Dr. Alex Ehsan, MD, PhD

Hematology/Oncology physician | Free cancer education in English & Spanish | Personal views only | No employer affiliation | Not medical advice

[Dr. Alex Ehsan, MD, PhD @DrAlexEhsan]

Thomas, I think your point about herding bias goes even deeper than repeatedly rediscovering HER2, EGFR, and the other familiar targets.

Coming from cancer biology, hematology, BMT, and CAR T, I have learned that some of the most important biology often lives in the patient who breaks our model.

The HER2 positive tumor that does not respond.

The biomarker negative patient who has an extraordinary response.

The apparently identical cancers that develop completely different mechanisms of resistance.

Maybe the next generation AI scientist should not simply search for new targets.

Maybe it should be trained to search aggressively for contradictions.

One AI proposes a biological hypothesis.

A second deliberately searches pathology, genomics, transcriptomics, treatment response, toxicity, and longitudinal clinical data for patients who prove that hypothesis wrong.

A third asks what experiment would most efficiently distinguish between the competing explanations.

Only the biology that survives that adversarial process moves forward.

AI can inherit not only target bias, but also our historical biases in which patients were biopsied, which assays were ordered, which endpoints we considered important, which negative studies were never published, and even how diseases were classified in the first place.

So perhaps truly novel biology requires more than novel data.

It requires an AI scientist willing to challenge the questions humans decided were worth asking.

Your background in oncohematology makes this especially interesting to me because cancer teaches this lesson repeatedly. The dominant clone today may not be the clone driving relapse tomorrow.

Could K Pro eventually become not just a system that knows more biology, but a system specifically designed to recognize when the accepted biological explanation is wrong?

That is the AI scientist I would want beside me at tumor board.

[Dr. Alex Ehsan, MD, PhD @DrAlexEhsan]

This may eventually become much bigger than a privacy solution.

Coming from cancer biology, hematology, BMT and cellular therapy, I keep thinking about all of the knowledge trapped inside individual hospitals.

One institution may have 40 patients with an unusual molecular subtype.

Another may have 15 patients who developed an unexpected toxicity.

Another may have discovered a resistance pattern after immunotherapy.

Individually, none of those datasets may be large enough to change practice.

But what if the question could travel instead of the patient data?

Imagine thousands of hospitals functioning as nodes of one learning oncology system.

An algorithm could ask every institution:

Which molecular signature predicted response?

Which patients developed resistance?

Which immune phenotype preceded toxicity?

Which rare cancer responded unexpectedly?

The raw patient records never need to leave the hospital, but the biological lesson could still become part of a global model.

Then take it one step further.

Could federated learning eventually become part of the clinical trial itself?

Instead of waiting years to build centralized databases, every participating hospital could continuously contribute learning from genomics, pathology, imaging, treatment response and toxicity while preserving local control of the data.

That could be especially powerful for rare cancers where no single institution will ever see enough patients.

The great opportunity may therefore not simply be connecting hospitals.

It may be turning thousands of isolated hospitals into one continuously learning scientific organism.

And the question I would ask Owkin is this:

Can federated AI eventually move beyond predicting outcomes and begin generating hypotheses across hospitals that humans would never have thought to ask?

That is where collective intelligence in medicine becomes truly extraordinary.

[Dr. Alex Ehsan, MD, PhD @DrAlexEhsan]

Longevity becomes much more interesting to me when we stop asking how many years we can add to life and start asking how many years of functional biology we can preserve.

Coming from cancer biology, BMT and cellular therapy, I have watched what happens when individual cellular systems lose resilience.

DNA damage accumulates.

Clonal populations emerge.

Immune surveillance weakens.

Mitochondria change.

Stem cell compartments become less regenerative.

Inflammation becomes chronic.

So I wonder whether aging should eventually be approached less as one disease and more as a progressive failure of multiple interacting biological networks.

The out of the box question is this.

Instead of developing one longevity drug at a time, could we someday measure a person’s dominant aging mechanisms and build a personalized intervention around them?

One person may primarily have immune aging.

Another mitochondrial dysfunction.

Another accelerated clonal hematopoiesis.

Another impaired DNA repair or cellular senescence.

Then longevity medicine would begin to resemble precision oncology.

Not one treatment for aging.

A biological map showing which system is failing first and an intervention designed around that individual trajectory.

To me, extending lifespan is interesting.

Extending healthspan by understanding why each person is aging is much more interesting.

[Dr. Alex Ehsan, MD, PhD @DrAlexEhsan]

Como oncólogo con formación en biología del cáncer, trasplante de médula ósea y terapia celular, lo que más me impresiona de Cas12a2 no es simplemente que pueda destruir una célula tumoral.

Es que una célula puede ser obligada a revelar su identidad molecular antes de morir.

Eso cambia completamente mi manera de pensar en CRISPR.

Hasta ahora hablamos mucho de CRISPR como una tijera molecular capaz de corregir DNA.

Cas12a2 plantea otra posibilidad.

Un sistema capaz de leer RNA y tomar una decisión.

Pero el cáncer rara vez se define por una sola señal.

Un tumor puede perder un antígeno, cambiar su expresión genética, desarrollar resistencia y evolucionar bajo presión terapéutica.

Entonces me pregunto si la próxima generación debería ir más allá de un solo interruptor.

¿Qué ocurriría si pudiéramos construir una lógica molecular que requiriera varias señales tumorales simultáneamente antes de activar la destrucción celular?

KRAS mutado presente.

Firma transcripcional tumoral presente.

Señales de tejido sano ausentes.

Solo entonces se activa el mecanismo.

En otras palabras, no simplemente CRISPR que reconoce una mutación.

CRISPR que pregunta primero:

Estoy realmente dentro de una célula cancerosa?

Y solamente después toma una decisión irreversible.

Eso podría transformar la precisión oncológica.

Si el cáncer cambia durante el tratamiento, ¿podríamos algún día actualizar esas reglas moleculares de reconocimiento a medida que el tumor evoluciona?

Quizás el futuro de CRISPR en oncología no sea solamente editar genes.

Quizás sea enseñar a las células a distinguir biológicamente entre “yo” y “enemigo”.

Ese concepto me parece

For our English speaking participants what impresses me most about Cas12a2 is not simply that it can destroy a cancer cell.

It is that a cell can be forced to reveal its molecular identity before it dies.

That changes the way I think about CRISPR.

Until now, we have often discussed CRISPR as a molecular scissors capable of correcting DNA.

Cas12a2 suggests another possibility.

A system that can read RNA and make a decision.

But cancer is rarely defined by a single signal.

A tumor can lose an antigen, change its gene expression, develop resistance, and evolve under therapeutic pressure.

So I wonder whether the next generation should go beyond a single kill switch.

What if we could build molecular logic that required several tumor signals to be present at the same time before cell destruction is activated?

Mutant KRAS present.

Tumor transcriptional signature present.

Healthy tissue signals absent.

Only then does the system activate.

In other words, not simply CRISPR that recognizes a mutation.

CRISPR that first asks:

Am I truly inside a cancer cell?

And only then makes an irreversible decision.

That could transform precision oncology.

If the cancer changes during treatment, could we someday update those molecular recognition rules as the tumor evolves?

Perhaps the future of CRISPR in oncology will not be only about editing genes.

Perhaps it will be about teaching cells to biologically distinguish between self and enemy.

[Dr. Alex Ehsan, MD, PhD @DrAlexEhsan]

Dr. Hamamoto, beautiful work.

Coming from cancer biology, hematology and cellular therapy, what interests me most is not simply reconstructing a three dimensional heart from ultrasound.

I wonder whether the next step is to reconstruct the future of that heart.

A fetal cardiovascular system is not static anatomy. It is a rapidly changing biological system whose circulation will be fundamentally reorganized at birth when the placenta is removed, pulmonary resistance falls, and the ductus and foramen ovale begin to close.

What if this technology eventually became a longitudinal fetal cardiovascular digital twin?

Combine serial ultrasound sweeps with Doppler flow, vessel dimensions, ventricular function and gestational age. Instead of asking only whether today's anatomy looks abnormal, the model could learn each fetus's trajectory and simulate what that circulation may do after birth.

Then AI could potentially help answer a much more clinically important question.

Not simply:

Is congenital heart disease present?

But:

Which fetus is likely to remain stable after delivery, which may deteriorate when fetal shunts close, and which should be delivered immediately beside pediatric cardiology and cardiac surgery?

Could the model generate a counterfactual normal circulation for the same gestational age and quantify exactly where that individual fetus begins to diverge from the expected developmental pathway?

That changes AI from an image recognition tool into a physiological forecasting system.

In oncology I have learned that recognizing disease is valuable.

Predicting where the biology is going next is far more powerful.

Could fetal cardiac imaging ultimately make that same transition?

[Dr. Alex Ehsan, MD, PhD @DrAlexEhsan]

I have followed Patrick Soon Shiong’s work for years, and there are very few people in oncology whose career makes me stop and think about what one physician scientist can build when he refuses to stay inside one discipline.

He began as a surgeon and scientist, helped change drug delivery through Abraxane, then kept moving upstream into tumor biology, innate immunity, NK cells, T cells and cellular therapy.

That progression means something to me.

My own path took me through hematology and oncology, bone marrow transplantation and CAR T. The longer I have worked around cancer, the more convinced I have become that the future will not come from asking which single drug kills a tumor best.

Cancer is an ecosystem.

The tumor, NK cells, T cells, macrophages, cytokines, metabolism, antigen presentation and the surrounding microenvironment are all interacting at the same time.

That is why I find Dr. Soon Shiong’s broader idea so compelling. ANKTIVA should not simply be thought of as another bladder cancer drug. The deeper scientific question is whether activation of IL-15 biology can become one component of a larger immune architecture in which we restore immune competence, improve tumor recognition and then combine that biology intelligently with other therapies.

What if someday we stop classifying immunotherapy primarily by tumor organ?

Instead of lung cancer therapy, bladder cancer therapy or pancreatic cancer therapy, imagine classifying patients according to their immune failure state.

One patient may have exhausted T cells.

Another may have inadequate NK-cell activity.

Another may have poor antigen presentation.

Another may have profound lymphopenia.

Another may have a suppressive myeloid environment.

Then treatment could be assembled around repairing the specific immune defect rather than simply treating the anatomical name of the cancer.

That, to me, would be a true immunotherapy platform.

Patrick's work sits exactly at the intersection of cancer biology, immunology and cellular therapy that has fascinated me throughout my career. Sorry for deviating from the stock talk!

[Dr. Alex Ehsan, MD, PhD @DrAlexEhsan]

Coming from cancer biology, BMT and CAR T, I agree that meticulous daily cell work is the foundation. But I wonder if the next step is to stop treating cell culture as a series of snapshots.

What if every culture had its own digital twin?

Instead of recording temperature, media changes, morphology and cell counts once or twice a day, we could continuously integrate imaging, metabolism, oxygen consumption, cytokines, cell state and growth kinetics into one evolving model.

Then the system would not simply tell us that a culture has gone wrong.

It might predict 24 to 48 hours earlier that the cells are drifting toward exhaustion, differentiation, contamination or loss of potency and recommend an intervention before the failure becomes visible.

For cellular therapies, that could be transformative.

Today we manufacture cells and then ask whether the final product passes testing.

The future may be continuous quality control where every hour of a cell’s life contributes to predicting what that cell will eventually do in a patient.

Maybe the most important assay will not be the final test.

[Dr. Alex Ehsan, MD, PhD @DrAlexEhsan]

This makes me wonder whether we are eventually going to discover that we are treating Huntington disease too late.

AMT 130 is trying to change the slope of a genetically programmed disease after clinical disease has already appeared.

But Huntington biology begins years before meaningful functional decline.

In oncology we learned a difficult lesson. Controlling the driver becomes progressively harder to translate into recovery after irreversible tissue damage has accumulated.

So what happens if a therapy like this truly modifies disease and we eventually move intervention upstream?

Could genetically confirmed individuals someday be followed with neurofilament light, imaging and other biomarkers until biology tells us that neuronal injury has begun, rather than waiting for symptoms to tell us?

The difficult part is that AMT 130 is a one time neurosurgical gene therapy and it lowers normal huntingtin along with mutant huntingtin. Treating an asymptomatic person therefore creates an entirely different risk benefit equation.

But if the signal continues to hold, I think the next great Huntington question may become not only:

Does gene therapy work?

It may become:

How early are we willing to treat a genetic disease when we already know where the biology is heading?

[Dr. Alex Ehsan, MD, PhD @DrAlexEhsan]

What strikes me about these Huntington trials is that we may be asking the wrong question if we view them as three competing drugs.

I have learned that complex diseases are rarely conquered by attacking only one layer of biology.

Votoplam is trying to lower huntingtin protein. SKY 0515 is particularly interesting because it lowers mutant huntingtin while also reducing PMS1, a modifier involved in somatic CAG expansion. Pridopidine approaches the disease from another direction through sigma 1 receptor biology and neuronal resilience.

That makes me wonder whether the future of Huntington disease eventually looks less like choosing a winner and more like combination oncology.

Could we simultaneously reduce the toxic protein, slow continued CAG expansion inside vulnerable neurons, and protect mitochondrial, synaptic and cellular function?

In cancer, we learned that suppressing one pathway often allows another mechanism of disease to continue.

Maybe Huntington disease will ultimately require the same conceptual shift.

Do we eventually move from one drug against HD to a rational combination that attacks the mutation, its downstream protein toxicity, and neuronal vulnerability at the same time?

That is the trial I would love to see someday.

[Dr. Alex Ehsan, MD, PhD @DrAlexEhsan]

What caught my attention here is not really the CO₂.

It is the creation of a second energy currency inside a living cell.

Coming from cellular therapy background, this immediately makes me wonder whether orthogonal cofactors could eventually become a form of metabolic firewall.

Today, when we engineer a cell, the new pathway still competes with the cell’s native biology for energy, reducing equivalents and substrates. That competition can create toxicity, instability and unpredictable behavior.

But imagine giving an engineered therapeutic cell its own private metabolic economy.

A synthetic pathway could run on a cofactor that normal human metabolism barely recognizes. CAR T cells, engineered immune cells or therapeutic microbes could potentially use that separate chemistry to power only the functions we designed.

Could we make an engineered cell dependent on that artificial cofactor so it survives and performs its therapeutic function only when we supply the corresponding metabolic input?

That would turn orthogonal metabolism into both an energy system and a safety switch.

We talk constantly about engineering genes.

Maybe the next frontier is engineering an entirely separate metabolism inside the same.
Imagine in our lifetime we may entangle metabolic engineering programmable cellular independence and biological containment.

[Dr. Alex Ehsan, MD, PhD @DrAlexEhsan]

Coming from BMT and CAR T, the number 10^16 is impressive, but it makes me think one step beyond protein design.

This is not CRISPR itself. It is an extraordinary way of designing and physically creating enormous DNA libraries.

But what happens when technologies like this converge with CRISPR, base editing and prime editing?

Imagine AI generating millions or billions of candidate regulatory sequences, receptors or genetic modifications, then gene editing placing selected designs directly into living cells. Those cells could be challenged with cancer, infection or another disease state, and the biological winners and failures could feed straight back into the next AI generation.

At that point we are no longer just designing molecules.

We are creating a closed evolutionary loop between AI, DNA synthesis, gene editing and living cells.

Could continuous variational synthesis eventually become the design engine upstream of CRISPR, where AI proposes the genetic possibilities, editing tests them in living biology, and the results teach the next generation what to build?

If that loop becomes fast enough, the real breakthrough may not be 10^16 sequences.

It may be learning how to navigate biological possibility itself.

GREAT JOB JURA BIO.

[Dr. Alex Ehsan, MD, PhD @DrAlexEhsan]

After years in oncology, I wonder if we are thinking about nanorobots too narrowly.

Why make them simply deliver chemotherapy?

Imagine a circulating machine that first interrogates the biology around it. Is this malignant tissue? Is it hypoxic? What antigens are present? Is there resistance? Only after several signals agree does it release therapy locally. Then it measures the response, communicates what happened, and safely disappears.

In oncology we currently biopsy, analyze, treat, wait, scan and adjust.

What if someday all five steps happened continuously inside the patient?

The real revolution would not be a smaller drug carrier.

It would be turning treatment into a closed loop biological system that can sense, decide, treat and learn in real time.

[Dr. Alex Ehsan, MD, PhD @DrAlexEhsan]

What this makes me think about is not just designing better proteins, but designing proteins that can make decisions.

After years in cancer biology, bone marrow transplant, CAR T and oncology, the hardest problem is rarely killing the cancer cell. It is killing the right cell, at the right time, in the right place, without injuring everything around it.

What if the next generation of AI designed proteins behaved more like biological logic gates?

A protein could remain inactive in normal tissue, recognize two or three features unique to a tumor microenvironment, activate only when those conditions are met, perform its function, then shut itself off.

That would move us beyond the traditional idea of one drug against one target.

We could begin designing therapeutics that sense context.

To me, that is where AI protein design becomes truly disruptive. Not just making new molecules, but creating molecules with conditional behavior.

The question I would ask is this: how close are we to designing proteins that do not simply bind a target, but actually make biologic decisions inside a patient?

[Dr. Alex Ehsan, MD, PhD @DrAlexEhsan]

Very cool, I always enjoy
MIT synthetic biology work using engineered bacteria. Coming from cancer biology and cellular therapy, this immediately caught my attention. We already think about cells as systems that sense signals, process information, and change behavior. This takes that concept another step by turning bacteria into components of a biological circuit.

The exciting part is not replacing silicon. It is giving living systems the ability to detect a problem and generate a programmed response.

But biology is much messier than a circuit board. What happens after hundreds of generations, changing temperatures, competing microbes, mutations, and environmental stress?

If researchers can make these circuits stable and predictable outside the laboratory, biological computing could become enormously important not only for agriculture, but eventually for medicine as well.

[Dr. Alex Ehsan, MD, PhD @DrAlexEhsan]

This is an ambitious intersection of genetics, reproductive medicine, and ethics.

Coming from cancer biology, hematology, and cellular therapy, I have watched precision medicine move from theory into treatments that once seemed impossible. Preventing severe inherited disease before birth could become another major step, but germline editing raises a different level of responsibility because the consequences may extend beyond one patient to future generations.

The science is exciting, but the hardest question may not be whether we can edit these genes. It may be deciding which conditions justify intervention, how we prove long term safety, and where society draws the line between preventing disease and selecting traits.

Cathy, how do you think this field should define that boundary before the technology moves faster than the ethical framework?

[Dr. Alex Ehsan, MD, PhD @DrAlexEhsan]

The science is fascinating because de-extinction is really a test of how far synthetic biology, gene editing, developmental biology, and AI can work together.

Coming from cancer biology and cellular therapy, I keep thinking about the same problem we face whenever we engineer living systems. Editing the genome is only one layer. Epigenetics, embryonic development, mitochondrial biology, gene regulation, and the environment all influence the final organism.

So the question for Colossal is bigger than whether we can create an elephant with mammoth-like traits.

How close can we actually get to reconstructing the biology of an extinct species rather than producing a modern proxy that simply looks and functions like one?

That distinction may end up teaching us as much about human genetics and regenerative medicine as it does about mammoths.

[Dr. Alex Ehsan, MD, PhD @DrAlexEhsan]

Coming from cancer biology, bone marrow transplantation, and CAR T, I have learned that an elegant mechanism is only the beginning. Translation into humans is where biology becomes unforgiving.

REP 0003 fascinates me because the strategy is fundamentally different from simply lowering circulating LDL. It attempts to give cells the machinery to degrade toxic excess intracellular free cholesterol, potentially attacking established plaque biology itself. The mouse data showing plaque regression are impressive, but we are still looking at preclinical evidence.

The question I would ask is whether human plaque macrophages will reproduce this degree of cholesterol clearance without creating new problems in intracellular sterol metabolism, inflammation, or plaque stability.

If that translates safely into humans, this could become much more than another cholesterol lowering therapy. It could change how we think about treating established atherosclerotic disease.

[Dr. Alex Ehsan, MD, PhD @DrAlexEhsan]

My background in cancer biology, bone marrow transplantation, and cellular therapy makes this especially interesting to me. We learned in oncology that getting cells to survive is only the first step. The harder questions are whether they remain functional, integrate safely, and continue working years later.

That is why the PET evidence of dopamine production may be more important than the headline 35 percent improvement. This was still a very small early study without a control group, so it deserves excitement with restraint.

The question I would love to see answered next is this. Five or ten years from now, will these transplanted neurons still produce dopamine and remain healthy, or will the biology driving Parkinson’s eventually affect the graft as well?

That is where regenerative medicine becomes truly fascinating.

[Dr. Alex Ehsan, MD, PhD @DrAlexEhsan]

@DeepLearn007 Hinton?

[Dr. Alex Ehsan, MD, PhD @DrAlexEhsan]

This is exactly the kind of conversation AI and drug development need more of. Organoids are especially interesting because they may help bridge the gap between elegant computational predictions and messy human biology.

My cancer biology and BMT background makes me wonder about the next step. If AI begins selecting compounds using organoid response data, how do we prevent the model from overlearning one snapshot of a tumor while missing clonal evolution, immune context, and treatment pressure?

That seems like a frontier question worth bringing to the dinner table.