231: The Future of Bone Marrow Biopsy: Omics and AI Integration

Paper Discussed in this Episode: Advancements in bone marrow biopsy: the role of omics and artificial intelligence in hematologic diagnostics. Maryam Alwahaibi and Nasar Alwahaibi. Front. Med. 2026; 13:1772478.

Episode Summary: In this journal club deep dive, we explore a paradigm shift in hematopathology, moving from 19th-century visual assessments to the cutting edge of precision medicine. We examine a 2026 review that unpacks how combining artificial intelligence with multi-omics technologies is transforming the traditional bone marrow biopsy from a static, subjective snapshot into a live, interactive, predictive 3D map. We ask: What happens when deep learning can predict underlying genetic mutations just by analyzing the visual shape and texture of a cell?.

In This Episode, We Cover:

The Breaking Point of Traditional Diagnostics: Why the 150-year-old gold standard of H&E staining and human visual assessment is hitting a biological and operational wall, plagued by subjectivity, high variability, and observer fatigue.

The Multi-Omics Multiverse: Moving beyond standard genomics to unpack the complex biological machinery of the marrow, including:
Epigenomics: The biological "switches," like DNA methylation, that control cell fate and can kick off malignant transformation without altering the underlying DNA sequence.
Lipidomics: How cellular fats form specialized signaling rafts that actively remodel the marrow's communication network.
Microbiomics (The Gut-Marrow Axis): How systemic inflammation driven by gut dysbiosis acts like a massive "traffic jam" that indirectly disrupts local bone marrow homeostasis and blood cell production.

AI as the Ultimate Analytical Partner: How artificial intelligence serves as a bridge between physical tissue morphology and high-dimensional molecular data. We discuss AI tools like MarrowQuant for objective cellularity mapping and the Continuous Index of Fibrosis (CIF) that replaces clunky human guesswork with a granular, predictive metric.

Predicting Genotype from Phenotype: The revolutionary capability of deep learning models to predict underlying genetic mutations (like TET2 or del 5q MDS) purely from the subvisual, spatial arrangement and shape of cells on a standard slide.

Roadblocks and Solutions: Why this technology isn't universally adopted yet. We break down the "black box" problem of AI, the brittleness of algorithms in different clinical settings, and how innovations like Federated Learning and Explainable AI (using heat maps) are overcoming these hurdles.

Key Takeaway: The integration of AI and multi-omics is redefining our understanding of bone marrow diseases. By uncovering invisible molecular machinery and objectively translating it through transparent algorithms, we are moving away from subjective human bottlenecks toward a highly personalized, predictive model of hematologic care.

Get the "Digital Pathology 101" FREE E-book and join us!

Transcript

Imagine uh imagine you're trying to manage traffic in this massive just sprawling metropolis.

Yeah. Yeah.

But the only tool you actually have to do your job is a static black and white photograph of the city skyline and it was taken from a helicopter at like noon.

So you can see the buildings basically.

Right. Exactly. You see the big skyscrapers, you get the general shape of downtown, but you have zero visibility into the actual traffic patterns. You can't see the power grid and you certainly don't know, you know, what emails are being sent inside those office buildings to actually coordinate the city's operations.

Yeah. Which is kind of essential if you're running a city.

Exactly. And for decades, traditional bone marrow biopsies have functioned a lot like that static photograph. You get this broad structural snapshot of a patient's cellular landscape, but you're well, you're largely blind to the underlying functional networks that are actually driving the disease.

Yeah, that's a perfect way to put it actually.

Welcome to the digital pathology podcast. for today's deep dive for all the trailblazers joining us today. So you know the pathologists, the oncologists, the hematology specialists out there, we are doing a journal club exploration of a paper that essentially takes that static photo and upgrades it into a live interactive predictive 3D map.

Right. And the paper we're looking at today is titled advancements in bone marrow biopsy the role of in artificial intelligence in humantologic diagnostics. It was published by Miam Alwahib and Nosar Alwahhibi. And this was in uh Frontiers in Medicine back in March of 2026.

I have it right here.

Yeah. And what this work really synthesizes for us and why it matters so much to the trailblazers listening is this stark reality that our traditional methods for assessing bone marrow are just hitting a wall. I mean these biopsies are the absolute cornerstone for diagnosing leukemias, lymphas, marrow failure conditions, but they're hitting a very hard biological and honestly operational ceiling.

Yeah. And I feel like before we get into the shiny new tech, we have to start by looking at how That old system is breaking down, right? Because that really sets the stakes for everything else.

Absolutely.

I mean, the gold standard has always been these conventional modalities, right? Histophology, flowcytometry, cytogenetics. A pathologist takes a bone marrow sample, uses an H& stain, which by the way, hemattoxylin and eosin, that dye technology is over 150 years old.

Yeah. It's literally 19th century technology,

right? It turns nucleoplasm pink, and then you just look down a microscope

and that visual assessment is just incredibly fraught because you know you're dealing with highly overlapping morphological features. The visual differences between a reactive state, so say a patient's marrow is just responding to a severe infection and early malignant changes like a developing mildis plastic syndrome.

They look almost identical sometimes.

Exactly. The differences can be incredibly subtle and because the entire interpretation relies on a human brain processing those subtle visual cues. The process is just inherently subjective. I mean, the medical literature is absolutely packed with data showing really high rates of intraobserver and intraobserver variability.

Wait, meaning two different doctors or even the same doctor on different days?

Both. You can hand the exact same slide to three highly trained, board-certified hematopathologists and literally get three slightly different interpretations of the blast count or, you know, the degree of dysplasia. Or like a pathologist might look at a slide on a Tuesday morning after their first cup of coffee, right? And then they look at that exact same slide on a Friday afternoon when they're experiencing severe cognitive fatigue and they just grade it completely differently.

Right. And plus scanning these slides manually is incredibly slow. So when you combine that slow subjective process with the current global shortage of expert bone pathologists, Yeah.

you just get massive clinical bottlenecks,

which is terrible for patient care.

It is. And it's especially disastrous for rare or you know differentiated tumors that just don't play by the standard visual rules anyway.

Okay, but let me push back on something here just playing devil's advocate for a second.

Sure.

We have relied on microscopes and train human eyes for decades, right? And we've cured or managed countless patients.

Why is this methodology suddenly insufficient right now? Like is the underlying biology of these diseases actually getting more complex or are standards for precision medicine just demanding way more than the human eye can deliver.

That's a great question. And what's fascinating here is that the underlying biology of a leukemia cell hasn't actually changed over the last 50 years.

Okay.

But our understanding of its molecular heterogeneity has absolutely exploded.

We now know that two leukemas that look morphologically identical under a microscope like down to the exact same shape, size, and staining pattern, they might actually be driven by completely different genetic mutations.

Oh wow. So they look the same but act totally differently.

Exactly. And more importantly, they might require completely different targeted therapies to achieve remission. So a human expert simply cannot manually quantify those highdimensional molecular complexities. I mean the human eye cannot see a genetic mutation, right? The human eye can't see a misfolded protein cascade,

right? If the human eye maxes out at the cellular level, then we have to drop down to the molecular level to see what's actually happening behind the scenes. And I think this is where the authors bring in the whole multiomics multiverse as they call it.

Yeah, the multiomics approach is key here

because to get past the bottleneck of that subjective visual analysis, the field had to develop technologies that interrogate the marrow at a microscopic really functional level.

Yeah. And these OMIX technologies, they systematically break down the marrow ecosystem layer by layer. The foundation of course is genomics and transcrytoics. We're moving far beyond the era of just sequencing a basic DNA blueprint. We are actively hunting for specific clonal mutations like single nucleotide polymorphisms or SNPs that tell us exactly where the DNA code broke

and then transcrytoics comes in.

Right? With transcrytoics, we're mapping out RNA splice variants. We can look at the active inflammatory signaling pathways to see not just what genetic code is broken, but what the cell is actually doing with that broken code in real time.

And the paper delves into epigenomics, too, which is where I think the mechanics gets super fascinating. Yeah. Because we aren't just looking at the code itself anymore. We are looking at DNA methylation and chromatin remodeling.

Yeah. The switches.

Exactly. They function as these physical roadblocks or access points on the DNA. There is the switches controlling lineage specification. You know, determining whether a hematopoetic stem cell turns into a red blood cell, a platelet or a white blood cell. When a methyl group attaches to the wrong spot and flips the wrong switch, it can completely silence a tumor suppressor gene, right? Which just kicks off malignant transformation without actually ultra the underlying DNA sequence at all.

That's spot on. And beyond the genome and the epiggenome, we get into what the authors highlight as the emerging frontiers. Lipidomics is a prime example of this,

which is crazy to me because lipids are just fat, right?

Yeah. Well, historically, yeah, lipids were viewed primarily as just passive structural components of cell membranes or like inert energy storage.

But lipidomics reveals that specific lipid species, things like ceramides or sphinol lipids, they actively regulate stromal hematop Poyetic communication.

Wait, really? They're sending signals.

Yes. They form these specialized rafts on the cell membrane that aggregate signaling receptors. By changing the lipid composition, they physically remodel how the marrow micro environment allows memal progenitors, osteoblasts, and immune cells to communicate with each other. They're essentially rewriting the entire communication network of the marrow.

That is wild. And speaking of wild communication networks, the section on microbiomics and the gut marrow axis completely shifts how we have to think about localized disease.

Oh, absolutely.

The paper details how systemic immune function and gut dispiosis, so severe imbalances in your gut bacteria, how they indirectly impact me homeostasis and even how a patient responds to chemotherapy. It operates a lot like a complex supply chain. You know, going back to our city analogy, you can have a local factory, the bone marrow, that is fully staffed and ready to produce blood cells, but if there is a massive traffic jam. Three towns over representing inflammation in the gut microbiome. That traffic jam delays the raw materials and sends stressed, agitated drivers into your local town.

I love that analogy. It's exactly like that. The systemic inflammation driven by the gut microbiome alters the cytoine balance that's circulating in the blood, which eventually of course hits the marrow. And the marrow is highly sensitive to inflammatory cytoines like interlucan 6 or TNF alpha.

Right?

Chronic exposure to those signals can push hematopoatic stem cells to exhaust themselves or bias them toward producing myoid cells over lymphoid cells. And for the trailblazer listening and trying to apply this to their practice, these mix layers aren't just esoteric biology, right? They fundamentally refine our diagnostic classifications. They detect targetable therapeutic vulnerabilities and minimal residual disease at levels that a traditional H& stain would completely miss.

But we run into a serious computational wall here, though, because provides an absolute mountain of molecular data, but a pathologist still needs a way to interpret the physical tissue morphology. Like we need a bridge between the digital data of a sequence genome and the physical reality of the tissue on the slide without getting trapped in those subjective human bottlenecks we established earlier.

Enter artificial intelligence. AI acts as that bridge serving as basically the ultimate analytical lab partner.

Okay, let's break that down. How does it apply to hisystologology meaning the actual solid tissue biopsies?

Well, in conditions like myop proliferative neoplasms, MPNs, One of the most critical diagnostic and prognostic tasks is grading the degree of bone marrow fibrosis. Traditionally, this is a highly subjective scoring system from 0 to three based on how thick the reticulin fibers look under the microscope,

which goes back to the human error problem.

Exactly. But AI models have now been trained to evaluate those reticuline changes pixel by pixel, generating a continuous index of fibrosis or CIF. So instead of a clunky 0, 1, 2, or three, you get a highly a granular continuous metric and this objective grading has been shown to predict disease progression in MPN's far more accurately than human estimation

and the paper highlights specific automated tools too right like Maroquin because when a human estimates cellularity they usually just scan the slide and guess saying something like wow looks like maybe 40% cellularity

yeah it's a very educated guess but still a guess

right but Maroquant automates the mapping of adiposites bone tbecula and vascular structures it systematically analyzes the entire slide mathematically subtracting the area taken up by fat and bone to standardize the cellularity assessment. It just removes the human guesswork completely.

It does. And from solid hisystologology, the authors moved to cytology. So the liquid bone marrow aspirate smears. Figure two in the paper outlines this comprehensive endtoend clinical workflow. And what really stands out there is how the AI doesn't just analyze cells, it actively manages quality control.

Yes. The moment a slide is digitized, the first thing the AI does is run a QC protocol. Right. Right.

It aggressively rejects crush artifacts, excessively thick areas of the smear and severe staining outliers, which is huge because if a human technician prepares a poor smear, a human pathologist might waste 20 minutes squinting at deformed cells just trying to force a diagnosis, the AI instantly flags the preparation as non-diagnostic.

Exactly. And if the slide does clear QC, the system moves to cell detection, segmentation, and classification, generating an automated differential uh histogram of cell types.

Oh. It flags suspicious morphological patterns and routes those specific regions of interest directly to the human expert for review. It functions as a highly sophisticated triage system and the AI possesses diagnostic capabilities that really push the boundaries of current pathology. The paper discusses models like displionet in myelvisisplastic syndromes. Identifying hypoganulated displastic neutrfils is clinically vital but notoriously difficult.

Why is this so hard?

Because a normal neutrfil has a cytoplasm packed with specific granules. In MDS, those granules can be sparse or absent, but the visual difference is incredibly subtle, especially if the slide stain is even slightly washed out. Disclaion detects this specific morphologic deficit objectively and consistently standardizing a task that routinely trips up human observers.

And it plays a massive role in acute primalic leukemia or APL too because APL is a true hematlogic medical emergency. Patients can rapidly develop disseminated intravascular coagul population. I mean, they can essentially bleed to death before a molecular test even comes back from the lab confirming the diagnosis.

It's terrifyingly fast.

Yeah. And APL is driven by a specific genetic swap, right? The T1517 transllocation. The AI can rapidly recognize the morphologic patterns associated with this transllocation like hyper granular cells with folded nuclei or bundles of hour rods in minutes. It reliably distinguishes APL from other acute myoid leukemias just by analyzing the geometry. and texture of the cells which allows clinicians to start life-saving all transretinoic acid therapy immediately.

Right? And if we connect this to the bigger picture, the algorithm is predicting genotype from phenotype. This is the paradigm shift. Deep learning models are linking the visual shape, texture, and spatial arrangement of cells directly to underlying genetic mutations.

That feels like magic.

It really does. A neural network processes pixel intensity gradients and highdimensional spatial relationships that our human visual cortex literally did not evolve to recognize. The paper notes AI can identify subvisidual hisystologic features strongly associated with TET2 mutations or del 5QMDS. You can look at the shape of a cell and say with high statistical probability, this cell is missing the long arm of chromosome 5.

Okay, let me stop you there because predicting genetic mutations from a simple slide scan sounds almost too good to be true. If AI can do this rapidly and cheaply, why isn't every community, hospital, and clinic on Earth running this software today? Like, does integrating AI actually save time or does it just create a new operational nightmare where pathologists spend all day babysitting algorithms, verifying these blackbox outputs and fighting with IT systems?

The friction points you're bringing up are exactly why this isn't universally adopted yet.

The blackbox nature of deep learning is a massive regulatory and clinical hurdle. A convolutional neural network might correctly identify a TET2 mutation, but it outputs a probability score without explaining its biological reasoning.

Right. It just says, "Trust me."

Yeah. And in medicine, you cannot simply tell a patient they have a specific leukemia because a proprietary algorithm said so. Pathologists require an auditable trail of evidence to sign off on a diagnosis, and blackbox models just don't provide that.

There's also a severe lack of generalizability, right? AI models are incredibly brittle when taken out of their training environments. You could have an algorithm trained on hundreds of thousands of pristine slides from a well-funded, high-end academic center, but it might completely collapse when you deploy it in a rural community hospital. hospital. Oh, totally. Because the scanner resolution is different or the glass slides are a millimeter thicker or the eosin stain has a slightly different pH causing a minor color shift.

Right. And the omic side has equally daunting hurdles, doesn't it?

Definitely. High throughput sequencing assays are expensive, technically complex, and demand massive biioinformatics infrastructure. Furthermore, a bone marrow biopsy is a physical procedure. If you get a dry tap or a hypocellular sample, extracting enough highquality RNA A or lipid material for multiomics analysis is incredibly difficult. A lowquality sample inputed into a high-end OMIS pipeline will just output biased incomplete molecular profiles.

And then you have the absolute headache of data integration because genomic data is static. Your DNA mutations are relatively fixed. But transcrytoic or metabolic data is highly dynamic. It fluctuates based on the time of day, the patient's stress levels or recent medications.

Yes. Exactly.

So mashing static structure al data together with temporal functional data and then trying to map that onto an AI generated spatial map of the tissue. I mean without established universal interpretation frameworks, it's a recipe for clinical paralysis. Are we looking at a bifurcated future where only elite well-funded academic centers can offer this level of precision diagnostics leaving community clinics relying on traditional microscopes?

The field recognizes that risk absolutely and is actively engineering solutions to democra advertise the technology to solve the AI generalizability problem. Researchers are shifting toward federated learning.

Oh, federated learning. That's where you train the model without moving the data. Right.

Exactly. Instead of a community hospital sending its sensitive raw patient data to a central server to train an algorithm which triggers massive data privacy and bandwidth issues. Federated learning sends the algorithm to the hospital.

Oh, okay.

The model trains locally on the community hospital's uniquely stained slides, updates its internal mathematical weights and then sends only those updated weights back to the central server. So the global model learns from the diverse messy reality of thousands of different labs without ever centralizing the raw data.

That's brilliant. And to solve the blackbox problem, the industry is forcing a shift toward explainable AI. Right? These are systems architected to basically show their work. Instead of just spinning out a diagnosis of MDS with 90% confidence, explainable AI generates a heat map overlay on the slide.

Yeah, the heat maps are game changers.

It highlights the exact morphologic features, the specific hypoganulated neutrfils or the abnormally clustered meggaariittes that actually drove the algorithm's prediction. It turns the AI from a mysterious oracle into a transparent assistant that the pathologists can verify and trust.

Right? And all of this development points toward the ultimate future vision outlined by the Alwahibes in the paper endtoend computational pathology pipelines. The authors point to spatial multiamic atlases specifically mentioning multipplexed imaging techniques like codeex

code detection by indexing.

Yes, codeex. It uses DNA barcoded antibodies to visualize dozens of different proteins on a single tissue section simultaneously without destroying the sample. It allows researchers to physically map genetic and proteomic alterations to their exact topographical niches in the marrow micro environment.

So you can actually see that the cells with the TET2 mutation are physically clustering around specific blood vessels while the health healthy cells are pushed to the periphery. You get the 3D map with the traffic patterns and the electrical grid fully illuminated. Going back to our very first analogy.

Exactly. It all comes full circle. To distill everything we've covered today for you, traditional bone marrow biopsies relying on visual inspection of H& stains are foundational but inherently limited by human subjectivity and visual constraints. Multi-omics technologies from the genomic code down to the lipidomic signaling rafts and the gut microbiome. of systemic influence reveal the invisible molecular machinery driving hematologic diseases.

An artificial intelligence is the engine that actually processes this overwhelming amount of data. It objectively quantifies tissue fibrosis, automates quality control on aspirate smears, and performs the highly complex mathematical translation required to link a cell's physical phenotype directly back to its underlying genomic mutations.

It is a complete paradigm shift in how we understand tissue.

It really is.

Which leaves us with a compelling philosophical question for you to Consider as we wrap up if artificial intelligence can eventually and accurately predict a patient's underlying molecular mutations purely by analyzing the visual shape, texture, and spatial arrangement of cells on a standard inexpensive H& slide? Does the traditional boundary between a morphologist and a molecular geneticist even exist in the future? As you navigate this evolving landscape as a trailblazer, will your value be defined by the physical tests you order or by your expertise in curating the digital algorithm? that interpret the tissue.

Man, that is a brilliant thread to pull on as we close out. From a static photograph of a city skyline to a living, interactive, predictive map of the human micro environment. Thank you to all you trailblazers for exploring this paper with us today on the digital pathology podcast. We will see you next time.