{
    "version": "https:\/\/jsonfeed.org\/version\/1.1",
    "title": "Konstantin Chernyshov: posts tagged bioinformatics",
    "_rss_description": "Konstantin Chernyshov, manager, bioinformatic, leader in biotech",
    "_rss_language": "en",
    "_itunes_email": "",
    "_itunes_categories_xml": "",
    "_itunes_image": "https:\/\/kostachernyshov.com\/blog\/pictures\/userpic\/userpic-square@2x.jpg?1731242071",
    "_itunes_explicit": "no",
    "home_page_url": "https:\/\/kostachernyshov.com\/blog\/tags\/bioinformatics\/",
    "feed_url": "https:\/\/kostachernyshov.com\/blog\/tags\/bioinformatics\/json\/",
    "icon": "https:\/\/kostachernyshov.com\/blog\/pictures\/userpic\/userpic@2x.jpg?1731242071",
    "authors": [
        {
            "name": "Konstantin Chernyshov",
            "url": "https:\/\/kostachernyshov.com\/blog\/",
            "avatar": "https:\/\/kostachernyshov.com\/blog\/pictures\/userpic\/userpic@2x.jpg?1731242071"
        }
    ],
    "items": [
        {
            "id": "46",
            "url": "https:\/\/kostachernyshov.com\/blog\/all\/bioinformatician-growth\/",
            "title": "Bioinformatician growth",
            "content_html": "<p>A bioinformatician on my team is always looking to grow and improve their skills. These checklists help identify the skills to develop and the knowledge to acquire.<\/p>\n<p>The checklists are the minimum required to move to the next level. The full set of requirements is discussed and agreed with the manager. You need to complete every checklist item to advance, but completing them all does not guarantee advancement: the manager may also set individual requirements.<\/p>\n<p>My team works in oncology and mainly deals with NGS data, so some items are specific to those areas. Some of the linked learning resources are in Russian.<\/p>\n<h2>Trainee<\/h2>\n<p>Knows the basics of molecular biology, statistics, and NGS analysis. Understands the null hypothesis, p-values, Fisher’s exact test, the chi-squared test, the t-test, the Mann–Whitney U test, and multiple testing correction.<\/p>\n<p>Can run a Jupyter notebook and use Python to read a data frame, filter its columns, and draw a histogram and a scatter plot.<\/p>\n<p>Learns the principle that working on something is not the same as getting it done. Shows relentless drive. Will revise test assignments and work tasks ten times if needed, persists from the very first tasks, and is not afraid of unfamiliar topics.<\/p>\n<p>Takes responsibility. The particular task matters less than the ability to take ownership and deliver a result. Volunteers to keep time at a team meeting, find the person responsible for the office and get a chair repaired, work out how to download a dataset, or set up an agent and write instructions for colleagues.<\/p>\n<p>Learns to organize their work. Until they gain experience, works on simple, routine tasks under close supervision from senior colleagues. Keeps track of their tasks and takes meeting notes without being reminded.<\/p>\n<h3>Studied<\/h3>\n<ul>\n<li><a href=\"https:\/\/stepik.org\/70\">The basics of molecular biology<\/a><\/li>\n<li>The basics of statistics, for example the two-part Stepik course: <a href=\"https:\/\/stepik.org\/lesson\/8095\/step\/1?unit=1371\">part one<\/a> and <a href=\"https:\/\/stepik.org\/course\/524\/promo\">part two<\/a><\/li>\n<li>The basics of <a href=\"https:\/\/stepik.org\/course\/67\/promo\">Python<\/a> or <a href=\"https:\/\/stepik.org\/course\/497\/syllabus\">R<\/a><\/li>\n<li>The basics of NGS, for example <a href=\"https:\/\/stepik.org\/1142\">on Stepik<\/a><\/li>\n<li>Clear written communication: headings, captions, links, and quotations in <a href=\"https:\/\/kinzhal.media\/howto-messenger\/\">messengers<\/a> and <a href=\"https:\/\/ilyabirman.ru\/meanwhile\/all\/neat-mail\/\">email<\/a><\/li>\n<li>Google Docs: formatting, styles, and access settings<\/li>\n<li>Google Slides<\/li>\n<\/ul>\n<h3>Read<\/h3>\n<ul>\n<li><a href=\"https:\/\/bureau.ru\/books\/fff\/demo\/4\">What it means to get something done<\/a><\/li>\n<li><a href=\"https:\/\/doi.org\/10.1016\/j.cell.2025.12.049\">Hallmarks of Cancer<\/a><\/li>\n<li><a href=\"https:\/\/kinzhal.media\/meetings-boss\/\">How to start running meetings well<\/a><\/li>\n<\/ul>\n<h3>Delivered<\/h3>\n<ul>\n<li>Completed a project under the supervision of senior colleagues and presented the results<\/li>\n<\/ul>\n<h2>Bioinformatician<\/h2>\n<p>An independent operator. Handles routine tasks independently and complex ones with a manager’s support, without constant supervision. Finds or requests the information they need, tracks deadlines, and raises concerns in time.<\/p>\n<p>Has at least one project of their own: a client project, an internal R&D project, a Confluence space, a blog, a channel, or a product.<\/p>\n<p>Learns to be flexible: resolves tensions between deadlines and quality through ingenuity and flexible scope, rather than by sacrificing deadlines, quality, or personal time.<\/p>\n<p>Learns to verify rather than trust. Can establish and explain that their own analysis, or work by a colleague, another department, or an agent, is based on real data and appropriate methods and does not overlook the obvious.<\/p>\n<p>Cohort analysis is a core skill. Can obtain a cohort and its annotations, assess data quality, analyze genetic events, and put the results into a presentation.<\/p>\n<p>Is well read and well prepared. There is no such thing as a bioinformatician who has not read Hallmarks of Cancer or cannot write a skill for Claude.<\/p>\n<p>Is not afraid of unfamiliar work. Will run an agent, write code and submit a pull request, start a Docker container, inspect a BAM file in a genome browser, run differential expression analysis, or put together a presentation.<\/p>\n<p>Learns to negotiate: understands the client’s problem and their own capacity, agrees on deadlines, and resolves conflicts.<\/p>\n<h3>Read<\/h3>\n<ul>\n<li><a href=\"https:\/\/www.anthropic.com\/engineering\/building-effective-agents\">Building effective agents by Anthropic<\/a><\/li>\n<li><a href=\"https:\/\/docs.google.com\/document\/d\/1fCDQgB5f5bHzB8bruompWff7jGkCL4mCEjlfpBRuY_o\/edit?usp=sharing\">30 selected papers<\/a> — make your own copy, mark what you have read, and use it for notes<\/li>\n<li><a href=\"https:\/\/clauswilke.com\/dataviz\/\">Claus O. Wilke. Fundamentals of Data Visualization<\/a><\/li>\n<\/ul>\n<h3>Studied<\/h3>\n<ul>\n<li>Working with an AI agent on a real task: defining the task, checking the result, and finding an error<\/li>\n<li>How bioinformatics tools work and where their limits lie: BWA, Strelka, Sequenza, kallisto, STAR, and STAR-Fusion<\/li>\n<li>Sequencing quality metrics<\/li>\n<\/ul>\n<h3>Delivered<\/h3>\n<ul>\n<li>Contributed to two projects led by senior colleagues<\/li>\n<li>Independently completed a research or cohort analysis project from start to finish<\/li>\n<li>Built something together with developers, taking ownership of the work<\/li>\n<li>Gave a lecture on an educational course<\/li>\n<\/ul>\n<h3>Collected<\/h3>\n<ul>\n<li>20 observations<\/li>\n<\/ul>\n<h2>Lead bioinformatician<\/h2>\n<p>Leads the team of independent operators. Represents bioinformatics to the client, from the initial request to the final result. Handles all aspects of the client relationship and takes responsibility for them.<\/p>\n<p>Knows the deadlines, writes client proposals, plans iterations, and manages the project. Solves problems by designing and building a real product. Takes responsibility for its usefulness, delivery date, and quality. Answers for every chart and comma in a paper or product. Finds flexible solutions.<\/p>\n<p>Chooses and supervises the people doing the work within the department and delegates tasks. Has projects of their own, a blog or channel, and maintains a section in Confluence.<\/p>\n<h3>Read<\/h3>\n<ul>\n<li>Genrikh Altshuller. Find an Idea<\/li>\n<\/ul>\n<h3>Managed<\/h3>\n<ul>\n<li>Independently explained a product feature to another department — developers, biologists, or clinicians — and handed it over for implementation<\/li>\n<li>Reviewed work by a specialist in another department, wrote feedback, checked the revisions, and accepted the result. In other words, made sure the work was done properly<\/li>\n<li>Independently sorted out a messy situation with another department: understood the conflict, motivated people, and resolved the organizational problem<\/li>\n<\/ul>\n<h3>Delivered<\/h3>\n<ul>\n<li>Attended five meetings with new clients and asked questions<\/li>\n<li>Wrote five task briefs and got them approved by the clients<\/li>\n<li>Delivered five practical tasks under their personal responsibility: presented an analysis to a physician, submitted work to a conference, or released a component update<\/li>\n<li>Solved a problem through other people’s work<\/li>\n<\/ul>\n<h3>Developed<\/h3>\n<ul>\n<li>Helped a trainee become a bioinformatician. Preferably two, or even three<\/li>\n<\/ul>\n<h3>Shared knowledge<\/h3>\n<ul>\n<li>Gave a public lecture<\/li>\n<\/ul>\n<h3>Standardized<\/h3>\n<ul>\n<li>Improved three things in the bioinformatics department’s work and agreed new procedures, standards, or principles<\/li>\n<li>Got five templates for replies, mailings, or pages approved<\/li>\n<\/ul>\n<h3>Collected<\/h3>\n<ul>\n<li>100 observations<\/li>\n<\/ul>\n<h2>Director<\/h2>\n<p>The defining quality is effectiveness: always knows what result is needed and how to achieve it. A director has many tasks, projects, employees, clients, and other concerns, so allocates time to influence everything that matters to the result.<\/p>\n<p>An entrepreneur and strategist. Rather than solving a task as a team lead does, sets tasks in a sequence that produces the greatest effect with the least waste.<\/p>\n<p>A standard setter. Improves procedures, processes, and tools. Each improvement benefits many projects.<\/p>\n<p>A magician. Helps managers and teams with simple, effective advice that makes them say, “Wait, we could do that?” Does not keep track of deadlines unless the lead bioinformatician brings them up. Is the final quality check, but not the main one: that responsibility belongs to the lead bioinformatician. Does not put up with crap and keeps pushing back against deadlines — and, in effect, against the lead bioinformatician.<\/p>\n",
            "date_published": "2026-10-18T16:11:53+04:00",
            "date_modified": "2026-10-18T16:11:39+04:00",
            "tags": [
                "bioinformatics",
                "portfolio"
            ],
            "_date_published_rfc2822": "Sun, 18 Oct 2026 16:11:53 +0400",
            "_rss_guid_is_permalink": "false",
            "_rss_guid": "46",
            "_rss_enclosures": [],
            "_e2_data": {
                "is_favourite": false,
                "links_required": [],
                "og_images": []
            }
        },
        {
            "id": "43",
            "url": "https:\/\/kostachernyshov.com\/blog\/all\/publications-in-2026\/",
            "title": "Publications in 2026",
            "content_html": "<p>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.<\/p>\n<h2>American Association for Cancer Research<\/h2>\n<p><b>“A comparative scoring framework for BCMA-, GPRC5D- and CD38-targeted therapies in multiple myeloma”<\/b><\/p>\n<p>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.<\/p>\n<p>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.<\/p>\n<div class=\"e2-text-picture\">\n<div class=\"fotorama\" data-width=\"2560\" data-ratio=\"1.3333333333333\">\n<img src=\"https:\/\/kostachernyshov.com\/blog\/pictures\/MM3.jpg\" width=\"2560\" height=\"1920\" alt=\"\" \/>\n<img src=\"https:\/\/kostachernyshov.com\/blog\/pictures\/MM.jpg\" width=\"2560\" height=\"1672\" alt=\"\" \/>\n<img src=\"https:\/\/kostachernyshov.com\/blog\/pictures\/MM2.png\" width=\"1288\" height=\"1494\" alt=\"\" \/>\n<\/div>\n<div class=\"e2-text-caption\"><a href=\"https:\/\/bostongene.com\/news-and-publications\/publications\/a-comparative-scoring-framework-for-bcma-gprc5d-and-cd38-targeted-therapies-in-multiple-myeloma\">BostonGene news<\/a>. <a href=\"https:\/\/aacrjournals.org\/cancerres\/article\/86\/7_Supplement\/2437\/779071\/Abstract-2437-A-comparative-scoring-framework-for\">AACR abstract<\/a>.<\/div>\n<\/div>\n<p>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. <a href=\"https:\/\/bostongene.com\/news-and-publications\/publications\/an-actionability-score-for-bcma-gprc5d-and-cd38-targeted-therapy-selection-in-multiple-myeloma\">Updated poster and abstract<\/a>.<\/p>\n<p><b>“Integrative application of Lymphly reveals shared molecular landscape across B-cell lymphomas”<\/b><\/p>\n<p>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.<\/p>\n<div class=\"e2-text-picture\">\n<div class=\"fotorama\" data-width=\"1920\" data-ratio=\"0.75\">\n<img src=\"https:\/\/kostachernyshov.com\/blog\/pictures\/lymphly-3.jpg\" width=\"1920\" height=\"2560\" alt=\"\" \/>\n<img src=\"https:\/\/kostachernyshov.com\/blog\/pictures\/lymphly-1.jpg\" width=\"2560\" height=\"1665\" alt=\"\" \/>\n<img src=\"https:\/\/kostachernyshov.com\/blog\/pictures\/lymply-2.png\" width=\"884\" height=\"626\" alt=\"\" \/>\n<\/div>\n<div class=\"e2-text-caption\"><a href=\"https:\/\/bostongene.com\/news-and-publications\/publications\/integrative-application-of-lymphly-reveals-shared-molecular-landscape-across-b-cell-lymphomas\">BostonGene news<\/a>. <a href=\"https:\/\/aacrjournals.org\/cancerres\/article\/86\/7_Supplement\/3877\/776908\">AACR abstract<\/a>.<\/div>\n<\/div>\n<p>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. <a href=\"https:\/\/bostongene.com\/news-and-publications\/publications\/cross-disease-stratification-by-genetic-classifier-lymphly-reveals-shared-molecular-programs-across-b-cell-malignancies\">Poster and abstract<\/a>.<\/p>\n<h2>American Society of Clinical Oncology<\/h2>\n<p class=\"foot\">The main clinical oncology conference in the US.<\/p>\n<p><b>“Examining the tumor microbiology and microenvironment of patients with triple negative inflammatory breast cancer receiving Keynote 522”<\/b><\/p>\n<p>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.<\/p>\n<p><a href=\"https:\/\/bostongene.com\/news-and-publications\/publications\/examining-the-tumor-microbiology-and-microenvironment-of-patients-with-triple-negative-inflammatory-breast-cancer-receiving-keynote-522\">ASCO abstract on BostonGene<\/a>.<\/p>\n<h2>European Hematology Association<\/h2>\n<p class=\"foot\">Europe’s leading hematology conference.<\/p>\n<p><b>“Risk-stratified patient selection in multiple myeloma for frontline treatment decision-making based on a transcriptomic classifier”<\/b><\/p>\n<p>We combined the ideas behind two transcriptomic multiple myeloma classifiers and built a consensus risk model.<\/p>\n<div class=\"e2-text-picture\">\n<div class=\"fotorama\" data-width=\"2560\" data-ratio=\"1.486643437863\">\n<img src=\"https:\/\/kostachernyshov.com\/blog\/pictures\/MM4.jpg\" width=\"2560\" height=\"1722\" alt=\"\" \/>\n<img src=\"https:\/\/kostachernyshov.com\/blog\/pictures\/MM5.jpg\" width=\"2560\" height=\"1182\" alt=\"\" \/>\n<\/div>\n<div class=\"e2-text-caption\"><a href=\"https:\/\/bostongene.com\/news-and-publications\/publications\/risk-stratified-patient-selection-in-multiple-myeloma-for-frontline-treatment-decision-making-based-on-a-transcriptomic-classifier\">BostonGene news<\/a>. <a href=\"https:\/\/library.ehaweb.org\/eha\/2026\/eha-2026\/4208381\/nikita.kotlov.risk-stratified.patient.selection.in.multiple.myeloma.for.html\">EHA abstract<\/a>.<\/div>\n<\/div>\n<p><b>“Overlaying TP53 loss and aneuploidy with distinct diffuse large B-cell lymphoma subtypes refines tumor classification and enhances risk stratification”<\/b><\/p>\n<p>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.<\/p>\n<div class=\"e2-text-picture\">\n<div class=\"fotorama\" data-width=\"2560\" data-ratio=\"1.4780600461894\">\n<img src=\"https:\/\/kostachernyshov.com\/blog\/pictures\/lymphly-5.jpg\" width=\"2560\" height=\"1732\" alt=\"\" \/>\n<img src=\"https:\/\/kostachernyshov.com\/blog\/pictures\/lymphly-6.png\" width=\"1116\" height=\"1382\" alt=\"\" \/>\n<\/div>\n<div class=\"e2-text-caption\"><a href=\"https:\/\/bostongene.com\/news-and-publications\/publications\/overlaying-tp53-loss-and-aneuploidy-with-distinct-diffuse-large-b-cell-lymphoma-subtypes-refines-tumor-classification-and-enhances-risk-stratification\">BostonGene news<\/a>.<\/div>\n<\/div>\n",
            "date_published": "2026-07-11T20:16:59+04:00",
            "date_modified": "2026-10-18T16:14:18+04:00",
            "tags": [
                "2026",
                "bioinformatics",
                "portfolio"
            ],
            "image": "https:\/\/kostachernyshov.com\/blog\/pictures\/MM3.jpg",
            "_date_published_rfc2822": "Sat, 11 Jul 2026 20:16:59 +0400",
            "_rss_guid_is_permalink": "false",
            "_rss_guid": "43",
            "_rss_enclosures": [],
            "_e2_data": {
                "is_favourite": false,
                "links_required": [
                    "jquery\/jquery.js",
                    "fotorama\/fotorama.css",
                    "fotorama\/fotorama.js"
                ],
                "og_images": [
                    "https:\/\/kostachernyshov.com\/blog\/pictures\/MM3.jpg",
                    "https:\/\/kostachernyshov.com\/blog\/pictures\/MM.jpg",
                    "https:\/\/kostachernyshov.com\/blog\/pictures\/MM2.png",
                    "https:\/\/kostachernyshov.com\/blog\/pictures\/lymphly-3.jpg",
                    "https:\/\/kostachernyshov.com\/blog\/pictures\/lymphly-1.jpg",
                    "https:\/\/kostachernyshov.com\/blog\/pictures\/lymply-2.png",
                    "https:\/\/kostachernyshov.com\/blog\/pictures\/MM4.jpg",
                    "https:\/\/kostachernyshov.com\/blog\/pictures\/MM5.jpg",
                    "https:\/\/kostachernyshov.com\/blog\/pictures\/lymphly-5.jpg",
                    "https:\/\/kostachernyshov.com\/blog\/pictures\/lymphly-6.png"
                ]
            }
        },
        {
            "id": "38",
            "url": "https:\/\/kostachernyshov.com\/blog\/all\/geomx-and-mycosis-fungoides\/",
            "title": "GeoMx and mycosis fungoides",
            "content_html": "<h2>Project on lymphoma<\/h2>\n<p>Mycosis fungoides, or cutaneous T-cell lymphoma, is a cancer in which deranged T lymphocytes start proliferating in the skin. The client carried out the wet part of the study using GeoMx technology and ordered the analysis from us at BostonGene. I led the project, worked out the annotations and the nuances of the technology and communicated with the client, while Lesha did the quality control and analysis.<\/p>\n<h2>GeoMx technology and files<\/h2>\n<p>Original article: <a href=\"https:\/\/www.nature.com\/articles\/s41587-020-0472-9\">Multiplex digital spatial profiling of proteins and RNA in fixed tissue, 2020<\/a><\/p>\n<div class=\"e2-text-picture\">\n<div class=\"fotorama\" data-width=\"721\" data-ratio=\"0.78369565217391\">\n<img src=\"https:\/\/kostachernyshov.com\/blog\/pictures\/geomx_технология.png\" width=\"721\" height=\"920\" alt=\"\" \/>\n<img src=\"https:\/\/kostachernyshov.com\/blog\/pictures\/ROI_на_слайде.png\" width=\"1101\" height=\"1280\" alt=\"\" \/>\n<\/div>\n<div class=\"e2-text-caption\">The essence of GeoMx technology and an image of a piece of skin from the study and the ROIs (regions of interest) on this piece.<\/div>\n<\/div>\n<p>The essence of GeoMx technology:<\/p>\n<ol start=\"1\">\n<li>There is a pre-prepared set of RNA probes for thousands of genes. These probes are connected by a UV linker to a barcode oligonucleotide. One gene we care about = one barcode.<\/li>\n<li>The biopsy is prepared, placed on a slide, treated with RNA probes, and hybridization takes place.<\/li>\n<li>Then the slide is placed into a machine with a camera, and on the microphoto shown on the machine’s screen, the researcher marks the region of interest.<\/li>\n<li>Then the machine blasts this outlined region with UV light, the UV linker falls apart, and the machine sucks up the barcodes.<\/li>\n<li>The barcodes are sequenced, even on a regular sequencer, and reads are produced.<\/li>\n<li>The reads are matched against a pre-prepared barcode–gene panel. The number of reads from each barcode is basically transformed into the signal intensity of a particular RNA.<\/li>\n<\/ol>\n<p>Therefore when we received raw sequencing files it became clear that this was not what we needed. We were sent oligo‑DNA barcodes that by themselves carry no information about genes and need to be matched to pairs of barcode‑gene via a special pipeline. The output files are Digital Count Conversion (DCC).<\/p>\n<div class=\"e2-text-picture\">\n<div class=\"fotorama\" data-width=\"742\" data-ratio=\"1.5888650963597\">\n<img src=\"https:\/\/kostachernyshov.com\/blog\/pictures\/мануал_geomx.png\" width=\"742\" height=\"467\" alt=\"\" \/>\n<img src=\"https:\/\/kostachernyshov.com\/blog\/pictures\/illumina_pipe_for_geomx.png\" width=\"1280\" height=\"720\" alt=\"\" \/>\n<\/div>\n<div class=\"e2-text-caption\">How DCC files are obtained from the GeoMx manual and the Illumina pipeline.<\/div>\n<\/div>\n<p>You can process fastq files yourself via Illumina’s BaseSpace <a href=\"https:\/\/www.illumina.com\/products\/by-type\/informatics-products\/basespace-sequence-hub\/apps\/nanostring-geomxr-ngs-pipeline.html\">GeoMx NGS Pipeline is distributed as a pipeline on BaseSpace<\/a>. It seems this doesn’t require special access, but it costs money and is not trivial at all. So I decided to write to the collaborator and ask if they already had DCC files. The DCC files were received the next day. Sometimes it’s easiest to simply ask.<\/p>\n<h2>Annotation<\/h2>\n<p>To ask the right questions about the cohort and then answer them you need to understand what you’re dealing with in the first place. In theory each patient may have several slides, and each slide may contain several ROIs. And patients should have different properties, something to compare with something.<\/p>\n<div class=\"e2-text-picture\">\n<img src=\"https:\/\/kostachernyshov.com\/blog\/pictures\/аннотация.png\" width=\"2014\" height=\"1071\" alt=\"\" \/>\n<div class=\"e2-text-caption\">The annotation to be grasped<\/div>\n<\/div>\n<p>After studying the table it became clear which dimensions are important:<\/p>\n<ul>\n<li>Patient — 1st dimension<\/li>\n<li>Diagnosis and Large‑cell transformation — 2nd dimension<\/li>\n<li>Segments — 3rd dimension<\/li>\n<li>ROI – region of interest or mask — 4th dimension<\/li>\n<\/ul>\n<p>Now this needs to be visualized.<br \/>\n<b>1st dimension.<\/b> We depicted the patient as a little man —<br \/>\n<b>2nd dimension.<\/b> We group patients by clinical parameters<\/p>\n<div class=\"e2-text-picture\">\n<img src=\"https:\/\/kostachernyshov.com\/blog\/pictures\/визуализация_аннотации1.png\" width=\"415\" height=\"172\" alt=\"\" \/>\n<div class=\"e2-text-caption\">It turned out that patients 4 and 9 are the same person. Since biopsies were taken at different times and from different locations, and the cohort is small, I decided to treat these two points as different patients.<\/div>\n<\/div>\n<p><b>3rd dimension.<\/b> In the annotation there are names “Epidermal”, “Sub‑epidermal”, “Dermis”. Therefore I decided that investigators took segments according to the principle “look at the skin at different depths”. I chose a picture of “skin in cross‑section” with the upper, middle and deep structures and hair follicles.<\/p>\n<div class=\"e2-text-picture\">\n<img src=\"https:\/\/kostachernyshov.com\/blog\/pictures\/кожа.png\" width=\"1215\" height=\"760\" alt=\"\" \/>\n<div class=\"e2-text-caption\">We placed the patients above the image of the skin and marked with circles which segments of what depth each patient has. We added the patient number to the skin and labeled it.<\/div>\n<\/div>\n<div class=\"e2-text-picture\">\n<img src=\"https:\/\/kostachernyshov.com\/blog\/pictures\/image-31.png\" width=\"665\" height=\"426\" alt=\"\" \/>\n<div class=\"e2-text-caption\">The arrangement of the patients is similar to a matrix so that patients 2 and 9, who have a segment from a follicle, are arranged above the follicle in the picture.<\/div>\n<\/div>\n<p>Now the data structure is clear. For example, only patient 5 has an Epidermal segment, or follicles appear only in two patients but they have different diagnoses.<\/p>\n<p>In the course of the exercise it turned out that segment names sometimes designated localization – “Epidermal”, and sometimes a property – “Macrotumor poor”. To clarify the annotation I prepared a table where categories for ROI were partially unified. We asked collaborators to fill in the orange cells and prepared a slide with an explanation of what was unclear.<\/p>\n<div class=\"e2-text-picture\">\n<img src=\"https:\/\/kostachernyshov.com\/blog\/pictures\/image-32.png\" width=\"1280\" height=\"721\" alt=\"\" \/>\n<div class=\"e2-text-caption\">Nice, isn’t it?<\/div>\n<\/div>\n<p>We wrote a letter to the client, explained the situation and asked them to clarify the annotation.<\/p>\n<h2>Analysis and results<\/h2>\n<p>Lesha immersed himself in the technology, selected tools for work and began the analysis: <a href=\"https:\/\/www.biorxiv.org\/content\/10.1101\/2022.05.26.493637v1\">Background modeling, Quality Control and Normalization for GeoMx RNA data with GeoDiff<\/a>. Having understood the technology, the necessary files and the structure of the cohort, the rest went without problems.<\/p>\n<p>We found that FMF has an overall inflammatory microenvironment with a predominance of components of innate immunity – T1 and M2 macrophages, NK cells and dendritic cells. And CMF is less inflamed, with a predominance of T‑helper and plasma cells.<\/p>\n<div class=\"e2-text-picture\">\n<img src=\"https:\/\/kostachernyshov.com\/blog\/pictures\/результаты.png\" width=\"1099\" height=\"559\" alt=\"\" \/>\n<div class=\"e2-text-caption\">Color coding is consistent in the left and right parts of the graphs and allows you to keep track of what was compared. One of hundreds of slides.<\/div>\n<\/div>\n<p>The client looked satisfied at the meeting, but then refused to pay us, and we forbade them to publish a paper with our analysis. We never learned the real reason. It happens.<\/p>\n",
            "date_published": "2026-05-03T23:25:59+04:00",
            "date_modified": "2026-05-03T23:23:19+04:00",
            "tags": [
                "2023",
                "bioinformatics",
                "portfolio"
            ],
            "image": "https:\/\/kostachernyshov.com\/blog\/pictures\/geomx_технология.png",
            "_date_published_rfc2822": "Sun, 03 May 2026 23:25:59 +0400",
            "_rss_guid_is_permalink": "false",
            "_rss_guid": "38",
            "_rss_enclosures": [],
            "_e2_data": {
                "is_favourite": false,
                "links_required": [
                    "jquery\/jquery.js",
                    "fotorama\/fotorama.css",
                    "fotorama\/fotorama.js"
                ],
                "og_images": [
                    "https:\/\/kostachernyshov.com\/blog\/pictures\/geomx_технология.png",
                    "https:\/\/kostachernyshov.com\/blog\/pictures\/ROI_на_слайде.png",
                    "https:\/\/kostachernyshov.com\/blog\/pictures\/мануал_geomx.png",
                    "https:\/\/kostachernyshov.com\/blog\/pictures\/illumina_pipe_for_geomx.png",
                    "https:\/\/kostachernyshov.com\/blog\/pictures\/аннотация.png",
                    "https:\/\/kostachernyshov.com\/blog\/pictures\/визуализация_аннотации1.png",
                    "https:\/\/kostachernyshov.com\/blog\/pictures\/кожа.png",
                    "https:\/\/kostachernyshov.com\/blog\/pictures\/image-31.png",
                    "https:\/\/kostachernyshov.com\/blog\/pictures\/image-32.png",
                    "https:\/\/kostachernyshov.com\/blog\/pictures\/результаты.png"
                ]
            }
        },
        {
            "id": "33",
            "url": "https:\/\/kostachernyshov.com\/blog\/all\/publications-in-2025\/",
            "title": "Publications in 2025",
            "content_html": "<p>In 2025, I lead a bioinformatics and NGS data analytics department. Over the year, our team supported 26 abstracts and 4 papers — we processed data, formulated and tested hypotheses, ran statistics, created figures, and wrote results and methods. Most of the abstracts were collaborations with clinics and universities, along with a few fully in-house studies. Participating in conferences helps us demonstrate our expertise, build new connections, and start new projects. Below are some of the most interesting examples.<\/p>\n<h2>Paper <a href=\"https:\/\/www.nature.com\/articles\/s41523-025-00723-0\" class=\"nu\">“<u>Breast Cancer Classifier...<\/u>”<\/a><\/h2>\n<p>There is a well-known breast cancer classification called “PAM50” — it was developed about thirty years ago and describes molecular-level differences and their relationship to patient prognosis. Everyone knows and uses it, but due to legal restrictions, we could not apply it directly. Our goal was to redesign a similar classification — conceptually close, no worse, and ideally better.<\/p>\n<p>I led the project: defined the problem, structured the narrative, broke it down into key blocks, selected the figures, grouped them into panels, and estimated the scope of work. The full analysis, writing, and final visualization were done by Polina Turova, the first author.<\/p>\n<p>We redesigned the classifier, showed that it performs on par with the original, introduced a new HER2-low subtype, and removed the Normal-like subtype. In 2023, this work was presented <a href=\"https:\/\/kostachernyshov.ru\/blog\/all\/publikacii-za-2023-god\/\">at the SABCS conference as an abstract and poster<\/a>, and in 2025 it was published in a solid journal.<\/p>\n<div class=\"e2-text-picture\">\n<div class=\"fotorama\" data-width=\"1280\" data-ratio=\"1.3333333333333\">\n<img src=\"https:\/\/kostachernyshov.com\/blog\/pictures\/image-23.png\" width=\"1280\" height=\"960\" alt=\"\" \/>\n<img src=\"https:\/\/kostachernyshov.com\/blog\/pictures\/image-24.png\" width=\"960\" height=\"1280\" alt=\"\" \/>\n<img src=\"https:\/\/kostachernyshov.com\/blog\/pictures\/image-25.png\" width=\"1853\" height=\"1034\" alt=\"\" \/>\n<img src=\"https:\/\/kostachernyshov.com\/blog\/pictures\/image-26.png\" width=\"1853\" height=\"1034\" alt=\"\" \/>\n<img src=\"https:\/\/kostachernyshov.com\/blog\/pictures\/image-27.png.jpg\" width=\"1964\" height=\"2560\" alt=\"\" \/>\n<img src=\"https:\/\/kostachernyshov.com\/blog\/pictures\/image-27.png-1.jpg\" width=\"1889\" height=\"2560\" alt=\"\" \/>\n<\/div>\n<div class=\"e2-text-caption\">To design the figures, I printed relevant papers and plots, arranged them physically on paper, then recreated them in Figma, describing what problem each figure and chart solves. After that, Polina filled the drafts with real data and finalized them. Images 5 and 6 are the final versions published in the journal.<\/div>\n<\/div>\n<h2>Paper <a href=\"https:\/\/www.nature.com\/articles\/s43856-025-00934-3\">”Validation of combined DNA and RNA analysis in oncology...”<\/a><\/h2>\n<p>This paper describes how BostonGene analyzes patient data. I contributed to one of the sections — proposing a methodology for analyzing genetic events in patients and how to visualize them.<\/p>\n<div class=\"e2-text-picture\">\n<div class=\"fotorama\" data-width=\"1911\" data-ratio=\"1.9381338742394\">\n<img src=\"https:\/\/kostachernyshov.com\/blog\/pictures\/image-27.png\" width=\"1911\" height=\"986\" alt=\"\" \/>\n<img src=\"https:\/\/kostachernyshov.com\/blog\/pictures\/image-28.png\" width=\"1895\" height=\"749\" alt=\"\" \/>\n<\/div>\n<div class=\"e2-text-caption\">Only clinically significant genetic events were included in the analysis, as defined by a dedicated BostonGene team. The visualization is an adaptation of <a href=\"https:\/\/www.nature.com\/articles\/s41586-020-1969-6\/figures\/2\">“Panorama of driver events” from another paper<\/a>.<\/div>\n<\/div>\n<h2>American Association for Cancer Research<\/h2>\n<p>A large oncology conference in the US, with a strong focus on translational medicine.<\/p>\n<div class=\"e2-text-picture\">\n<img src=\"https:\/\/kostachernyshov.com\/blog\/pictures\/image-37.png\" width=\"1280\" height=\"960\" alt=\"\" \/>\n<div class=\"e2-text-caption\">“Pan-cancer analysis of TRIM37 copy-number and development of fit-for-screening in situ hybridization tools”. <a href=\"https:\/\/www.abstractsonline.com\/pp8\/#!\/20273\/presentation\/4837\">AACR abstract book<\/a>.<\/div>\n<\/div>\n<h2>American Society of Clinical Oncology<\/h2>\n<p>The main clinical oncology conference in the US.<\/p>\n<div class=\"e2-text-picture\">\n<div class=\"fotorama\" data-width=\"1773\" data-ratio=\"1.6805687203791\">\n<img src=\"https:\/\/kostachernyshov.com\/blog\/pictures\/image-29.png\" width=\"1773\" height=\"1055\" alt=\"\" \/>\n<img src=\"https:\/\/kostachernyshov.com\/blog\/pictures\/image-30.png\" width=\"797\" height=\"878\" alt=\"\" \/>\n<\/div>\n<div class=\"e2-text-caption\">“Molecular subtyping and insights into sarcoma biology and prognosis”. <a href=\"https:\/\/ascopubs.org\/doi\/10.1200\/JCO.2025.43.16_suppl.11536\">ASCO abstract book<\/a>.<\/div>\n<\/div>\n<h2>American Society of Hematology<\/h2>\n<p>The largest conference on lymphomas and blood disorders in the world.<\/p>\n<div class=\"e2-text-picture\">\n<div class=\"fotorama\" data-width=\"1560\" data-ratio=\"1.5522388059701\">\n<img src=\"https:\/\/kostachernyshov.com\/blog\/pictures\/image-33.png\" width=\"1560\" height=\"1005\" alt=\"\" \/>\n<img src=\"https:\/\/kostachernyshov.com\/blog\/pictures\/image-34.png\" width=\"835\" height=\"971\" alt=\"\" \/>\n<\/div>\n<div class=\"e2-text-caption\">“Refining diffuse large B-cell lymphoma subtyping using lymphly, an evidence-based classifier”. <a href=\"https:\/\/bostongene.com\/news-and-publications\/publications\/refining-diffuse-large-b-cell-lymphoma-subtyping-using-lymphly-an-evidence-based-classifier\">BostonGene news<\/a>. <a href=\"https:\/\/doi.org\/10.1182\/blood-2025-2557\">ASH conference site<\/a>.<\/div>\n<\/div>\n<h2>San-Antonio Breast Cancer Symposium<\/h2>\n<p>The largest conference on breast cancer in the world.<\/p>\n<div class=\"e2-text-picture\">\n<div class=\"fotorama\" data-width=\"1782\" data-ratio=\"1.7802197802198\">\n<img src=\"https:\/\/kostachernyshov.com\/blog\/pictures\/image-35.png\" width=\"1782\" height=\"1001\" alt=\"\" \/>\n<img src=\"https:\/\/kostachernyshov.com\/blog\/pictures\/image-36.png\" width=\"1177\" height=\"770\" alt=\"\" \/>\n<\/div>\n<div class=\"e2-text-caption\">“Tumor-associated macrophage (TAMs) and cancer-associated fibroblasts (CAFs) profiles in invasive lobular carcinoma (ILC) vs no special type (NST)”. <a href=\"https:\/\/bostongene.com\/news-and-publications\/publications\/tumor-associated-macrophage-tams-and-cancer-associated-fibroblasts-cafs-profiles-in-invasive-lobular-carcinoma-ilc-vs-no-special-type-nst\">BostonGene news<\/a>.<\/div>\n<\/div>\n",
            "date_published": "2026-04-29T05:27:42+04:00",
            "date_modified": "2026-04-29T05:15:22+04:00",
            "tags": [
                "2025",
                "bioinformatics",
                "portfolio"
            ],
            "image": "https:\/\/kostachernyshov.com\/blog\/pictures\/image-23.png",
            "_date_published_rfc2822": "Wed, 29 Apr 2026 05:27:42 +0400",
            "_rss_guid_is_permalink": "false",
            "_rss_guid": "33",
            "_rss_enclosures": [],
            "_e2_data": {
                "is_favourite": false,
                "links_required": [
                    "jquery\/jquery.js",
                    "fotorama\/fotorama.css",
                    "fotorama\/fotorama.js"
                ],
                "og_images": [
                    "https:\/\/kostachernyshov.com\/blog\/pictures\/image-23.png",
                    "https:\/\/kostachernyshov.com\/blog\/pictures\/image-24.png",
                    "https:\/\/kostachernyshov.com\/blog\/pictures\/image-25.png",
                    "https:\/\/kostachernyshov.com\/blog\/pictures\/image-26.png",
                    "https:\/\/kostachernyshov.com\/blog\/pictures\/image-27.png.jpg",
                    "https:\/\/kostachernyshov.com\/blog\/pictures\/image-27.png-1.jpg",
                    "https:\/\/kostachernyshov.com\/blog\/pictures\/image-27.png",
                    "https:\/\/kostachernyshov.com\/blog\/pictures\/image-28.png",
                    "https:\/\/kostachernyshov.com\/blog\/pictures\/image-37.png",
                    "https:\/\/kostachernyshov.com\/blog\/pictures\/image-29.png",
                    "https:\/\/kostachernyshov.com\/blog\/pictures\/image-30.png",
                    "https:\/\/kostachernyshov.com\/blog\/pictures\/image-33.png",
                    "https:\/\/kostachernyshov.com\/blog\/pictures\/image-34.png",
                    "https:\/\/kostachernyshov.com\/blog\/pictures\/image-35.png",
                    "https:\/\/kostachernyshov.com\/blog\/pictures\/image-36.png"
                ]
            }
        },
        {
            "id": "32",
            "url": "https:\/\/kostachernyshov.com\/blog\/all\/aacr-2026-the-next-step-for-foundation-models-toward-real-world\/",
            "title": "AACR 2026. The next step for foundation models toward real-world products",
            "content_html": "<div class=\"e2-text-picture\">\n<img src=\"https:\/\/kostachernyshov.com\/blog\/pictures\/image-75.png\" width=\"1704\" height=\"550\" alt=\"\" \/>\n<div class=\"e2-text-caption\">the path from embedding to product is long and difficult<\/div>\n<\/div>\n<p>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.<\/p>\n<p>I gathered some thoughts from the AACR 2026 and formulated like this:<\/p>\n<p><b>How to use embeddings and imputation → to get one step closer to a product.<\/b><\/p>\n<h2>Synthetic histology slides → against class imbalance<\/h2>\n<p>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.<\/p>\n<p>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.<\/p>\n<p><a href=\"https:\/\/www.biorxiv.org\/content\/10.64898\/2025.12.29.692457v1.full\">The paper about the BERGERON and CRAWFORD models<\/a> from Stanford.<\/p>\n<div class=\"e2-text-picture\">\n<div class=\"fotorama\" data-width=\"1024\" data-ratio=\"1.3333333333333\">\n<img src=\"https:\/\/kostachernyshov.com\/blog\/pictures\/image-63.png\" width=\"1024\" height=\"768\" alt=\"\" \/>\n<img src=\"https:\/\/kostachernyshov.com\/blog\/pictures\/image-64.png\" width=\"1024\" height=\"768\" alt=\"\" \/>\n<img src=\"https:\/\/kostachernyshov.com\/blog\/pictures\/image-65.png\" width=\"1024\" height=\"768\" alt=\"\" \/>\n<\/div>\n<div class=\"e2-text-caption\">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.<\/div>\n<\/div>\n<h2>Attention analysis → for finding contextual, model-derived hypotheses<\/h2>\n<p>Model attention analysis is the calculation of features that the model relies on more strongly than others when making a decision.<\/p>\n<p>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.<\/p>\n<div class=\"e2-text-picture\">\n<div class=\"fotorama\" data-width=\"1024\" data-ratio=\"1.3333333333333\">\n<img src=\"https:\/\/kostachernyshov.com\/blog\/pictures\/image-66.png\" width=\"1024\" height=\"768\" alt=\"\" \/>\n<img src=\"https:\/\/kostachernyshov.com\/blog\/pictures\/image-67.png\" width=\"1024\" height=\"768\" alt=\"\" \/>\n<img src=\"https:\/\/kostachernyshov.com\/blog\/pictures\/image-68.png\" width=\"1024\" height=\"768\" alt=\"\" \/>\n<img src=\"https:\/\/kostachernyshov.com\/blog\/pictures\/image-69.png\" width=\"1024\" height=\"768\" alt=\"\" \/>\n<\/div>\n<div class=\"e2-text-caption\">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.<\/div>\n<\/div>\n<h2>Benchmarking pressure → for directed model development<\/h2>\n<p>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.<\/p>\n<p>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.<\/p>\n<h3>Multimodality → works better than a single modality<\/h3>\n<p>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.<\/p>\n<div class=\"e2-text-picture\">\n<img src=\"https:\/\/kostachernyshov.com\/blog\/pictures\/image-70.png\" width=\"1024\" height=\"768\" alt=\"\" \/>\n<\/div>\n<h3>More data → the model works better<\/h3>\n<p>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.<\/p>\n<div class=\"e2-text-picture\">\n<img src=\"https:\/\/kostachernyshov.com\/blog\/pictures\/image-71.png\" width=\"1024\" height=\"768\" alt=\"\" \/>\n<\/div>\n<h3>Text descriptions of slides → for better quality with a smaller training set<\/h3>\n<p>Yana Lipkova gave a review of benchmark reviews for pathology models.<\/p>\n<p>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.<\/p>\n<div class=\"e2-text-picture\">\n<img src=\"https:\/\/kostachernyshov.com\/blog\/pictures\/image-72.png\" width=\"1024\" height=\"768\" alt=\"\" \/>\n<div class=\"e2-text-caption\">CONCH<\/div>\n<\/div>\n<h3>Working with context and autoregression → for better protein function prediction<\/h3>\n<p>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.<\/p>\n<p>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.<\/p>\n<div class=\"e2-text-picture\">\n<div class=\"fotorama\" data-width=\"1024\" data-ratio=\"1.3333333333333\">\n<img src=\"https:\/\/kostachernyshov.com\/blog\/pictures\/image-73.png\" width=\"1024\" height=\"768\" alt=\"\" \/>\n<img src=\"https:\/\/kostachernyshov.com\/blog\/pictures\/image-74.png\" width=\"1024\" height=\"768\" alt=\"\" \/>\n<\/div>\n<div class=\"e2-text-caption\">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.<\/div>\n<\/div>\n",
            "date_published": "2026-04-28T01:03:22+04:00",
            "date_modified": "2026-04-28T00:58:20+04:00",
            "tags": [
                "2026",
                "bioinformatics"
            ],
            "image": "https:\/\/kostachernyshov.com\/blog\/pictures\/image-75.png",
            "_date_published_rfc2822": "Tue, 28 Apr 2026 01:03:22 +0400",
            "_rss_guid_is_permalink": "false",
            "_rss_guid": "32",
            "_rss_enclosures": [],
            "_e2_data": {
                "is_favourite": false,
                "links_required": [
                    "jquery\/jquery.js",
                    "fotorama\/fotorama.css",
                    "fotorama\/fotorama.js"
                ],
                "og_images": [
                    "https:\/\/kostachernyshov.com\/blog\/pictures\/image-75.png",
                    "https:\/\/kostachernyshov.com\/blog\/pictures\/image-63.png",
                    "https:\/\/kostachernyshov.com\/blog\/pictures\/image-64.png",
                    "https:\/\/kostachernyshov.com\/blog\/pictures\/image-65.png",
                    "https:\/\/kostachernyshov.com\/blog\/pictures\/image-66.png",
                    "https:\/\/kostachernyshov.com\/blog\/pictures\/image-67.png",
                    "https:\/\/kostachernyshov.com\/blog\/pictures\/image-68.png",
                    "https:\/\/kostachernyshov.com\/blog\/pictures\/image-69.png",
                    "https:\/\/kostachernyshov.com\/blog\/pictures\/image-70.png",
                    "https:\/\/kostachernyshov.com\/blog\/pictures\/image-71.png",
                    "https:\/\/kostachernyshov.com\/blog\/pictures\/image-72.png",
                    "https:\/\/kostachernyshov.com\/blog\/pictures\/image-73.png",
                    "https:\/\/kostachernyshov.com\/blog\/pictures\/image-74.png"
                ]
            }
        },
        {
            "id": "31",
            "url": "https:\/\/kostachernyshov.com\/blog\/all\/aacr-2026-trends\/",
            "title": "AACR 2026 trends",
            "content_html": "<div class=\"e2-text-picture\">\n<img src=\"https:\/\/kostachernyshov.com\/blog\/pictures\/aacr-general-scheme.png\" width=\"2332\" height=\"786\" alt=\"\" \/>\n<\/div>\n<p>AACR is a large annual oncology conference. This year it took place in San Diego, and 22,000 participants attended the conference.<\/p>\n<p>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.<\/p>\n<p>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.<\/p>\n<div class=\"e2-text-picture\">\n<img src=\"https:\/\/kostachernyshov.com\/blog\/pictures\/booth.png\" width=\"1024\" height=\"768\" alt=\"\" \/>\n<div class=\"e2-text-caption\">BostonGene came to the conference<\/div>\n<\/div>\n<h2>Conference Opening<\/h2>\n<p>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?<\/p>\n<div class=\"e2-text-picture\">\n<img src=\"https:\/\/kostachernyshov.com\/blog\/pictures\/prostate-cancer.png\" width=\"1024\" height=\"768\" alt=\"\" \/>\n<div class=\"e2-text-caption\">Challenges in oncology still exist. Prostate cancer develops resistance to AR inhibitors by changing its histology.<\/div>\n<\/div>\n<div class=\"e2-text-picture\">\n<img src=\"https:\/\/kostachernyshov.com\/blog\/pictures\/image-38.png\" width=\"1024\" height=\"768\" alt=\"\" \/>\n<div class=\"e2-text-caption\">Drug development is not standing still. CAR-T therapies for solid tumors are being actively tested.<\/div>\n<\/div>\n<div class=\"e2-text-picture\">\n<div class=\"fotorama\" data-width=\"1024\" data-ratio=\"1.3333333333333\">\n<img src=\"https:\/\/kostachernyshov.com\/blog\/pictures\/image-39.png\" width=\"1024\" height=\"768\" alt=\"\" \/>\n<img src=\"https:\/\/kostachernyshov.com\/blog\/pictures\/image-40.png\" width=\"1024\" height=\"768\" alt=\"\" \/>\n<img src=\"https:\/\/kostachernyshov.com\/blog\/pictures\/image-41.png\" width=\"1024\" height=\"768\" alt=\"\" \/>\n<img src=\"https:\/\/kostachernyshov.com\/blog\/pictures\/image-42.png\" width=\"1024\" height=\"768\" alt=\"\" \/>\n<\/div>\n<div class=\"e2-text-caption\">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.<\/div>\n<\/div>\n<h2>News from the World of Large Biological Models<\/h2>\n<p>The session was chaired by Jakob Nikolas Kather, a rising star and professor of clinical AI at the University of Dresden, Germany.<\/p>\n<div class=\"e2-text-picture\">\n<img src=\"https:\/\/kostachernyshov.com\/blog\/pictures\/image-43.png\" width=\"1024\" height=\"768\" alt=\"\" \/>\n<div class=\"e2-text-caption\">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.<\/div>\n<\/div>\n<p>Faisal Mahmood, the father of AI in pathology, presented several new models from his lab.<\/p>\n<div class=\"e2-text-picture\">\n<div class=\"fotorama\" data-width=\"1024\" data-ratio=\"1.3333333333333\">\n<img src=\"https:\/\/kostachernyshov.com\/blog\/pictures\/image-44.png\" width=\"1024\" height=\"768\" alt=\"\" \/>\n<img src=\"https:\/\/kostachernyshov.com\/blog\/pictures\/image-45.png\" width=\"1024\" height=\"768\" alt=\"\" \/>\n<\/div>\n<div class=\"e2-text-caption\">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.<\/div>\n<\/div>\n<div class=\"e2-text-picture\">\n<div class=\"fotorama\" data-width=\"1024\" data-ratio=\"1.3333333333333\">\n<img src=\"https:\/\/kostachernyshov.com\/blog\/pictures\/image-46.png\" width=\"1024\" height=\"768\" alt=\"\" \/>\n<img src=\"https:\/\/kostachernyshov.com\/blog\/pictures\/image-47.png\" width=\"1024\" height=\"768\" alt=\"\" \/>\n<img src=\"https:\/\/kostachernyshov.com\/blog\/pictures\/image-48.png\" width=\"1024\" height=\"768\" alt=\"\" \/>\n<img src=\"https:\/\/kostachernyshov.com\/blog\/pictures\/image-49.png\" width=\"1024\" height=\"768\" alt=\"\" \/>\n<\/div>\n<div class=\"e2-text-caption\">A new histology model, a new model for histology + transcriptomics, a new model for histology + genomics, and a new model for spatial transcriptomics.<\/div>\n<\/div>\n<p>Bo Wang, the creator of scGPT, also presented two new models from his lab.<\/p>\n<div class=\"e2-text-picture\">\n<div class=\"fotorama\" data-width=\"1024\" data-ratio=\"1.3333333333333\">\n<img src=\"https:\/\/kostachernyshov.com\/blog\/pictures\/image-50.png\" width=\"1024\" height=\"768\" alt=\"\" \/>\n<img src=\"https:\/\/kostachernyshov.com\/blog\/pictures\/image-51.png\" width=\"1024\" height=\"768\" alt=\"\" \/>\n<\/div>\n<div class=\"e2-text-caption\">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.<\/div>\n<\/div>\n<p>Jure Leskovec presented Biomni — an agentic system designed for multi-step work on biological questions. We should test it.<\/p>\n<div class=\"e2-text-picture\">\n<div class=\"fotorama\" data-width=\"1024\" data-ratio=\"1.3333333333333\">\n<img src=\"https:\/\/kostachernyshov.com\/blog\/pictures\/image-52.png\" width=\"1024\" height=\"768\" alt=\"\" \/>\n<img src=\"https:\/\/kostachernyshov.com\/blog\/pictures\/image-53.png\" width=\"1024\" height=\"768\" alt=\"\" \/>\n<img src=\"https:\/\/kostachernyshov.com\/blog\/pictures\/image-54.png\" width=\"1024\" height=\"768\" alt=\"\" \/>\n<img src=\"https:\/\/kostachernyshov.com\/blog\/pictures\/image-55.png\" width=\"1024\" height=\"768\" alt=\"\" \/>\n<\/div>\n<div class=\"e2-text-caption\">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.<\/div>\n<\/div>\n<h2>Poster Session<\/h2>\n<p>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.<\/p>\n<div class=\"e2-text-picture\">\n<div class=\"fotorama\" data-width=\"1024\" data-ratio=\"1.3333333333333\">\n<img src=\"https:\/\/kostachernyshov.com\/blog\/pictures\/image-56.png\" width=\"1024\" height=\"768\" alt=\"\" \/>\n<img src=\"https:\/\/kostachernyshov.com\/blog\/pictures\/image-57.png\" width=\"1024\" height=\"768\" alt=\"\" \/>\n<img src=\"https:\/\/kostachernyshov.com\/blog\/pictures\/image-58.png\" width=\"1024\" height=\"768\" alt=\"\" \/>\n<img src=\"https:\/\/kostachernyshov.com\/blog\/pictures\/image-59.png\" width=\"1024\" height=\"768\" alt=\"\" \/>\n<\/div>\n<div class=\"e2-text-caption\">1, 2 — immunology and experimental therapy; 3 — AI section; 4 — clinical research: prognostic biomarkers, molecular cancer classifiers, population studies, liquid biopsy.<\/div>\n<\/div>\n<p>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.<\/p>\n<div class=\"e2-text-picture\">\n<div class=\"fotorama\" data-width=\"1024\" data-ratio=\"1.3333333333333\">\n<img src=\"https:\/\/kostachernyshov.com\/blog\/pictures\/image-60.png\" width=\"1024\" height=\"768\" alt=\"\" \/>\n<img src=\"https:\/\/kostachernyshov.com\/blog\/pictures\/image-61.png\" width=\"768\" height=\"1024\" alt=\"\" \/>\n<\/div>\n<\/div>\n<p>The main conclusion is that we can work on this area where AI meets the trend of advancing new types of therapies.<\/p>\n",
            "date_published": "2026-04-28T00:48:23+04:00",
            "date_modified": "2026-04-28T00:47:19+04:00",
            "tags": [
                "2026",
                "bioinformatics"
            ],
            "image": "https:\/\/kostachernyshov.com\/blog\/pictures\/aacr-general-scheme.png",
            "_date_published_rfc2822": "Tue, 28 Apr 2026 00:48:23 +0400",
            "_rss_guid_is_permalink": "false",
            "_rss_guid": "31",
            "_rss_enclosures": [],
            "_e2_data": {
                "is_favourite": false,
                "links_required": [
                    "jquery\/jquery.js",
                    "fotorama\/fotorama.css",
                    "fotorama\/fotorama.js"
                ],
                "og_images": [
                    "https:\/\/kostachernyshov.com\/blog\/pictures\/aacr-general-scheme.png",
                    "https:\/\/kostachernyshov.com\/blog\/pictures\/booth.png",
                    "https:\/\/kostachernyshov.com\/blog\/pictures\/prostate-cancer.png",
                    "https:\/\/kostachernyshov.com\/blog\/pictures\/image-38.png",
                    "https:\/\/kostachernyshov.com\/blog\/pictures\/image-39.png",
                    "https:\/\/kostachernyshov.com\/blog\/pictures\/image-40.png",
                    "https:\/\/kostachernyshov.com\/blog\/pictures\/image-41.png",
                    "https:\/\/kostachernyshov.com\/blog\/pictures\/image-42.png",
                    "https:\/\/kostachernyshov.com\/blog\/pictures\/image-43.png",
                    "https:\/\/kostachernyshov.com\/blog\/pictures\/image-44.png",
                    "https:\/\/kostachernyshov.com\/blog\/pictures\/image-45.png",
                    "https:\/\/kostachernyshov.com\/blog\/pictures\/image-46.png",
                    "https:\/\/kostachernyshov.com\/blog\/pictures\/image-47.png",
                    "https:\/\/kostachernyshov.com\/blog\/pictures\/image-48.png",
                    "https:\/\/kostachernyshov.com\/blog\/pictures\/image-49.png",
                    "https:\/\/kostachernyshov.com\/blog\/pictures\/image-50.png",
                    "https:\/\/kostachernyshov.com\/blog\/pictures\/image-51.png",
                    "https:\/\/kostachernyshov.com\/blog\/pictures\/image-52.png",
                    "https:\/\/kostachernyshov.com\/blog\/pictures\/image-53.png",
                    "https:\/\/kostachernyshov.com\/blog\/pictures\/image-54.png",
                    "https:\/\/kostachernyshov.com\/blog\/pictures\/image-55.png",
                    "https:\/\/kostachernyshov.com\/blog\/pictures\/image-56.png",
                    "https:\/\/kostachernyshov.com\/blog\/pictures\/image-57.png",
                    "https:\/\/kostachernyshov.com\/blog\/pictures\/image-58.png",
                    "https:\/\/kostachernyshov.com\/blog\/pictures\/image-59.png",
                    "https:\/\/kostachernyshov.com\/blog\/pictures\/image-60.png",
                    "https:\/\/kostachernyshov.com\/blog\/pictures\/image-61.png"
                ]
            }
        }
    ],
    "_e2_version": 4171,
    "_e2_ua_string": "Aegea 11.4 (v4171e)"
}