An analysis that looks at the cells of Miriam’s tumor one by one, instead of averaging all of them. These are the first results.
Support for her medical team’s decisions. Not a diagnosis or medical advice.
Status as of 18 October 2026
Part of the analysis already has results. They are published block by block, once verified and approved by Miriam. The rest is still under way. Go to the results
Technique and samples described here: verified
A tumor is not the same mass everywhere. It has tumor cells of several kinds, immune cells, blood vessels and supporting tissue. The usual tests grind up a piece and measure the average.
measures each cell on its own. What it reads is in each one, and that says what type it is and what it is doing.
Metaphor
The smoothie: you know it has strawberry, but not how many strawberries went in or whether one was off.
The fruit bowl: you see each piece, what type it is and how many of each there are.
Two samples that already existed, stored in , which is how a biopsy is kept.
From the liver biopsy of August 2026, the only one done on the liver so far.
From a bone biopsy that contained tumor.
Each one is a small piece of one lesion. Whatever comes out describes that piece, and the two lesions may not resemble each other.
The test is run by Paloma Cejas’s lab at the Dana-Farber Cancer Institute, where she is co-director of the Center for Functional Cancer Epigenetics.
What makes it special is the tissue. Until recently, reading cell by cell needed fresh tissue. With , the lab does it on , the way hospitals store biopsies. That makes it possible to study real patient samples, like these.
This is translational research: taking a lab technique to patient samples, so that what is learned can one day help in the clinic.
“The more we analyze real patient samples, the better our chances of identifying robust biomarkers that can ultimately benefit more patients.”
The lab runs the test and its own analysis. The figures in “Results” come from a separate analysis that Polaris, Miriam’s system, runs on those same data: they are not the lab’s figures. Neither those figures nor the explanations on this page speak for the lab.
This is what the analysis may be able to answer. These are questions, not findings.
Cells are grouped by how similar their switched-on genes are. The result is something like a census: tumor cells (and whether there is more than one type), immune cells, supporting cells.
A partial answer. It shows which immune cells are inside the sample and which activation signals they carry. Some of the molecules through which a tumor shows itself to the immune system, like the , have no probe and this test does not read them.
Among what was looked for: HER3, 5T4, LIV-1, TROP2, Nectin-4, B7-H3, DLL3 and SSTR2. The question is whether their gene is switched on and in what fraction of the tumor cells, a detail that a test on the whole tissue does not give. In the liver, that fraction cannot be counted yet. If a gene is not in the probe panel, it cannot be looked at. They are named as possible to ask about.
This test measures which genes are expressed; it does not read mutations. The come from another test, the sequencing of the . What the cell-by-cell analysis can add is context: in which cells each candidate’s gene is switched on, if it is. inferred
What came out
Each block is published once it has been verified and Miriam has approved it, so they may appear one at a time. Everything describes these two samples, and each one is a small piece of one lesion. The figures come from the analysis that Polaris, Miriam’s system, runs on the lab’s data, not from the lab’s report. When something is the lab’s opinion, the page says so.
Yes. Two biopsies were read cell by cell: about 20,600 cells from the liver and about 2,500 from the bone, a particularly difficult tissue that has to be decalcified before it can be read. After quality control, 19,047 remain in the liver and 1,993 in the bone. According to the lab, both gave excellent-quality data; the bone carries more background noise, as is normal for decalcified tissue.
20,600 cells
2,500 cells
Two tumor genes stand out in both biopsies: NY-BR-1 and 5T4. They are measured in the tumor cells that can be identified most clearly, a selected group rather than the whole tumor. In that group, NY-BR-1 is detected in just over half of the cells and 5T4 in one in four.
Part of that signal is background noise from the sample, so the real figure is a good deal lower. In the bone, 5T4 is still borderline against the non-tumor cells from the same sample. HER3 shows up too, but healthy liver cells have it just as much or more.
Two genes much sought after in other tumors, PRAME and MAGE-A4, do not appear in the RNA above background in either biopsy, and knowing that also helps in choosing the path.
58 in every 100
of the group of 993 cells
54 in every 100
of the group of 68 cells
25 in every 100
of the group of 993 cells
26 in every 100
of the group of 68 cells
PRAME
0 in every 100
MAGE-A4
0 in every 100
There are very few T cells, the defense cells that can attack the tumor: around 1 in 100 cells in the liver biopsy, or fewer. This is what is called a “cold” tumor. The test also shows that, in the RNA, the machinery the tumor uses to show its proteins to the defenses comes out lower than in healthy cells. Knowing this helps point to which type of immune treatment makes the most sense.
Both metastases share the same main genes, and the amplification of the FGFR1 gene, already known from DNA, shows up as active in both. The copy-number profile read in the RNA is consistent with both coming from the same clone; exome sequencing of the liver may be able to confirm it.
FGFR1 active
FGFR1 active
Illustrative diagram: the dots are not real cells and their position is not data.
The tumor is mostly the luminal breast type. In the RNA, the estrogen receptor program is active, although in the liver the stain (immunohistochemistry) for that receptor came out negative: a striking finding Miriam’s medical team has in front of them.
Some of the tumor cells also show neuroendocrine features (CHGA, SYP and INSM1), scattered within the tumor without forming a separate group. In the liver, the INSM1 signal is real; according to the lab, it is not something they usually see in other breast cancer cases, though they do not yet fully understand what it means.
The genes sought in classic neuroendocrine tumors (SSTR2 and DLL3) barely appear in the RNA, although the May DOTATOC PET did show uptake in several bone lesions.
Luminal tumor cell
With neuroendocrine features
Illustrative diagram: neither the number of dots nor their position is data.
These are the steps the analysis follows, in the planned order. If the results change the plan, the list will change too.
3 of 8 steps taken
The bar counts the steps on this list that have already been taken. It does not measure progress toward any medical goal.
Two small pieces from two lesions, stored in paraffin.
Each cell was read one by one, to see which genes it has switched on.
They are on this page, block by block and each with its seal. More will be added.
The analysis measures RNA. Staining the tissue shows whether the protein is actually there. This depends on how much tissue is left. In the glossary:
This is the test that reads the tumor’s own mutations. The cell-by-cell analysis does not. In the glossary:
HLA varies from person to person, a bit like an immune ID card: it is what cells use to show pieces of protein to the immune system. Knowing her type helps show which therapies might fit. In the glossary:
Combine the three pieces to see which candidate targets there could be for a personalized therapy.
The decision belongs to her medical team, with all the information in front of them.
Status of the steps as of 17 October 2026.
Miriam’s medical team sees it first. For them it is one more piece of information. This page does not recommend any treatment, and it will not.
What has already come out is above, with its seal: verified, inferred or unverified. The rest will be added block by block.
If you analyze single-cell data or work on personalized therapy and want to help, the “How to help” page explains how.
How to help