ATOMIC trial | Counter Abstract: The HR That Moves When You Confirm the Biomarker
The Source Report | Oncology AI Lab | Allen Li, MD
The Source Report Video of ATOMIC trial
ATOMIC is the trial the adjuvant dMMR colon cancer field has been waiting for. First phase 3 randomized data testing immunotherapy after surgery. HR 0.50 for DFS. NCI-sponsored. Three hundred community sites. Already in the NCCN guidelines. The result is real, and it changes practice. But the magnitude of the number depends on what you know about the denominator.
3 Things the Abstract Does Not Tell You
1. The NNT to prevent one DFS event at 3 years is 10. The absolute risk reduction is 10.1 percentage points.
2. Central confirmation of dMMR showed 14% of control-arm patients were not confirmed dMMR, versus 8.7% in the atezolizumab arm. That imbalance dilutes the control arm. The HR for DFS among centrally confirmed dMMR patients is 0.55.
3. Thirty percent of control patients received 3 months or less of FOLFOX, the duration IDEA showed is inferior for FOLFOX, compared with 22% in the atezolizumab arm. Control patients who stopped chemo early had no further treatment. Atezolizumab patients who stopped early still had months of immunotherapy ahead.
Counter Abstract
Background: The ATOMIC trial asks whether adding a PD-L1 inhibitor to standard adjuvant chemotherapy improves disease-free survival in resected stage III dMMR colon cancer. The design answers a specific question: whether atezolizumab plus 6 months of mFOLFOX6, followed by 6 months of atezolizumab monotherapy, improves DFS over 6 months of mFOLFOX6 alone. It does not test atezolizumab without chemotherapy, and it does not address the 3-month chemotherapy duration that many community oncologists now use for low-risk stage III disease based on IDEA data.
Methods: 712 patients with resected stage III dMMR colon cancer (determined by local IHC) were randomized 1:1 to atezolizumab plus mFOLFOX6 for 6 months followed by atezolizumab monotherapy for 6 months (12 months total) versus mFOLFOX6 alone for 6 months. The control arm was the US standard of care at both design and publication. The trial was open-label. The primary endpoint was DFS, pre-specified in the protocol. DFS is a validated surrogate for OS in adjuvant colon cancer with chemotherapy regimens. The trial was NCI-sponsored with Genentech as pharmaceutical collaborator. Central confirmation of dMMR was retrospective, and 11.5% of enrolled patients were not centrally confirmed as dMMR, with an imbalance between arms: 8.7% not confirmed in the atezolizumab arm versus 14.3% in the control arm.
Results: At a median follow-up of 40.9 months, 3-year DFS was 86.3% in the atezolizumab arm versus 76.2% in the control arm (HR 0.50, 95% CI 0.35 to 0.73, P<0.001). The absolute risk reduction (ARR) at 3 years is 10.1 percentage points. The number needed to treat (NNT) to prevent one DFS event at 3 years is 10. The trial was stopped at the second pre-specified interim analysis after 127 of 165 planned events, crossing the efficacy boundary. Among centrally confirmed dMMR patients, the HR for DFS is 0.55 (95% CI 0.38 to 0.82).
A pre-specified subgroup analysis by chemotherapy duration showed that patients receiving more than 6 cycles (more than 3 months) of mFOLFOX6 had benefit (HR 0.41, 0.27 to 0.64), while patients receiving 6 or fewer cycles (3 months or less) showed no benefit (HR 0.97, 0.44 to 2.11). Thirty percent of control patients received 3 months or less of mFOLFOX6, the duration the IDEA trial showed is inferior for FOLFOX, compared with 22% of atezolizumab patients. Control patients who stopped chemotherapy early had no further treatment. Atezolizumab patients who stopped chemotherapy early continued on immunotherapy monotherapy.
Grade 3 or 4 adverse events occurred in 84.1% of the atezolizumab arm versus 71.9% of the control. Six grade 5 events occurred in the atezolizumab arm versus two in the control; two were treatment-related. OS has not reached significance (HR 0.90, 95% CI 0.55 to 1.47, 64 events).
Conclusion: ATOMIC demonstrates a clinically and statistically significant DFS improvement with the addition of atezolizumab to adjuvant mFOLFOX6 in dMMR stage III colon cancer. The NNT of 10 at 3 years is favorable and justifies practice change for patients who are candidates for immunotherapy. Two structural features merit consideration when interpreting the magnitude of benefit: a 5.6-percentage-point imbalance in non-centrally-confirmed dMMR patients between arms (more non-dMMR patients in the control arm), and an 8-percentage-point imbalance in patients receiving 3 months or less of chemotherapy (more abbreviated chemo in the control arm with no treatment safety net, versus ongoing immunotherapy in the experimental arm). Neither invalidates the result, but it is worth noting that the HR for DFS among centrally confirmed dMMR patients is 0.55.
Practice Landscape: Atezolizumab is not currently FDA-approved for adjuvant colon cancer. An FDA submission based on these data is anticipated. NCCN guidelines have already incorporated the ATOMIC results and extended them to T4bN0 stage II dMMR tumors. In the neoadjuvant space, the NICHE-2 study (nivolumab plus ipilimumab, 4 weeks preoperative, single-arm phase 2) showed 95% major pathologic response and 100% 3-year DFS in dMMR colon cancer. The phase 3 AZUR-2 trial comparing perioperative dostarlimab monotherapy to adjuvant chemotherapy in dMMR colon cancer is enrolling. ATOMIC establishes the first randomized adjuvant evidence. Head-to-head comparisons with neoadjuvant strategies do not yet exist.
How the AI Read This Trial
The structured appraisal correctly extracted the NNT, identified the clean statistical design, flagged the OS confounding from post-recurrence immunotherapy crossover, and pulled the centrally confirmed sensitivity analysis showing the attenuated HR of 0.55. It scored the trial accurately as well-designed with a genuine result.
It missed two things I caught. First, it treated the 11.5% central dMMR discordance as an aggregate population finding without examining the arm-level split. The 5.6-point imbalance between arms, with the control carrying more non-confirmed patients, creates a directional bias the AI did not name. Second, it identified the chemotherapy-duration subgroup interaction but did not connect it to the IDEA trial data showing 3 months of FOLFOX is inferior. That connection reframes the subgroup from a biological curiosity to a structural design feature: the asymmetric treatment duration creates a safety net in the experimental arm that the control arm does not have.
The AI also framed immature OS as a trial limitation when, in the curative adjuvant setting, DFS is the clinically meaningful endpoint and does not require OS confirmation to justify practice change. That is a framing judgment a practicing oncologist makes automatically. The structured checklist does not.
The Evidence Literacy Point
ATOMIC is a strong trial with a real result. The practice recommendation is clear: add atezolizumab for your dMMR stage III colon cancer patients who are candidates for immunotherapy. What the abstract compression removes is the calibration. The denominator under the hazard ratio matters. It always does.
The Source Report: Counter Abstract Methods
This appraisal was produced using The Source Report, a structured AI-assisted critical appraisal methodology developed at the Oncology AI Lab. The primary publication, supplemental appendix, and statistical analysis plan were analyzed by Claude (Anthropic) operating under a versioned skill file (v7.4) that executes a mandatory 23-step checklist covering control-arm adequacy, endpoint integrity, subgroup analysis, post-progression therapy, informative censoring, abstract framing, and trial quality scoring. The AI extracts absolute and relative benefit measures, calculates number needed to treat (NNT) and ESMO-MCBS grade, and identifies findings present in the data but absent from the abstract (delta). All findings are verified against the source documents through a pre-output verification gate before any content is produced. The analysis runs in two phases with a mandatory expert checkpoint between them. At checkpoint, the physician (A.L.) reviews all AI outputs, corrects errors in framing and interpretation, selects the primary delta, and provides the clinical bottom line.
Disclaimer: This content is for medical education and evidence literacy purposes only. It is not medical advice. Treatment decisions should be made by the treating physician in consultation with the patient based on individual clinical circumstances. The Source Report is an independent educational series. It is not affiliated with the trial sponsors, investigators, or any pharmaceutical company. The authors have no financial relationships with Genentech, Roche, or the Alliance for Clinical Trials in Oncology.
The chemotherapy-duration subgroup analysis discussed in this appraisal was pre-specified but exploratory. It was not powered for independent hypothesis testing and should not be used in isolation to guide treatment decisions. The centrally confirmed dMMR analysis is a sensitivity analysis, not the primary result.
This appraisal used AI assistance (Claude, Anthropic) for data extraction and structured analysis. All AI outputs were reviewed and corrected by a board-certified medical oncologist. The AI does not make clinical recommendations.