Konstantin Chernyshov: posts tagged 2026

Cancer Wars. Episode IV — A New Hope

Sat, 17 Oct 2026 08:59:39 +0400

Sixteen Years

In 1998, a hospital in Glasgow performed two kidney transplants, one in a man and one in a woman. These are complex operations followed by demanding treatment, but fortunately both kidneys took and functioned perfectly normally. A year and a half later, however, a routine examination revealed a small tumor in the woman’s breast. A few weeks later, several more small tumors appeared, this time under her skin. Tests showed that they were melanoma metastases. In March 2000, the patient died. Two months after her death, a tumor was found in the transplanted kidney of the man who had received his transplant at the same time. Once again, a biopsy showed melanoma.

The Scottish donor registry connected the two stories. Both kidneys had come from the same woman, who had died in May 1998, apparently from a hemorrhage beneath the membranes surrounding the brain. And here was the strange part: she had had melanoma sixteen years before her death. She had undergone surgery and treatment, then lived for fifteen years without signs of disease, attending follow-up appointments at a local clinic. Eventually, she was discharged as apparently healthy. A year after that discharge, she died, and the story of the transplanted kidneys began.

Sixteen years. Could melanoma cells really have survived somewhere in her body all that time, then traveled into the recipients with the transplanted kidneys? Where had those cells been hiding? What had kept them in check? And why did they wake up in someone else’s body?

MacKie R. M. et al. Fatal Melanoma Transferred in a Donated Kidney 16 Years after Melanoma Surgery. New England Journal of Medicine, 2003.

At about the same time, across the Atlantic in St. Louis, a young researcher named Vijay Shankaran and his colleagues were investigating this question experimentally. They compared tumor development in mice with intact and impaired immune defenses. The immune defect arose because the mice lacked RAG2, an important protein needed for T and B lymphocytes to mature. All the mice received the same dose of a carcinogen, but sarcomas developed more often in the group with impaired immunity: 30 of 52 immunodeficient mice developed sarcomas, compared with 11 of 57 ordinary mice. Without mature lymphocytes, tumors appeared more frequently and sooner. The immune system really did influence their fate.

Shankaran V. et al. IFNγ and lymphocytes prevent primary tumour development and shape tumour immunogenicity. Nature, 2001.

But how does a mobile immune cell recognize that one of the body’s own cells has become dangerous? And where does its order to act come from? To reach that order, we need to go back thirty years, to a cell that did not even have a name yet.

A Cell Without a Job

In 1970, a young physician named Ralph Steinman joined Zanvil Cohn’s laboratory at Rockefeller University. Cohn studied macrophages, large cells of the innate immune system that engulf microbes and cellular debris. Steinman was interested in what happened next: once a macrophage had swallowed something, what happened to the cellular remains and proteins, and how did an encounter with foreign material activate lymphocytes?

When he joined Cohn, macrophages were considered likely intermediaries between an antigen and a lymphocyte. One idea at the time was that a macrophage captured a foreign protein, preserved it, and presented it to lymphocytes for recognition.

For two years, Steinman watched macrophages and ran all kinds of experiments. He wanted to find a foreign protein on a macrophage’s surface. But the results did not fit the original idea comfortably. The protein could indeed be detected inside the cell, but once inside, it ended up in vesicles called lysosomes, where it was broken down. Macrophages readily engulfed the material under study, yet the expected transfer of an intact protein to lymphocytes could not be seen. The authors described this difficulty in the paper’s conclusion.

Steinman began looking for another cell and turned to the spleen. Earlier work had reproduced immune responses in cultures of spleen cells, but purified lymphocytes needed help from additional cells that adhered to their surfaces. In 1972, in this mixture, he noticed unusual, mobile cells with branching extensions.

These strange cells stretched out thin processes, drew them back, and changed shape again. According to a biographical account, the first observation took place in 1972. Steinman called Cohn over to his microscope. Cohn’s wife, Fern, later recalled her husband coming home with news of Ralph’s important discovery.

A year later, Steinman and Cohn published their description of the new cell type. Because of the branching processes, they called the cells dendritic, from the Greek word for tree.

Colleagues doubted that such a rare cell could play a decisive role. But in 1978, Steinman and his colleague Margit Witmer showed that dendritic cells were at least a hundred times better than other spleen cells at initiating a lymphocyte response.

By then, Steinman had a student of his own: Michel Nussenzweig.

Passing On the Find

A hundredfold advantage in a test tube did not yet establish a leading role in a living spleen. Dendritic cells made up only a small fraction of its cells, and any result could be attributed to their neighbors. To find out who was actually doing the work, the researchers had to remove dendritic cells, and only dendritic cells, from the mixture. The laboratory produced an antibody called 33D1 that bound specifically to dendritic cells. Together with complement, blood proteins that puncture the membranes of antibody-coated cells, it killed dendritic cells right in the test tube while leaving the others intact. The treated spleen cells were mixed with lymphocytes from a genetically different mouse. In such a mixture, lymphocytes respond to foreign tissue markers, much as they do during transplant rejection. In the 1983 study, removing dendritic cells reduced this reaction by 75–90 percent. Adding purified dendritic cells restored it, and an addition amounting to just half a percent of the stimulating cells was enough.

Steinman R. M. et al. Dendritic cells are the principal stimulators of the primary mixed leukocyte reaction in mice. Journal of Experimental Medicine, 1983.

Gradually, a picture emerged of how the body assembles its immune defenses. Among the many T lymphocytes, only a small fraction can attack other cells. They must receive a signal, multiply, and develop into cells that can reach the relevant tissue. A dendritic cell connects these events: it collects material, presents fragments of it to a lymphocyte, and, given the appropriate signals, initiates a response.

The Display Window

But what exactly does a dendritic cell show a lymphocyte? A T lymphocyte cannot look inside another cell. It sees only what is displayed on the surface, where molecules called MHC play a central role. In humans, these molecules are called HLA. How this encounter worked remained a mystery for a long time.

The answer came from crystallography. At Harvard, in Don Wiley’s laboratory, Pamela Bjorkman and her colleagues spent several years trying to obtain crystals of the HLA-A2 molecule. In 1987, they published its structure. The top of the molecule turned out to be a groove, with two long helices along the edges of a flat floor. And inside that groove was something nobody had put there: on the electron density map, the cleft was occupied by foreign material that the molecule had picked up while it was still in the cell. Later, this material was identified as short protein fragments: peptides.

The display window had come into view. A cell continually cuts up some of its proteins into peptides and displays them outside, in the grooves of MHC class I molecules. A T lymphocyte checks what is on display without looking inside. If a viral or altered protein appears in the cell, a fragment of it can appear in the window.

Bjorkman P. J. et al. Structure of the human class I histocompatibility antigen, HLA-A2. Nature, 1987. Hildner K. et al. Batf3 deficiency reveals a critical role for CD8α+ dendritic cells in cytotoxic T cell immunity. Science, 2008.

A dendritic cell can do more: on MHC class I, it can display material captured and processed from another cell, as well as material from inside itself. This pathway was named cross-presentation. It allows killer T lymphocytes to be prepared for an encounter with a tumor before they reach it themselves.

How important this was for tumor defense was tested in 2008. Kai Hildner and his colleagues used mice lacking the Batf3 gene. These animals had almost none of one branch of dendritic cells specialized for cross-presentation. A fibrosarcoma that ordinary mice rejected grew in these animals; at the same time, the initiation of specific cytotoxic T-cell responses was impaired. In this transplantation model, the defenses failed to assemble without the necessary intermediaries.

From a strange cell under a microscope came an explanation of how many cells begin to act together: one finds a clue, another receives an order, and their descendants go on the attack.

But remember that the tumor arose from the body’s own cells. What happens if the body decides to attack itself? When that happens, a particular group of diseases develops: autoimmune diseases. But that is not our case. We know that something has changed inside the malignant cell. What on its display window might reveal that change?

A Clue from the Body Itself

In the late 1970s, while Varmus and Bishop in San Francisco were discovering that the viral oncogene src was present in the normal cells of birds and mammals, Lloyd Old’s group in New York, on the other side of the country, was studying features on the outside of tumor cells. They were looking for molecules on human malignant cells that the patient’s own immune system could recognize. The method was called autologous typing. Researchers took tumor cells and lymphocytes from the same patient and watched to see whether one would attack the other. The patient’s own cells against the patient’s own cells, with no foreign tissue and no mice.

A young German physician named Alexander Knuth took part in this work. Among the thirteen melanoma patients the group studied, the strongest reaction came from a man identified by the initials A.V. His T lymphocytes killed cells from his own tumor, grown in a culture called SK-MEL-29. In 1981, Knuth returned to Mainz and continued the same work. There, a second such patient was found, a woman identified in the scientific papers as MZ-2. She had stage IV melanoma, with metastases in the kidney, ovaries, lymph nodes, and spleen. Repeated operations and chemotherapy had not produced complete remission. But her T lymphocytes, like A.V.’s, recognized her tumor cells. Remember both patients: in each, the immune system had found its own clue in the tumor.

They recognized it, but what exactly were they recognizing? To find out, in 1984 Alexander Knuth followed the advice of his former mentor Lloyd Old and traveled to Brussels to meet another scientist, Thierry Boon. Boon’s group had developed a method that used activated T cells to find the tumor gene responsible for activating those same lymphocytes.

Boon and Knuth began the search. They took two types of cells from patient MZ-2: killer T cells and melanoma cells. From the killer T cells, they grew clones that were best at recognizing and destroying the melanoma cells. In the second stage, they continued to bring these selected killer T cells into contact with melanoma, but now selected only the rare melanoma cells that escaped their killers by losing the target. There were now three kinds of cells: killer T cells, ordinary melanoma cells, and escape variants. The third stage was the largest. They made a library from the ordinary melanoma: DNA was chopped into large pieces and inserted into hundreds of thousands of plasmids, circular DNA molecules used as carriers. They introduced these plasmids in batches into the escape cells. The idea was this: if introducing a plasmid caused a killer T cell to recognize and destroy the escaping melanoma again, that cosmid must contain the gene that produced the molecule displayed on MHC-I. And they found such a gene.

van der Bruggen P. et al. A gene encoding an antigen recognized by cytolytic T lymphocytes on a human melanoma. Science, 1991.

What would you expect from the gene they had found? After everything we know about mutations in cancer, the answer seems obvious: an altered protein would give the tumor away. In 1991, Pierre van der Bruggen, Alexander Knuth, Thierry Boon, and their colleagues published the gene. They named it MAGE, from melanoma antigen. It was the first gene for a human tumor antigen recognized by the patient’s own T lymphocytes. And the surprising thing was that it was not mutated! It was exactly the same in the patient’s normal cells. Only four years later would the mystery begin to clear: in a normal body, this gene is silent almost everywhere except in immature sperm cells in the testis. In melanoma, it had switched on at the wrong time and in the wrong place. What gave the tumor away was not an altered protein, but one of the body’s own proteins appearing where it should not.

And what about the first patient? Work with his cells continued in Mainz, and in 1995 Thomas Wölfel’s group found what A.V.’s lymphocytes were targeting. This time, it really was a mutation, in the CDK4 gene, whose protein helps a cell move toward copying its DNA. At position 24, cysteine had replaced arginine. The substitution was present in the melanoma and absent from the patient’s lymphocytes. Inside the cell, it interfered with CDK4 binding to the inhibitory protein p16, weakening one of the brakes on cell division. Outside the cell, a fragment of that same altered protein, displayed on HLA-A2, became a target for the patient’s T cells.

A mutation that weakened an internal brake on division simultaneously left a trace for external surveillance. A target arising from a mutation is called a neoantigen. The cell escaped one constraint and became more visible to another.

Wölfel T. et al. A p16INK4a-insensitive CDK4 mutant targeted by cytolytic T lymphocytes in a human melanoma. Science, 1995.

Two clues, two ways to give yourself away. MZ-2’s melanoma switched on something that should have stayed silent. A.V.’s melanoma broke something that should have worked. Both traces end up in the same display window, and the immune system can read both. If the Glasgow donor’s melanoma had even one such clue, there was something for the guards to see.

De Smet C. et al. The activation of human gene MAGE-1 in tumor cells is correlated with genome-wide demethylation. PNAS, 1996. Wölfel T. et al. A p16INK4a-insensitive CDK4 mutant targeted by cytolytic T lymphocytes in a human melanoma. Science, 1995.

We can now assemble the chain of a targeted attack. A dendritic cell acquires tumor material and helps activate the appropriate T lymphocytes. Their descendants arrive in the tissue and inspect peptides on the surfaces of cells. On finding its target, a cytotoxic lymphocyte delivers a lethal strike.

But what if the tumor removes the display window itself: the MHC-I molecules? This is where the story of recognition took an unexpected turn.

A Line in a Dissertation

In the autumn of 1981, Klas Kärre was working on his dissertation at the Karolinska Institute. He was interested in natural killer cells, or NK cells. They could destroy certain tumor cells without prior immunization, but the rule by which they chose their targets remained unclear. What common features linked the very different cells they attacked?

Kärre later recalled struggling with a single sentence. He was trying to describe susceptible cells and noticed that it was much easier to state what resistant cells possessed. The question turned around: what feature allows a cell to avoid attack?

MHC-I already occupied a central place in ideas about recognition. Kärre proposed that its presence might serve as a signal holding an NK cell back from attacking. The disappearance of a familiar feature would then change a cell’s fate. To notice danger, it was not always necessary to find a new foreign molecule. Sometimes, losing a normal feature of the self could be enough.

The idea still had to be made to work in an experiment. A twenty-two-year-old student, Hans-Gustaf Ljunggren, joined Kärre. His summer project was to test whether NK cells could recognize tumor cells that had lost certain molecules belonging to their own organism.

The researchers took a mouse lymphoma resistant to NK cells and obtained variants with sharply reduced amounts of MHC-I. In the first stage of the experiment, the original and altered lymphoma cells multiplied at the same rate. Losing surface molecules did not take away the tumor’s ability to live and divide. In the second stage, after the lymphoma cells were injected into ordinary mice, the variants met different fates: the original lymphoma cells grew into a new tumor, while cells with low MHC-I died and did not establish new lesions. In the third stage, the researchers suppressed NK-cell activity. The previously rejected variants then began forming tumors again.

Ljunggren H.-G., Kärre K. Host resistance directed selectively against H-2-deficient lymphoma variants. Journal of Experimental Medicine, 1985.

This principle became known as missing self. Normal MHC-I molecules send an inhibitory signal to an NK cell; as their numbers decline, that inhibition is lost. Today, we know that activating signals, including signs of cellular stress, also matter for NK-cell activation. But the direction of Kärre’s thinking has endured. The immune system can detect a disturbance through the disappearance of something that ought to be present.

The two branches of defense looked at the same surface in different ways. A T lymphocyte read the peptide on display; an NK cell took account of whether normal identification molecules remained. For a tumor lineage, these requirements made uncomfortable neighbors. Hiding from one means of detection could mean becoming more vulnerable to another.

Two Hundred Days

The mechanisms we have found let us picture the destruction of a dangerous clone: recognition, assembly of the response, arrival of lymphocytes, and cell death. But what was happening where no growing tumor could be seen?

The Glasgow donor’s story shows that cells with malignant potential could persist after sixteen years without clinical signs of melanoma. It raises a question: had the immune system been keeping them in check all that time?

Catherine Koebel and her colleagues modeled such a situation. They gave ordinary mice a small dose of the carcinogen 3-methylcholanthrene. Some animals developed growing sarcomas. The others remained under observation. Two hundred days passed, more than half a year. Some mice still had no detectable tumor; in others, a small, stable mass remained at the injection site.

On day 200, the researchers began giving one group a control immunoglobulin that did not affect immunity, and another group antibodies against lymphocytes. In effect, they switched off several important components of immune control. They did not give the carcinogen again. After this intervention, growing tumors appeared in seven of fifteen mice in the second group, compared with one of thirteen in the control group.

What happened forced a new reading of the preceding two hundred days. At least some of the animals had harbored a population capable of growing all along. Its presence became visible once the defenses were removed.

Cells taken from small, stable masses produced progressive tumors when transferred into immunodeficient mice. The long period of quiet had not made them safe. Within the masses themselves, researchers found fewer signs of cell division, more signs of cell death, and more T lymphocytes than in growing sarcomas. Together, these results supported a picture of dynamic equilibrium: the immune response was holding the tumor population’s growth in check.

After two hundred days of apparent quiet, another outcome of immune defense had emerged: it could contain a dangerous population without destroying it completely. This kind of work is easy to miss if total destruction of the enemy is the only thing that counts as victory.

Koebel C. M. et al. Adaptive immunity maintains occult cancer in an equilibrium state. Nature, 2007.

For an evolving tumor, this meant that acquiring the ability to multiply was not enough. Its descendants also had to survive the pressure exerted by the surrounding organism. The lineage’s fate depended on whether it could produce more living cells than it lost. The immune system was part of that calculation.

Back to Glasgow

Now let us return to the donor. The kidney transplant replayed Koebel’s experiment in a human body, without the design or the controls. To allow the kidneys to take, the recipients were given drugs that suppressed their immune responses. The defenses were removed, much as antibodies had removed them on day 200, and melanoma grew within months.

In the second recipient, immunosuppression was withdrawn and interferon, which itself strengthens the immune response, was prescribed. After thirteen weeks, signs of rejection appeared. The kidney was removed, and extensive masses of dead melanoma cells were found inside it. Two years after the melanoma was detected, the man was alive and had no clinical signs of a tumor. When the guards returned, they began killing again.

MacKie R. M. et al. Fatal Melanoma Transferred in a Donated Kidney 16 Years after Melanoma Surgery. New England Journal of Medicine, 2003.

A single case cannot prove that the immune system was what held the melanoma in check for sixteen years. But the laboratory findings fit that interpretation. Those sixteen quiet years were not empty. This defense had no witnesses because it was doing its job.

The Student’s Voice

On December 7, 2011, Michel Nussenzweig delivered the Nobel Lecture on dendritic cells. He was the same student who had helped Steinman show that a tiny fraction of spleen cells did the crucial work. Now he was telling the story of the discovery on his teacher’s behalf.

Ralph Steinman himself had died of pancreatic cancer three days before the Nobel Prize was announced. The disease had been diagnosed four years earlier. Steinman was treated with surgery, chemotherapy, and experimental immunotherapy based on dendritic cells that he had developed himself. We cannot say what role each part of the treatment played, but four years was a relatively long time for pancreatic cancer, especially then.

Alongside his story stand two others. Remember the two patients Alexander Knuth and Thierry Boon had worked on? MZ-2 and A.V. lived for more than thirty years without their disease returning. The researchers themselves linked this to the strong and diverse T-cell responses in both patients.

The body has defenses that keep working after the internal brakes of its cells have failed. These defenses receive information about what is inside cells, notice the loss of normal features, assemble a response, and destroy their targets. Sometimes, all their work leaves behind is the absence of a growing tumor.

Publications in 2026

Sat, 11 Jul 2026 20:16:59 +0400

In 2026, I am leading the Bioinformatics Department at BostonGene. In addition to our core work — client projects — we run a small research program within the department to deepen our expertise in oncology and build expertise in other areas, such as autoimmune diseases. This year, I led one complete research project; on the others, I served as a mentor and oversaw quality control.

American Association for Cancer Research

“A comparative scoring framework for BCMA-, GPRC5D- and CD38-targeted therapies in multiple myeloma”

I am the first author here: I came up with and substantiated the idea, designed the experiments, assembled the team, delegated parts of the work, pulled the results together, designed the visualizations, wrote the text, and presented the study at the conference.

We proposed the Actionability Score — an NGS-based approach for assessing which target and which type of immunotherapy may be best suited for a patient with multiple myeloma: BCMA- or GPRC5D-targeted therapy, CAR-T, a T-cell engager, or an anti-CD38 antibody. The score accounts for target activity, resistance features, and the state of the tumor microenvironment. The results looked promising in a small retrospective cohort, but the approach still needs to be validated in larger cohorts.

BostonGene news. AACR abstract.

Update, 8 September 2026: an updated poster, “An Actionability Score for BCMA-, GPRC5D-, and CD38-targeted therapy selection in multiple myeloma”, is listed in Clinical Lymphoma, Myeloma & Leukemia. It develops the AACR work; I remain the first author. Updated poster and abstract.

“Integrative application of Lymphly reveals shared molecular landscape across B-cell lymphomas”

Lymphly was originally developed for diffuse large B-cell lymphoma. We applied it to other B-cell lymphomas and found shared genetic and transcriptomic patterns across diagnoses. For example, individual Lymphly subtypes were linked to follicular, Burkitt, marginal zone, and primary mediastinal B-cell lymphomas.

BostonGene news. AACR abstract.

Related work, 8 September 2026: “Cross-disease stratification by genetic classifier Lymphly reveals shared molecular programs across B-cell malignancies”, listed in Clinical Lymphoma, Myeloma & Leukemia. The analysis includes 4,353 DLBCL and 2,203 other B-cell lymphoma samples. I am a co-author. Poster and abstract.

American Society of Clinical Oncology

The main clinical oncology conference in the US.

“Examining the tumor microbiology and microenvironment of patients with triple negative inflammatory breast cancer receiving Keynote 522”

Together with the MD Anderson team, we studied tumor samples from patients with triple-negative inflammatory breast cancer receiving the KEYNOTE-522 regimen. Patients with a pathologic complete response had higher PD-L1 expression and a trend toward greater T-cell abundance. Non-responders showed angiogenic changes; some post-treatment samples showed increased fibrosis or a reduced immune component. The cohort was small, so these observations need further validation. I am a co-author of the abstract.

ASCO abstract on BostonGene.

European Hematology Association

Europe’s leading hematology conference.

“Risk-stratified patient selection in multiple myeloma for frontline treatment decision-making based on a transcriptomic classifier”

We combined the ideas behind two transcriptomic multiple myeloma classifiers and built a consensus risk model.

BostonGene news. EHA abstract.

“Overlaying TP53 loss and aneuploidy with distinct diffuse large B-cell lymphoma subtypes refines tumor classification and enhances risk stratification”

We added two independent markers of genomic instability to Lymphly: loss of TP53 function and a quantitative measure of aneuploidy. The really cool result is that patients with aneuploidy but no TP53 loss have the best survival! Their genome is a mess, but the cell’s protective mechanisms are still trying to make it self-destruct.

BostonGene news.

AACR 2026. The next step for foundation models toward real-world products

Tue, 28 Apr 2026 01:03:22 +0400

the path from embedding to product is long and difficult

In the world of large models, generating embeddings and imputing data are only early steps. The final product-level steps, such as response prediction, are much more difficult. The hard part is to understand and implement the intermediate steps between imputation and an actual real-world product.

I gathered some thoughts from the AACR 2026 and formulated like this:

How to use embeddings and imputation → to get one step closer to a product.

Synthetic histology slides → against class imbalance

One interesting application of synthetic histology slides is generating slides for rare disease subtypes to overcome class imbalance. The problem is that a rare subtype is… rare. And when training a model, there are not enough samples of the rare subtype for the model to learn it well.

The proposed approach is to generate histology slides of the rare subtype at the dataset preparation stage. In fact, synthetic embeddings of the rare subtype are enough, even without visualization. After adding this extra set of synthetic samples, the main model learns the rare subtype better and predicts it with higher quality.

The paper about the BERGERON and CRAWFORD models from Stanford.

In the images: 1. The general scheme of the two models. BERGERON creates embeddings of real and synthetic histology slides, while CRAWFORD visualizes them. 2. Synthetic embeddings do not batch separately from real ones. 3. The accuracy of rare subtype prediction increased by 0.1 after adding synthetic data.

Attention analysis → for finding contextual, model-derived hypotheses

Model attention analysis is the calculation of features that the model relies on more strongly than others when making a decision.

At the level of a single slide, we can ask: what exactly did the model see in this region that is not present elsewhere? At the cohort level: how do low-risk slides differ from high-risk slides? How do malignant regions on slides differ from normal regions? In other words, we compare not one whole slide against another whole slide, but the most interesting regions specifically.

In the images: 1. Visualization of model attention. It is visible that chromatin in the center of the nucleus is the most important feature for high risk. 2. Patches are arranged by decreasing gradient of the feature that is important for risk prediction. Note that the full slide is not shown — only small patches are shown. 3. For samples with paired spatial transcriptomics, gene activity could also be studied. 4. Differential expression analysis is performed specifically for the interesting patches, not across the whole slide, and with a risk-based contrast.

Benchmarking pressure → for directed model development

Every serious talk about new models included the claim: “our model performs better than other methods on benchmarks, look.” Without this validation, the rest of the talk was no longer perceived as convincing.

But it was even more interesting when teams drew a concrete conclusion from this and took the next step, using benchmark insights to improve the model.

Multimodality → works better than a single modality

Faisal Mahmood, the father of all AI for pathology, shows in the TANGLE paper that if a model is trained on both histology and transcriptomics, the results are better than with histology alone.

More data → the model works better

Bioptimus, a French company and creator of commercial AI models, released the M-optimus model and showed that as the training set increases, model performance also improves, and the plateau has not yet been reached. Although I was not impressed by the result: when the training set increased from 8× to 16×, the correlation of imputed spatial expression increased from 0.136 to 0.140? What even is that, lol.

Text descriptions of slides → for better quality with a smaller training set

Yana Lipkova gave a review of benchmark reviews for pathology models.

It was interesting that the CONCH model shows performance comparable to top models while using a training set that is an order of magnitude smaller. This was achieved because CONCH was trained on slides with text descriptions, unlike most modern models, which are trained only on images.

CONCH

Working with context and autoregression → for better protein function prediction

The author of scGPT, Bo Wang, presented the BioReason-Pro work. This is a pipeline that helps describe the function of a protein when its sequence is known, but no wet-lab experiment has been performed yet. The first stage of this pipeline is GO-GPT, a model that generates preliminary hypotheses about protein functions based on the protein sequence.

GO-GPT is an example of directed model development: benchmarks and experience from CAFA-like evaluations showed that flat prediction of Gene Ontology labels is not enough. The authors took the next step: they turned GO annotation into a task of sequential term generation — in other words, they added autoregression so that the architecture itself better matches the structure of Gene Ontology.

In the images: 1. GO-GPT beats previous SOTA approaches on benchmarks. 2. The architecture of the whole BioReason pipeline. As I understood it, the two upper arrows are GO-GPT, and there is also text embedding preparation; then a giga-prompt is assembled from this and passed into a third model.

AACR 2026 trends

Tue, 28 Apr 2026 00:48:23 +0400

AACR is a large annual oncology conference. This year it took place in San Diego, and 22,000 participants attended the conference.

In 2026, two trends were clearly visible at AACR: impressive results from large biological models, and many new experimental drugs being developed through classical approaches. AACR 2026 tried to find points of connection between these two trends.

Since BostonGene is taking part in this race from the side of the first trend, offering its expertise and tools to help advance therapies through clinical trial phases, I focused mainly on the AI sessions.

BostonGene came to the conference

Conference Opening

The opening immediately set the tone for the whole conference. Three speakers showed that challenges in oncology still remain and discoveries happen every year, while the arms race does not stop either: new drugs are also being developed every year. The question is: can AI help us with this?

Challenges in oncology still exist. Prostate cancer develops resistance to AR inhibitors by changing its histology.

Drug development is not standing still. CAR-T therapies for solid tumors are being actively tested.

How can AI help us? The SPARC model: understanding the activity of molecular patterns from images and predicting risk. The Boltz model: personalized selection of targeted therapy by modeling the interaction between a drug molecule and a mutated target protein. The MIRAI model: AI-based recognition of mammography images and 5-year risk assessment.

News from the World of Large Biological Models

The session was chaired by Jakob Nikolas Kather, a rising star and professor of clinical AI at the University of Dresden, Germany.

Jakob said that pathology foundation models are something from 2020. 2026 is the year of agents that work with databases of clinical records and answer doctors in human language.

Faisal Mahmood, the father of AI in pathology, presented several new models from his lab.

Faisal understood that it was time to move further and presented Apollo — a large foundation model whose main modalities are standard clinical records about the patient and results of routine laboratory and instrumental tests, such as blood tests and X-rays. The model can understand risk factors, predict disease progression, and predict treatment outcomes.

A new histology model, a new model for histology + transcriptomics, a new model for histology + genomics, and a new model for spatial transcriptomics.

Bo Wang, the creator of scGPT, also presented two new models from his lab.

X-Cell: a new improved version of scGPT. BioReason: a model for predicting the function of proteins for which the amino acid sequence is already known, but no wet-lab experiment has been performed.

Jure Leskovec presented Biomni — an agentic system designed for multi-step work on biological questions. We should test it.

The agent can work directly with bioinformatics tools and search databases. Examples of use: 1. the agent designed a wet-lab protocol; 2. the agent went through 458 tables and files from a study and generated hypotheses that can then be tested.

Poster Session

The following trend was visible at the poster sessions. The sections about new classes of therapies and first-in-human trials were extremely crowded. The AI sections had moderate interest. But the classical sections on clinical studies and bioinformatics were rather quiet.

1, 2 — immunology and experimental therapy; 3 — AI section; 4 — clinical research: prognostic biomarkers, molecular cancer classifiers, population studies, liquid biopsy.

I presented two works. The first one was about identifying the right therapy for a patient with multiple myeloma. The second one was about shared genetic patterns between diffuse large B-cell lymphoma and lymphomas of other histologies.

The main conclusion is that we can work on this area where AI meets the trend of advancing new types of therapies.