HARMONi-GI1: ivonescimab plus GemCis beats durvalumab plus GemCis on OS in first-line biliary tract cancer (HR 0.75), but it is a single-region Chinese interim analysis with imbalanced subsequent therapy

Data as of 2026-10-26 (Taipei time). [Fact] statements are sourced; [Inference] statements are this article's judgment.

1. One-line conclusion

[Fact] At HARMONi-GI1's prespecified interim analysis, ivonescimab plus gemcitabine/cisplatin vs durvalumab plus gemcitabine/cisplatin gave mOS 14.9 vs 12.1 months (HR 0.75; 95% CI 0.61–0.92; P=0.0031), PFS HR 0.61 and ORR 53.8% vs 32.5% (Akeso press release). [Inference] This is the first Phase 3 trial in BTC to win OS on top of the TOPAZ-1 standard, and China's standard of care may change; but for sponsors, whether the data extrapolate to the West depends on the US EA2251 trial, and this HR should not be used directly to design Western trials (Summit).

For sponsors: comparator assumptions for first-line BTC trials should allow for a PD-1/VEGF bispecific becoming China's new standard; the durvalumab arm's mOS of 12.1 months (CancerNetwork) is the latest control benchmark.

2. What happened

[Fact] HARMONi-GI1 (AK112-309, NCT06591520) is a randomized, double-blind, multicenter Phase 3 trial conducted in China. 682 patients with unresectable or postoperatively recurrent BTC, no prior systemic therapy and ECOG 0–1 were randomized 1:1 to ivonescimab 20 mg/kg q3w (n=340) or durvalumab 1500 mg q3w (n=342), each with GemCis for up to 8 cycles, followed by maintenance; stratification factors were disease status and primary site (intrahepatic, extrahepatic bile duct, gallbladder); the primary endpoint was OS, with PFS and ORR as hierarchically tested key secondary endpoints; median follow-up was 14.3 months (CancerNetwork; OncoDaily).

[Fact] OS: mOS 14.9 (95% CI 12.9–NE) vs 12.1 (11.0–13.4) months, stratified HR 0.75 (0.61–0.92), one-sided P=.0031; 12-month OS 59.7% vs 50.7%, 18-month 44.0% vs 32.5%. PFS: 8.4 vs 5.8 months, HR 0.61 (0.51–0.72), P<.0001. ORR 53.8% vs 32.5%, DCR 91.5% vs 77.8% (CancerNetwork). In the ECOG 1 subgroup OS HR was 0.71 (0.56–0.91), and in patients ≥65 years OS HR was 0.67 (0.47–0.95) (Akeso).

[Fact] All-cause grade 3–4 TEAEs 82.0% vs 72.3%, serious TEAEs 43.4% vs 43.7%, fatal TEAEs 3.2% vs 3.8% (CancerNetwork); any-grade bleeding 9.7% vs 6.5%, grade ≥3 bleeding 1.5% vs 1.8% (Akeso).

For sponsors: "IO plus GemCis" is no longer the ceiling in first-line BTC, and comparator assumptions for the next BTC trial should allow for a PD-1/VEGF bispecific becoming the new standard.

3. Why it succeeded

[Inference] Mechanism: ivonescimab blocks both PD-1 and VEGF (Akeso). The trial has no arm that blocks only one pathway, so the separate contributions of PD-1 and VEGF cannot be distinguished; the 21.3-percentage-point ORR difference (Akeso) shows only higher overall antitumor activity.

[Inference] The comparator is strong: durvalumab plus GemCis is the international standard set by TOPAZ-1. Akeso cites an mOS of 12.8 months for the TOPAZ-1 durvalumab arm (Akeso), and this trial's control arm reached 12.1 months (CancerNetwork), within a reasonable range, so there is no "weak control" problem.

[Inference] But two points call for caution. First, subsequent therapy was imbalanced: 45.9% of the durvalumab arm received subsequent anticancer therapy vs 31.8% of the ivonescimab arm, including targeted therapy 25.1% vs 14.4% (CancerNetwork). More subsequent therapy in the control arm should in theory "dilute" the OS difference, so the imbalance makes the OS conclusion conservative, but it also means care after study treatment differed between arms, and the reason will need the full paper. Second, this is an interim analysis with 14.3 months' median follow-up (CancerNetwork); the upper bound of the HR's confidence interval is 0.92, and the HR may regress as data mature. On safety, grade 3–4 TEAEs were almost 10 percentage points higher (CancerNetwork), mainly hematologic (Akeso).

[Inference] Geography: all patients came from China (NCT06591520; Summit). The etiologic mix of BTC in China (for example liver fluke- and hepatitis B-related intrahepatic cholangiocarcinoma) differs from the West, and we could not find this trial's etiologic or molecular distribution; that is the key gap for extrapolation.

For sponsors: HARMONi-GI1's effect size applies only to a Chinese population, and its OS is likely conservative (the control arm received more subsequent therapy, CancerNetwork); when designing BTC trials, run sensitivity analyses on effect-size assumptions by etiologic mix and subsequent-therapy rates rather than plugging in HR 0.75.

4. Regulatory and commercial implications

[Fact] The first-line BTC indication has been proposed for NMPA priority review (Akeso); ivonescimab has held 3 lung cancer indications in China since 2024 (CancerNetwork). Summit states that HARMONi-GI1 is single-region data generated and analyzed independently by Akeso (Summit). [Fact] FDA precedent: ODAC voted 14 to 1 that the China-only ORIENT-11 could not be applied to the US population (OncLive), and the FDA then issued a CRL recommending a multiregional trial with an OS endpoint (Lilly).

[Inference] HARMONi-GI1 improves on ORIENT-11 in two ways: its primary endpoint is OS, and its comparator, durvalumab plus GemCis, is one of the current US standards; but the "single country" and "no FDA involvement" weaknesses remain (OncLive). Summit has opened the US EA2251 trial (Summit), effectively conceding that HARMONi-GI1 alone is not enough for a US filing. Commercially, the first-line BTC position of AZ's Imfinzi and Merck's Keytruda will be hit first in China; in the West, the earliest change waits for EA2251 (estimated primary completion 2028-12, ClinicalTrials.gov).

For sponsors: single-country Chinese OS data can support an NMPA filing, but a US filing still needs multiregional or US data (Lilly); global development budgets should include the time and cost of a confirmatory trial up front.

5. Recommendations for the next trial and OncoLattice

[Inference] First, sample sizes for new BTC trials can no longer rest solely on the historical mOS of durvalumab plus GemCis; if a sponsor's drug will compete with ivonescimab, consider an adaptive design that can switch comparators mid-trial, or enroll enough patients outside China from the start. Second, subsequent therapy should be specified in the protocol and recorded case by case; HARMONi-GI1's gap of 31.8% vs 45.9% (CancerNetwork) will be an FDA review focus. Third, OncoLattice site strategy: if a client wants to use Chinese data to accelerate global development, US or European sites should be in the same protocol from the start, to avoid repeating ORIENT-11 (OncLive); BTC is a rare cancer, US enrollment relies on cooperative-group networks, and EA2251, led by ECOG-ACRIN, is an example (Summit).

For OncoLattice: however strong a single-country Chinese Phase 3 is, a US filing still needs multiregional or US OS data, and clients' global plans should account for that time and cost.

6. Open questions and what to watch

• Etiologic and molecular distribution (FGFR2, IDH1, HER2, etc.) and details of subsequent therapy: expected in the full LBA8 paper, whose journal and date we could not find; registry at NCT06591520.
• Mature OS and final analysis: see the primary completion date in the ClinicalTrials.gov record (NCT06591520).
• Outcome of NMPA priority review: watch Akeso announcements (Akeso).
• Whether results replicate in a US population: NCT07765433 (EA2251), covered in a separate critique.

The next thing to watch is subsequent therapy and etiologic stratification in the full paper: if the OS benefit holds in patients without liver fluke- or HBV-related disease, the case for extrapolating to the West becomes much stronger.

Evidence limitations

The current data come from Akeso's press release and conference coverage, with no peer-reviewed paper yet; this is an interim analysis at 14.3 months' median follow-up (CancerNetwork).

Advice for readers: when citing HR 0.75, also note "single region in China, interim analysis", and do not use it as the effect-size assumption for Western BTC trials before EA2251 reads out.

Note: differences from the original relay

• n=682, mOS 14.9 vs 12.1 (HR 0.75), PFS HR 0.61 and ORR 53.8% vs 32.5% match the sources (CancerNetwork).
• "7% of the control arm used traditional Chinese medicine as later therapy": not found. None of Akeso's press release, CancerNetwork or OncoDaily gives a TCM rate; what we found was subsequent anticancer therapy 31.8% vs 45.9%, targeted 14.4% vs 25.1% and immunotherapy 14.4% vs 21.1% (CancerNetwork). This article does not use the 7% figure.
• The press release gives only n and part of the design; patient numbers per arm and follow-up come from CancerNetwork and OncoDaily, not from Akeso's original text.