from the 5′ and 3′ termini were extracted and converted to FASTA format. Motifs of lengths 8 bp, 10 bp, and 12 bp were searched using the positive sequences as input and the negative sequences as background. HOMER calculates the statistical significance of each motif relative to the background sequences, and motifs with p-values < 0.01 were considered putatively associated with pCR, providing a biological interpretation of Fragmentia-AI Response's predictive features.

2.6. Statistical analysis

Statistical analyses were performed using R (v4.5.0) and Python (v3.11.8) (see Supplementary Methods for details). Categorical variables were compared using Chi-square or Fisher's exact tests, and continuous variables were analyzed using the Wilcoxon rank-sum test or Kruskal-Wallis test, as appropriate. Ordinal variables were assessed for trends using the Jonckheere-Terpstra test. Receiver operating characteristic (ROC) analysis was performed with the “pROC” R package, and the area under the curve (AUC) was calculated to evaluate the discriminatory ability of each predictor. Sensitivity and specificity were calculated at predefined thresholds, and corresponding 95% confidence intervals (CIs) were derived using the Wilson score method. Event-time distributions were estimated using Kaplan-Meier methodology with the “survival” and “survminer” R packages, and log-rank tests were used to assess differences in survival distributions. Unless otherwise specified, all statistical tests were two-sided, with p < 0.05 considered statistically significant.

3. Results

3.1. Patient characteristics

The discovery cohort included 102 patients with LARC, with a median age of 58.0 years (range 33-73) and 76.5% male (Table 1). Pathological tumor regression grades 0, 1, 2, and 3 were observed in 26, 12, 51, and 13 patients, respectively, corresponding to a pathological complete response (pCR) rate of 25.5% (26/102). Although all patients were metastasis-free at baseline, the majority presented with T3 tumors (87.3%, 89/102), with similar distributions between pCR and non-pCR groups. Only two patients had no nodal involvement. Most patients were classified as stage IIIB, with no significant differences between response groups. Mesorectal fascia (MRF) positivity and extramural venous invasion (EMVI) were observed in 24.5% and 32.4% of patients, respectively, with comparable frequencies across pCR and non-pCR subgroups. As of the last follow-up, 20 patients experienced recurrence. The estimated 1-year and 2-year recurrence-free survival (RFS) rates were 86.2% and 81.1%, respectively (Fig. S1A). As expected, patients achieving pCR had significantly longer RFS compared with non-pCR patients (log-rank p = 0.025) (Fig. S1B).

Table 1.

Baseline characteristics of patients in the discovery cohort.

	All (N = 102)	Non-pCR (N = 76)	pCR (N = 26)	P-value
Sex				0.114
Female	24 (23.5%)	21 (27.6%)	3 (11.5%)
Male	78 (76.5%)	55 (72.4%)	23 (88.5%)
Age (range)	58.0 [33.0, 73.0]	59.0 [38.0, 73.0]	56.0 [33.0, 73.0]	0.197
TRG				<0.001
0	26 (25.5%)	0 (0%)	26 (100%)
1	12 (11.8%)	12 (15.8%)	0 (0%)
2	51 (50.0%)	51 (67.1%)	0 (0%)
3	13 (12.7%)	13 (17.1%)	0 (0%)
Tumor (T)				0.712
2	2 (2.0%)	1 (1.3%)	1 (3.8%)
3	89 (87.3%)	67 (88.2%)	22 (84.6%)
4	11 (10.8%)	8 (10.5%)	3 (11.5%)
Nodes (N)				0.136
0	2 (2.0%)	1 (1.3%)	1 (3.8%)
1	26 (25.5%)	23 (30.3%)	3 (11.5%)
2	74 (72.5%)	52 (68.4%)	22 (84.6%)
Metastasis (M)				∼
0	102 (100%)	76 (100%)	26 (100%)
Clinical stage				0.566
IIA	2 (2.0%)	1 (1.3%)	1 (3.8%)
IIIA	1 (1.0%)	1 (1.3%)	0 (0%)
IIIB	67 (65.7%)	48 (63.2%)	19 (73.1%)
IIIC	32 (31.4%)	26 (34.2%)	6 (23.1%)
MRF				0.794
Negative	77 (75.5%)	58 (76.3%)	19 (73.1%)
Positive	25 (24.5%)	18 (23.7%)	7 (26.9%)
EMVI				>0.999
Negative	69 (67.6%)	51 (67.1%)	18 (69.2%)
Positive	33 (32.4%)	25 (32.9%)	8 (30.8%)

Abbreviations: TRG, tumor regression grade; pCR, pathological complete response; MRF, mesorectal fascia; EMVI, extramural venous invasion.

The external validation cohort comprised 96 patients, of whom 16 achieved pCR, and 80 were non-pCR following nCRT (Table S1). The median age at diagnosis was similar to that of the discovery cohort. Unlike the discovery cohort, which had a higher proportion of male patients and stage III disease, the external cohort exhibited a relatively balanced distribution in terms of sex and clinical stage (p = 0.01 and p < 0.001, respectively). Nevertheless, no significant differences in clinical characteristics were observed between subgroups.

3.2. Model development and training performance

The discovery cohort was divided into a training set (N = 62) and a temporally separated validation set (N = 40) according to enrollment date. Plasma samples were collected at five perioperative time points for longitudinal monitoring (Fig. 1). Among these, the pre-surgery time point (T4) was selected for model development, as it represents the immediate post-nCRT tumor status most relevant to pathological response assessment. We developed an initial LLM framework for predicting pCR by fine-tuning the pre-trained DNABERT-2 framework, which is a transformer-based model originally designed for nucleotide sequences, employing BPE for sequence tokenization, as well as multi-head self-attention layers and feedforward networks to capture long-range dependencies.

Fig. 1.

Study design

A total of 102 patients with locally advanced rectal cancer (LARC) who received neoadjuvant chemoradiotherapy (nCRT) followed by surgery were enrolled. Serial plasma samples were collected at five time points: T1 (pre-nCRT, baseline), T2 (before the 15th fraction of nCRT), T3 (before the 25th fraction of nCRT), T4 (post-nCRT, pre-surgery), and T5 (post-surgery). cfDNA was extracted and subjected to targeted sequencing using a 422-gene panel. Fragmentia-AI Response, a DNABERT-2-powered large language model, was trained on cfDNA samples collected at T4 to predict pathological complete response (pCR), with the training and validation cohorts determined by enrollment date. An independent cohort of 96 patients receiving the same treatment regimen was used for external validation. The association between LLM-based prediction scores and patient outcomes, including nCRT response, defined by post-surgical pathological assessment, and recurrence-free survival, was assessed using perioperative plasma samples.

To assess model performance across fragment counts, we subsampled different numbers of cfDNA fragments per sample. As shown in Fig. S2A, mapped read counts per sample were on the order of 10⁷. Predictive performance, measured by AUC, generally increased with the number of reads and reached its highest level when all available reads were used (Fig. S2B and C). The “all-reads” configuration was therefore adopted for final inference.

We then tracked model convergence throughout training. As the number of epochs increased, the loss steadily decreased, and the AUC gradually improved until stabilizing around epoch 784 (Fig. S2D and E), indicating effective adaptation of the pre-trained sequence representations to the pCR prediction task. The model selected at epoch 784 was designated Fragmentia-AI Response and was used for subsequent validation analyses.

3.3. Model performance in training and validation cohorts

The model achieved an AUC of 0.909 (95% CI: 0.812-1.000) in the training set and 0.903 (95% CI: 0.785-1.000) in the validation set (Fig. 2A). Prediction scores were significantly higher in patients achieving pCR compared with non-pCR patients (Wilcoxon p < 0.001) and inversely correlated with TRG (Jonckheere-Terpstra test p < 0.001) (Fig. 2B and C). No significant differences in prediction scores were observed across subgroups stratified by sex, age, MRF status, or EMVI (Fig. S3). In contrast, conventional variant allele frequency (VAF)-based metrics derived from mutation profiling showed limited ability to discriminate between pCR and non-pCR patients, yielding substantially lower AUC values compared to Fragmentia-AI Response (Table S2). Integrating VAF features and Fragmentia-AI Response output in a stacked ensemble model resulted in only marginal improvement in predictive performance (Fig. S4), suggesting that Fragmentia-AI Response had already captured the majority of cfDNA-derived molecular signals associated with pathological response to nCRT.

Fig. 2.

Model performance evaluation

(A) Receiver operating characteristic (ROC) curves of Fragmentia-AI Response for predicting pathological complete response (pCR) in the training and validation cohorts. (B, C) Distribution of prediction scores in patient subgroups stratified by pathological response (B) and tumor regression grade (TRG) (C) in the training and validation cohorts. (D) ROC curves of Fragmentia-AI Response for predicting pCR in the external validation cohort. (E) Box plots comparing the distribution of prediction scores between pCR and non-pCR patients in the external cohort. (F) Sensitivity and specificity of Fragmentia-AI Response at the predefined cutoff (0.4939) corresponding to 95% specificity in the training cohort.

To explore model generalizability, we evaluated Fragmentia-AI Response in an independent external cohort of 96 LARC patients. Despite a slightly lower AUC (0.835; 95% CI: 0.754-0.916), likely reflecting differences in cohort composition, the model consistently produced higher prediction scores in pCR patients, supporting its reproducibility across independent datasets (Fig. 2D and E). Using the predefined cutoff corresponding to 95% specificity in the training cohort (0.4939), sensitivities were 0.73 (95% CI: 0.48-0.89), 0.64 (95% CI: 0.35-0.85), and 0.44 (95% CI: 0.23-0.67) in the training, validation, and external cohorts, respectively (Fig. 2F).

3.4. Perioperative dynamics of model prediction scores and associations with patient outcomes

We also evaluated Fragmentia-AI Response using plasma samples collected at other perioperative time points. Interestingly, the model's predictive performance varied across time points and appeared to correlate with their temporal proximity to surgery (Fig. 3A and B; Fig. S5). Specifically, the AUC was lowest at baseline (T1), likely reflecting greater ctDNA variability before treatment initiation. AUCs gradually improved at on-treatment time points, with T3 performing slightly better than T2, and peaked at the pre-surgery sample (T4) used for model development. Post-surgery samples (T5) maintained relatively high predictive performance despite reduced ctDNA abundance after tumor removal, possibly reflecting the persistence of minimal residual disease (MRD)-associated cfDNA signals in non-pCR patients that continued to distinguish response groups.

Fig. 3.

Perioperative prediction scores and their associations with pathological response and patient survival

(A) Area under the curve (AUC) of Fragmentia-AI Response applied to cfDNA samples collected at different perioperative time points. Error bars represent 95% confidence intervals. (B) Distribution of Fragmentia-AI Response prediction scores across time points for patients stratified by pathological response. Gray lines connect longitudinal samples from the same patient, illustrating individual score trajectories. (C, D) Kaplan-Meier curves of recurrence-free survival for patient subgroups stratified by Fragmentia-AI Response score levels using plasma samples collected at T4 (C) and T5 (D). (E) Sankey diagram showing patient distribution across subgroups and recurrence status. (F) Kaplan-Meier curves of recurrence-free survival among patient subgroups defined by T4-T5 changes. Group 1 represents patients with high T4 scores, group 2 includes those with low T4 and high T5 scores, and group 3 comprises patients with low scores at both time points. (G) Swimming plot depicting relapse timelines for patients who experienced recurrence (N = 20).

Furthermore, both high T4 and T5 scores were significantly associated with improved RFS (log-rank p = 0.012 and p = 0.024, respectively) (Fig. 3C and D; Fig. S6). Based on the dynamic changes in scores between these two time points, patients were classified into four categories: high-high (N = 21), high-low (N = 1), low-high (N = 22), and low-low (N = 58) (Fig. 3E). Given the limited number of high-low patients, we combined them with the high-high group as “Group 1”, while defining low-high as “Group 2” and low-low as “Group 3”. RFS differed among the three groups (log-rank p = 0.025), with Group 1 showing the most favorable outcomes, Group 2 showing intermediate outcomes, and Group 3, characterized by persistently low scores, exhibiting the least favorable outcomes (Fig. 3F).

Importantly, all patients who eventually relapsed had low T4 scores, consistent with early detection of recurrence risk, whereas approximately 20% of relapsed patients exhibited high T5 scores but experienced later relapse, underscoring the prognostic value of temporal score dynamics (Fig. 3G). Collectively, these findings suggest that Fragmentia-AI Response captures both immediate treatment response and residual disease signals, providing additional prognostic information for patient risk stratification.

3.5. Model interpretability and biological insights

We next sought to explore the biological features underlying the predictive scores generated by our model. To identify cfDNA fragments most likely contributing to pCR, fragments were ranked by combined Fragmentia-AI Response attention scores and multilayer perceptron outputs. The top 1% of fragments were defined as response-associated and compared with a randomly selected 1% of fragments in both pCR and non-pCR groups. Analysis of fragment length revealed that the top 1% fragments in both groups were predominantly 50-100 bp, markedly shorter than the canonical ctDNA fragment size of ∼166-167 bp observed in the random fragments (Fig. 4A; Fig. S7). Considering the potential involvement of subnucleosomal structures, we analyzed the proportion of subnucleosomal-sized particles (subNSP) across pCR and non-pCR groups. The subNSP fraction was significantly higher in pCR samples across all three datasets, suggesting that these fragments likely contribute to the model's predictive signal (Fig. 4B). GC content analysis showed that the top 1% fragments exhibited a peak around 40%, whereas random fragments displayed a bimodal distribution with an additional minor peak near 60% (Fig. 4C; Fig. S8). These findings indicate that cfDNA fragments contributing to high predictive scores are enriched for short, low-GC sequences, consistent with nucleosome-depleted regions.

Fig. 4.

Characterization of cfDNA fragments contributing to high predictive scores

(A) Length distribution of cfDNA fragments, comparing the top 1% fragments ranked by model scores with a randomly selected 1% of fragments. (B) Proportion of subnucleosomal-sized particles (subNSP) between pCR and non-pCR patients across datasets. (C) GC content distribution of top 1% versus random 1% fragments. (D, E) Motif enrichment analysis of response-associated sequences. Target sequences were the top 1% fragments from pCR cases, and background sequences were randomly selected 1% fragments from non-pCR cases. Motifs of 8 bp, 10 bp, and 12 bp that were significantly enriched (p < 0.01) in upstream (D) and downstream (E) regions flanking each fragment are shown.

To further characterize sequence determinants associated with high predictive scores, motif discovery analysis was performed using HOMER on 20-bp sequences from the 5′ and 3’ termini of the top 1% fragments from pCR cases, with randomly selected 1% fragments from the non-pCR group serving as background controls. Several motifs were significantly enriched in the top 1% fragments relative to the background (p < 0.01), both upstream and downstream of fragment ends (Fig. 4D and E), providing a biological interpretation of the predictive features captured by Fragmentia-AI Response. Collectively, these analyses indicate that high predictive scores for pCR are associated with a distinctive subset of cfDNA fragments that are shorter, have characteristic GC content, display subnucleosomal protection patterns, and are enriched for specific sequence motifs. These features suggest a mechanistic link between chromatin organization, fragmentomic signatures, and therapeutic response.

4. Discussion

In this study, we developed and validated Fragmentia-AI Response, an LLM-based framework that directly learns predictive representations from cfDNA sequences to enable minimally invasive prediction of response to nCRT in LARC. Trained on preoperative plasma, Fragmentia-AI Response demonstrated robust discrimination of pCR in internal training and validation cohorts and outperformed conventional VAF-based mutation metrics, with integration of VAF features providing only marginal gains. Importantly, T4 scores were also associated with recurrence-free survival, with lower scores indicating higher biological treatment resistance and greater relapse risk. Applied to an independent external cohort, Fragmentia-AI Response retained discriminatory ability, supporting the feasibility of leveraging high-dimensional cfDNA sequence and fragmentomic patterns for both response prediction and recurrence risk stratification.

A key advantage of Fragmentia-AI Response lies in its ability to exploit nucleotide-level and fragmentomic patterns without relying on prior mutation calls. By fine-tuning a transformer architecture, the model learns complex, context-dependent features that correlate with tumor-derived cfDNA and treatment-induced molecular changes, while circumventing limitations of mutation-based assays such as tumor heterogeneity and low sequencing depth. Our interpretability analyses revealed that the cfDNA fragments most strongly associated with high predictive scores are predominantly short (50-100 bp), enriched for low-GC content, exhibit subnucleosomal protection patterns, and harbor specific sequence motifs. These features likely reflect not only the intrinsic chromatin organization of the tumor but also the DNA damage and chromatin remodeling induced by radiotherapy and chemotherapy. Radiochemotherapy generates DNA strand breaks and induces cell death (e.g., apoptosis), preferentially exposing linker DNA or nucleosome-depleted regions that are more vulnerable to nuclease digestion. The resulting short fragments are therefore enriched from these open chromatin regions, while motif enrichment may highlight regulatory sites undergoing treatment-induced changes in binding or accessibility. Consistent with this, prior studies have shown that nucleosome footprints in cfDNA correlate with chemotherapy response. For example, Yang et al. demonstrated that pre-treatment nucleosome positioning signatures in cfDNA at transcription start sites differed between responders and non-responders to neoadjuvant chemotherapy in breast cancer [33]. Together, these findings provide a mechanistic rationale for how Fragmentia-AI Response's learned fragmentomic features may capture treatment-associated chromatin alterations, linking fragment patterns to therapeutic response. The consistent outperformance over VAF-based metrics across internal cohorts further underscores the added value of leveraging fragmentomic and sequence-level information. Moreover, the strong association between Fragmentia-AI Response scores and TRG indicates that cfDNA fragmentomic features serve as sensitive molecular correlates of treatment efficacy.

The temporal dynamics of Fragmentia-AI Response scores provide additional insights into tumor response and residual disease across perioperative time points. Predictive performance was lowest at T1, likely reflecting low tumor-derived cfDNA contribution and higher pre-treatment variability. Scores at T2 and T3 showed modest predictive ability, suggesting that even without specific training on these time points, Fragmentia-AI Response can partially capture early signals of tumor response and help identify non-responders during nCRT, which could inform adaptive treatment strategies. Performance peaked at T4, where cfDNA most directly represents post-treatment tumor status, while postoperative (T5) samples maintained moderately high predictive ability despite substantial reduction in ctDNA levels, providing information on minimal residual disease and potential relapse risk. Regarding recurrence-free survival, T4 scores showed slightly stronger associations with outcomes than T5, as cfDNA at T4 reflects residual tumor activity after nCRT, with low scores indicating biological treatment resistance and a higher likelihood of relapse. T5 ctDNA levels are near the detection limit following tumor resection, and while they can still detect residual disease in non-pCR patients, the prognostic signal is somewhat attenuated. Factors contributing to these differences include training on T4 data, higher tumor-derived cfDNA abundance at T4, dilution and biological noise at T5 due to surgery, and transient false negatives or delayed ctDNA release postoperatively. Collectively, these observations explain the dynamic AUC patterns and highlight the potential clinical value of applying a T4-trained model across multiple perioperative time points to inform early response assessment, preoperative decision-making, and postoperative surveillance.

Fragmentia-AI Response represents a new class of cfDNA-based biomarkers that exploit high-dimensional fragmentomic and sequence information from NGS data. Because this approach does not depend on tumor-informed mutation panels, it may be applicable across patients with heterogeneous mutational landscapes or low ctDNA abundance, circumstances in which conventional mutation-based assays often fail. Clinically, early identification of patients likely to achieve pCR could support organ-preserving strategies, thereby sparing selected patients from the morbidity of major surgery and improving quality of life. Indeed, studies have suggested that organ-preservation strategies in LARC patients with complete response can achieve outcomes comparable to surgery, with fewer functional complications and better QOL [[34], [35], [36], [37]]. In this context, the predefined cutoff of 0.4939 was intentionally selected in the training cohort to prioritize specificity, reflecting its intended use as a high-specificity component of multimodal pCR assessment rather than as a standalone classifier. This operating strategy was designed to minimize false classification of patients with residual disease as having achieved pCR, which would be particularly consequential if used to inform organ-preservation decisions. Sensitivity decreased from 0.73 in the training cohort to 0.64 in the internal validation cohort and 0.44 in the external cohort, while specificity in the external cohort was 0.89, indicating that the operating characteristics of the fixed training-derived threshold were not fully maintained across cohorts. These findings define the current clinical positioning of Fragmentia-AI Response: the predefined cutoff may provide complementary information for pCR assessment, but its external-cohort sensitivity does not support standalone use for treatment decisions such as omission of surgery. Prospective multicenter validation and threshold calibration will therefore be important to improve generalizability while preserving the intended high-specificity operating profile. Conversely, recognizing non-responders before surgery may facilitate timely treatment adaptation, such as intensification of neoadjuvant regimens, earlier surgical intervention, or enrollment into clinical trials of novel agents. Furthermore, monitoring model prediction scores at perioperative time points may help stratify relapse risk and tailor postoperative surveillance or adjuvant treatment intensity.

Several limitations should be acknowledged. The sample size is modest, and further validation in larger, prospective multicenter cohorts will be required to confirm generalizability. Additionally, although DNABERT-2-based tokenization effectively captures sequence-level patterns, the model remains a “black box,” and the biological underpinnings of its learned representations require further exploration. In particular, the mechanistic validity of the learned fragmentomic features, especially those suggestive of subnucleosomal protection, requires experimental substantiation, such as mapping cfDNA footprints to nucleosome occupancy or chromatin accessibility profiles in matched tumor tissues. Finally, extending this framework beyond fragmentomics to integrate complementary features, such as methylation profiles, fragmentation endpoints, and chromatin accessibility, may further enhance predictive power and biological insight.

5. Conclusion

In summary, Fragmentia-AI Response represents a proof-of-concept demonstration that transformer-based LLMs can decode cfDNA sequence patterns to predict therapeutic response and residual disease in rectal cancer. This mutation-agnostic, data-driven framework expands the scope of cfDNA-based precision oncology and provides a foundation for developing generalizable, real-time predictors of treatment efficacy across cancer types.

Ethics approval and consent to participate

The study was approved by the Ethics Committee of Nanjing Geneseeq Medical Laboratory (NSJB-MEC-2025-16). Written informed consent was obtained from each patient before sample collection.

Consent for publication

Not applicable.

Availability of data and materials

The code supporting the model analysis and figure generation is publicly available on GitHub (https://github.com/cancer-oncogenomics/Fragmentia-AI-Response). Other data used and/or analyzed during the current study are available from the corresponding author upon reasonable request, subject to applicable institutional and ethical requirements.

Author's contributions

K.M.: Conceptualization, formal analysis, investigation, writing – original draft; S.W.: Conceptualization, formal analysis, investigation, writing – original draft; G.S.: Formal analysis, investigation, writing – review & editing; H.B.: Investigation, writing – review & editing; Y.S.: Resources, supervision, writing – review & editing; Z.Z.: Resources, supervision, funding acquisition, writing – review & editing.

Funding

This study was supported by the Natural Science Foundation of Hunan Province (No. 2024JJ5568). The funders had no role in the study design, data collection and analysis, or the decision to publish this manuscript.

Declaration of competing interest

S.W., G.S., H.B. and Y.S. are employees of Nanjing Geneseeq Technology Inc. The remaining authors declare no competing interests.

Appendix A. Supplementary data

The following is the Supplementary data to this article: