Health AI · Oncology informatics
Triomics
An oncology-tuned language model that reads the whole cancer chart and hands research coordinators a cited worklist before clinic opens. Adopted at Memorial Sloan Kettering, Mount Sinai and Yale, with a Series B closed in May 2026 and a published record narrower than the marketing around it.
Forty-five minutes a chart
Memorial Sloan Kettering keeps as many as 1,800 clinical trials open at once. More than 400 research coordinators work that portfolio, and the way they find candidates is by reading. Staging details. Biomarkers. Prior therapies. Then they retype it into systems that do not talk to each other.
“Copy-and-paste has become an unfortunate but common bridge today,” said Joe Lengfellner, who runs MSK’s Clinical Research Innovation Consortium, in November 2025.
MSK puts the cost of prescreening one patient against a large trial portfolio at up to 45 minutes, longer when the record includes clinician notes and PDFs. Multiply that by a clinic schedule and screening stops being a task. It becomes a rationing decision about which patients get looked at. In mid-2024 MSK ran a head to head competition across multiple vendors to see whether software could take it over.
Where the information hides
Roughly 80 percent of what decides whether a cancer patient fits a trial sits in unstructured text: progress notes, pathology, imaging narratives, outside records, scanned faxes. Software has queried the structured fields for years. The text was the wall.
The company’s own peer-reviewed paper opens by noting that only about 7 percent of adults participate in cancer clinical trials, and points at physician screening burden as a major cause.
“And missing a trial can mean someone losing months or years of time with loved ones,” Sarim Khan said in 2025.
A biotech researcher and an Adobe AI lead
Sarim Khan and Hrituraj Singh were college friends. Khan went on to biotech research at MIT, where he wanted a faster route to the data his own work needed. Singh went on to lead generative AI efforts at Adobe. They founded Triomics in 2021 and went through Y Combinator that winter. Khan runs the company; Singh runs the models and is a named author on both the matching paper and the 2025 abstraction preprint.
“Oncology is the hardest place to build AI, yet the most important,” Singh said when the Series B was announced.
Milwaukee first, then the top ten
The first adopter was the Medical College of Wisconsin Cancer Center, which began running OncoLLM across four disease teams in July 2024. The company’s central validation work was done with MCW investigators and published in npj Digital Medicine that October.
The rest arrived quickly. MSK announced a two-phase rollout on November 4, 2025, after a retrospective bake off. Yale Cancer Center and Smilow Cancer Hospital agreed that same month to screen their full active trial portfolio. Mount Sinai Tisch Cancer Center deployed system wide in January 2026, reaching Queens, Brooklyn, South Nassau, Morningside and West as well as the flagship. In May 2026 Battery Ventures led a $22 million Series B, joined by Texas Oncology subsidiary Precision Health Informatics.
What is proven, and what is still claimed
| Evidence | What the record shows | Source type |
|---|---|---|
| npj Digital Medicine, Oct 28, 2024 | The PRISM paper. OncoLLM is a 14B parameter model fine-tuned from Qwen-1.5 14B on one cancer center’s EHR data. Criterion-level accuracy 63 percent, against 53 percent for GPT-3.5 Turbo and 68 percent for GPT-4. Citations correct 86.71 percent of the time. | Public record |
| The 95 percent figure | The company product page says the paper demonstrated 95 percent accuracy in matching patients to trials. In the paper, one expert checked 10 trials OncoLLM ranked first that were not the trial the patient enrolled in, found the patient eligible for 9, and the authors wrote that this effectively raises top-3 accuracy to about 95 percent. A 10 case adjudication, not a headline accuracy. | Differs from the paper |
| Benchmark against doctors | The paper says LLMs can almost match qualified medical doctors here. The human bar is agreement among five physician annotators: 64 percent across all five, 70 percent for the two best. OncoLLM scored 63 percent, so the comparison is to physician agreement, not a gold standard. | Public record |
| ASCO 2025, abstract e13706 | MCW, December 2024 to January 2025. 2,277 patients across 38 tumor types, 100 percent of those seen. 49 flagged for one trial, ALLIANCE-A222004-ANOREXIA. 22 of the 49 accepted and watch-listed. Reported as a 40 percent increase in matches. | Public record |
| Surgery, March 2026 | MCW gastrointestinal surgical oncology, July to December 2024. 514 patients evaluated, 34 matches, 9 enrollments. Of the 25 that did not convert: 9 ineligible, 5 declined, 4 provider discretion, 7 unexplained. A 26.5 percent conversion rate. | Public record |
| HARMON-E preprint, Dec 2025 | Abstraction evaluated on more than 400,000 notes and scanned PDFs from 2,250 patients. Average F1 0.93, 100 of 103 variables above 0.85. Posted to arXiv, not peer reviewed. | Public record |
| How the outcome numbers are sourced | Coverage of the Series B says npj Digital Medicine documented a 67 percent cut in chart review, 40 percent more matches and 30 percent more enrollments. None of those numbers appear in that paper. The 40 percent is from the ASCO 2025 abstract, the accrual figure from a separate ASCO Quality Care Symposium abstract, and the chart review figure is a company estimate, put at 67 percent in the release and 70 percent on the website. | Misattributed |
| Cost advantage | Company executives told Fierce Healthcare the model was 40 times cheaper than proprietary models. The paper reports about $170 against about $6,055 for the same 980 patient-trial pairs, which its authors call about 35-fold, or $0.17 against $6.18 per pair. | Differs from the paper |
| Deployments | MSK, Mount Sinai and MCW published the work under their own mastheads. Yale leaders are quoted in a company release. MD Anderson and Texas Oncology appear only in company statements, as do the website figures of 1.8 million records processed and 10 or more NCI partners. | Mixed, partner and company |
| Author independence | Six of the thirteen npj Digital Medicine authors are Triomics employees and a seventh consults for the company. MSK discloses institutional financial interests related to Triomics, and its senior vice president for clinical research sits on the company advisory board. | Public record |
| Regulatory and registry footprint | No FDA 510(k) or PMA record. No registered trial naming the company. No NIH or NSF award. The peer-reviewed work states it received no specific grant from any funding agency, so the validation was self-funded. | Not found |
Read plainly: the engineering is real, published, and unusually specific about its own limits. The record supports that an oncology-tuned model screens every scheduled patient instead of a fraction, surfaces candidates a human screen missed, and does it far below frontier-model cost. It does not yet support the round numbers used to sell that. The strongest outcome data comes from one cancer center, one supportive care trial, and a surgical pilot in which three quarters of the matches did not become an enrollment. That last paper is the most useful thing the company has published, because it counts what did not happen.
What to watch
- Results from MSK and Mount Sinai. Both said they would publish, neither has, and both dwarf the Wisconsin pilot behind the current numbers.
- Whether HARMON-E clears peer review, and whether its 0.93 F1 holds on a second institution’s notes.
- Enrollment, not matching. Matches are cheap to raise; the Surgery paper shows where they leak.
- Whether the company restates the 95 and 67 percent figures to match what was measured.
- A Form D for the Series B. None had been filed as of September 22, 2026.
In their words
“We have seen medical records [with] thousands of pages of information,”
Sarim Khan, co-founder and CEO, to TechCrunch, 2026 · Independent
“And missing a trial can mean someone losing months or years of time with loved ones.”
Sarim Khan, Microsoft Bay Area blog, 2025 · Partner blog
“Copy-and-paste has become an unfortunate but common bridge today.”
Joe Lengfellner, senior director of clinical research IT, Memorial Sloan Kettering, Healthcare IT News, 2025 · Independent
“The analysis was unambiguous: The approach works as intended and can widen access by reducing false negatives in screening.”
Joe Lengfellner, Memorial Sloan Kettering, Healthcare IT News, 2025 · Independent
“Since implementing the OncoLLM platform in July, we’ve been able to evaluate all the patients in the Disease Oriented Teams (DOTs) that are part of our pilot project.”
Anai Kothari, MD, MS, co-principal investigator, Medical College of Wisconsin Cancer Center · Site investigator
“By deploying an AI platform trained specifically for oncology, we can identify trial opportunities earlier, more consistently, and more equitably, allowing clinicians to focus on meaningful conversations with patients rather than manual chart review.”
Karyn Goodman, MD, MS, associate director of clinical research, Mount Sinai Tisch Cancer Center, January 2026 · Customer release
“Getting a model to reason reliably across thousands of pages of notes, pathology, imaging and evolving trial criteria, and show its work, is what separates a demo from software that clinicians actually use.”
Hrituraj Singh, co-founder and CTO, Series B release, May 2026 · Company release
Related companies
Sources
- Public recordPRISM: Patient Records Interpretation for Semantic clinical trial Matching
- Public recordEnterprise-level AI screening for oncology trials, e13706
- Public recordUnderstanding unrealized trial enrollments after patient-to-trial matching
- Public recordHARMON-E, arXiv 2512.19864
- Public recordForm D, Triomics, Inc., CIK 0001962627
- IndependentMemorial Sloan Kettering innovates clinical trials with AI
- IndependentTriomics nabs $22M to bring oncology AI to cancer centers
- IndependentWhere participation in cancer trials lacks, Triomics could step in
- IndependentHow Triomics uses AI to help doctors treat cancer more quickly
- CustomerMSK to deploy Triomics’ platform
- CustomerMount Sinai launches AI trial-matching platform
- CustomerTrailblazers in innovation
- CompanyTriomics raises $22 million
- CompanyTriomics website, PRISM and Harmony pages
Profile researched and written by Healthcare Discovery. Last updated September 29, 2026.
