Cancer is a terrible disease, and also one that we all know too well.

It is not a new problem, rather one that exists since thousands of years & is studied in unimaginable detail.

Then why do people still die of cancer?

Let's start understanding this by taking a step back.

It’s 1938, and Public Health Services are advising people that detecting and treating cancers early will save their lives.


Now fast-forward nowadays. We hear the exact same core message from the Public Health Services of our times, gradually and consistently backed up by more and more data.

How and why is public health advice from 1938 so relevant today, 85 years later? Does this mean we made no progress in curing cancer?

Certainly not.

Let me show you why👇

By zooming into cancer medical care in 1938, we quickly realize how rudimentary it actually was.

Think of radical mastectomy: a truly invasive procedure developed by William Stewart Halsted at the end of the 19th century, in which the whole breast, all lymph nodes under the arm, and the chest wall muscle were removed.

Radical mastectomies were still the norm following breast cancer diagnosis until 1970s.

Nowadays, such surgery is almost never done, and has been replaced by local removal of the tumor (lumpectomy), followed or preceded by multiple options of additional therapies.

Now, let’s better understand the real impressive transformations that cancer science & medical practice have undergone in the last decades. I’ve chosen 10 such milestones (1938-2023) to emphasize below.

This list is by no means exhaustive, rather simply a conceptual backbone.

1.

1941 - Hormonal Therapy for Prostate Cancer

Charles Huggins discovers that lowering testosterone production by removing the testicles or by giving estrogen leads to shrinkage of prostate tumors.

Hormonal therapy continues to be a main treatment for prostate cancer nowadays.

2.

1958 - Combination Chemotherapy

James F. Holland, Emil Freireich, and Emil Frei propose that using multiple chemotherapy agents simultaneously will make it more difficult for tumors to become resistant.

Their hypothesis builds upon generalizing the current state of art in tuberculosis treatment, in which the combination of antibiotics was more efficient than single agents.

Nowadays, combination chemotherapy is a widespread standard of care in cancer centers around the world.

3.

1978 - Tamoxifen

An anti-estrogen drug, originally developed as birth control treatment, is approved by the FDA for the treatment of estrogen receptor positive (ER+) advanced breast cancer.

In the following years, Tamoxifen was shown to also be effective for early disease.

Tamoxifen is a targeted treatment: it specifically targets only a particular type of cancerous cells (unlike chemotherapy).

It is delivered as a tablet, and is currently the gold standard treatment for ER+ breast cancers (80% of all newly-diagnosed breast cancers).

4.

1998 - Trastuzumab

FDA approves Herceptin (Trastuzumab), a monoclonal antibody that inhibits the tumor-stimulating effects of the over-expressed amplicon HER2, for the treatment of metastatic HER2+ breast cancers (and later on also for treating early-stage HER2+ disease).

Therapies involving Trastuzumab (usually in addition to chemotherapy) currently remain the gold standard in treating HER2+ breast cancers.

5.

2001 - Imatinib

FDA approves Imatinib (Gleevec) for the treatment of chronic myeloid leukemia (CML), transforming a cancer death sentence into a manageable condition.

Back in 1960s, Peter Nowell and David Hungerford (then a researcher at the University of Pennsylvania) found that CML patients had an abnormally short chromosome (later named the Philadelphia chromosome).

Later on, Nora Heisterkamp at NCI identifies the reason why the Philadelphia chromosome was shorter: when this chromosome forms, two genes get fused together, forming the BCR-ABL fusion protein.

Further, when this fusion happens in blood cells, it causes CML.

The drug Gleevec, first conceptualized by Brian Drucker at Oregon University, blocks this fusion, hence only targeting faulty cells that display it.

5 years after the first phase 1 clinical trial of Imatinib (1998), 98% of treated CML patients were disease-free.

Nowadays, Imatinib is a potential curative treatment: if a CML patient is in remission after 2 years of Imatinib treatment, their life expectancy is the same as a person who doesn’t have cancer.

In May 2001, this monumental discovery makes it on the cover of TIME magazine.

6.

2006 - Gardasil

On a completely different line of research, the first vaccine for the prevention of a human cancer is approved by the FDA in 2006.

Gardasil was originally developed by researchers at the University of Queensland in Australia.

It initially protected against infections from four Human Papillomavirus (HPV) strains: HPV16 and 18, causing 70% of cervical cancers, and HPV 6 and 11, causing 90% of genital warts.

The current form of this vaccine is Guardasil 9, which protects against 5 additional HRV types.

Harald zur Hansen, who contributed to proving that genital HPV infections can causally lead to cervical cancer, is awarded the Nobel Prize in 2008 for this discovery

The approval of Guardasil is an important step also as it is a prevention measure, rather than a cancer treatment

7.

2011 - Ipilimumab

In all fairness, the last decade belongs to immunotherapy, which aims to boost the natural body’s defense system to eliminate cancerous cells

The idea of unleashing the immune system against cancer goes many centuries back, but most progress is very recent

Ipilimumab is the first FDA-approved immune checkpoint inhibitor, which are monoclonal antibodies that release the breaks of the immune system to trigger immune-mediated anti-tumor responses.

Immune checkpoints themselves are molecules that work to maintain immune tolerance.

But these molecules are often used by tumor cells to their advantage, to evade immunosurveillance.

Ipilimumab activates the immune system by targeting CTLA-4, a co-inhibitory molecule expressed on T cells which negatively regulates T cell activation.

8.

2014 - Pembrolizumab (Keytruda)

The first PD-1 immune checkpoint inhibitor approved by the FDA.

PD-1 is expressed on the surface of several immune cell types (including T cells) and acts as a negative regulator of immune response.

Its ligand PDL-1 is expressed in normal tissues and regulates immune tolerance by suppressing T cell-mediated immune proliferation and cytokine secretion when binding to PD-1.

The interaction between PD-1 and PDL-1 diminishes immune activity & helps prevent autoimmune disease.

But, tumor cells hijack this mechanism and abnormally express PDL-1 to escape immune surveillance.

This is why antibodies targeting PD-1 or PDL-1 have been approved for the treatment of multiple cancer types.

Blockage of the PD-1 pathway has been very successful in the clinic.

In 2017, a conceptual shift in cancer immunotherapy takes place: Keytruda is approved for general use in any tumor with a certain genetic profile (based on tumor mutational burden), rather than specific sites.

9.

2017 - Kymirah (Tisagenlecleucel)

First chimeric antigen receptor (CAR) T cell therapy approved by FDA: T cells genetically modified & expanded ex vivo (outside the patient) to mount personalized tumor attacks, by targeting neoantigens (mutations displayed on cancer cells).

These engineered fighters are then infused back into the patient, with the idea of selectively attacking the cells they were designed to target.

A pioneering early-stage study coupled non-viral CAR-T therapy with CRISPR editing, showing promise for increased clinical efficacy.

https://twitter.com/simocristea/status/1594736869211869187?s=20

As of today, multiple cancer types show sustained clinical responses to immunotherapy.

But real drawbacks still exist

1) immunotherapy is expensive
2) response rates remain limited
3) there is no consistent satisfactory understanding of factors predicting these responses.

10.

2020 - AI in cancer

This decade belongs to artificial intelligence, which is quickly transforming the way we understand and interact with our own biology.

Large AI models can now efficiently learn the image features separating cancerous lesions from normal ones from large amounts of labeled data.

Such models are already deployed in medical centers across the world to assist with clinical cancer diagnosis.

The pace of change is extremely rapid, and powerful AI models will soon be able to

1) assist in efficient drug design
2) accurately predict patient-specific personalized responses to various therapies
3) flag high-risk conditions for clinical management, as a preventive measure

Ok, this all sounds truly impressive.

But how much does it matter in the end?

How do all these great therapeutic discoveries factor in for the one metric that we humans care about most: the time we have to live?

The harsh truth is: not great.

Despite impressive progress in cancer research, the drop in death rate from cancer is not as impressive.

This becomes even more apparent when compared with what used to be the main source of death in the developed world (cardiovascular disease), for which the drop has been twice as large.

Trend holds also for 🇨🇭, a best-case scenario for healthy lifestyle & optimal healthcare.

There are lots of complex reasons to explain these differences in magnitudes, some more intuitive than others.

Nevertheless, it does leave us thinking.

And it begs the question: why is progress against cancer so difficult?

The answer is that cancer is a relentless evolutionary process.

It evolves to escape most therapies we use against it, no matter how potent.

Every tumor is its own different word, and everything cancers want is to live.

Fighting an evolutionary desire to live is hard.

Cancer is a living process, but not a foreign one.

It is a living process living its own parallel life, as part of ours.

Gradually, it hijacks and overtakes our own inner working mechanisms.

This is why, conceptually, cancer is a very hard problem.

Because, in a sense, the enemy is ourselves.

It’s all different pieces of the same puzzle.

Now let’s get back to where we started.

1938: early cancers can be cured.

85 years later: can we now cure all early cancers?

Unfortunately not, and it’s also really, really complicated.

But, do we understand better today than in 1938 why this is?

Definitely.

And this takes us so much closer to where we want to be.

This was it for today's journey.

I will regularly post such essays on my newly-opened Substack, to make genomics knowledge more accessible.

Please subscribe if this is of interest to you.

https://simocristea.substack.com/p/why-we-cant-yet-cure-cancer

Why curing cancer is so damn hard

The 2 main forces driving cancer evolution:

1. mutation. new progeny arise, increasing diversity

2. selection. populations with higher fitness (capacity to reproduce) overtake the other lineages by frequency

Tumors evolve via an iteratively malignant dance among these two. 

🤔All fields and industries are regularly disrupted. Today, we all own a smartphone.

How will the next disruption change Biology?

Accurate CRISPR editing &amp; designer babies? Fancy prediction models behind medical decisions?

Probably not. 

Rather, something totally different 🎁 

🤔 Why Biology will have its “AI moment”:

Biology is inherently messy, noisy &amp; complex.

Maths is inherently clean &amp; precise.

Closed-form solutions seldom work for Biology.

The uninterpretable &amp; non-mechanistic “dark” side of Deep Learning might just be the spirit of Biology. 

The science of #immunotherapy can cure a patient's otherwise incurable cancer.

But sometimes immunotherapy fails completely

Shockingly, we hardly know why.

A meta-analysis of #Genomics & #Transcriptomics in >1,000 immunotherapy-treated patients aims to better understand why🧵

This 2021 @CellCellPress paper is one of the best #DataScience #Bioinformatics resources out there for understanding the genetic determinants of response to immune checkpoint inhibitors (ICIs).

https://www.cell.com/cell/fulltext/S0092-8674(21)00002-7?_returnURL=https%3A%2F%2Flinkinghub.elsevier.com%2Fretrieve%2Fpii%2FS0092867421000027%3Fshowall%3Dtrue

Some context:

PD-1 & PD-L1 inhibitors are examples of ICIs.

ICI is a type of immunotherapy that un-blocks the immune system & allows it to mount attacks🤺

It does it by inhibiting checkpoints (s.a. PD-1 & PD-L1): proteins that keep the immune system from attacking its own self

This meta-analysis gathered 1,008 ICI-treated patients from 12 cohorts (with clinical outcome) with the goal of identifying predictors of immune checkpoint inhibition sensitization.

It reprocessed all data with a uniform #Bioinformatics pipeline to maximize cohort comparability.

Key takeaways from a thorough analysis👇

1. Putting all this data together
2. Consistent pipeline for re-analysis
3. Literature browsing for biomarkers (723 articles yielding 55 unique biomarkers)
4. Predictive capacity of individual biomarkers
5. Multivariate predictive model

Immune checkpoint blockade biomarkers can be grouped into four categories (used throughout the paper):

1. neoantigen-related (surrogate for T-cell reactivity)
2. immune evasion capabilities (driving resistance)
3. host factors (HLA)
4. immune infiltration capabilities

After re-preprocessing all data, the predictive capacity of each biomarker was assessed separately (odds ratio on the side).

Clonal TMB (number of non- synonymous mutations estimated to be present in every cancer cell) & total TMB had strongest effect sizes across all 12 studies

Some context:

Tumor mutational burden (TMB) is a known immunotherapy response biomarker, but its applicability across cancer types is controversial.

Recent works showed that high TMB is a viable predictor in tumors only if CD8+ T cells positively correlate with neoantigen load.

https://twitter.com/simocristea/status/1598773246887104517?s=20&t=fLqHHj4NF-4yX3mQDUlj9g

Biomarker analysis in the second category (immune escape) showed no significant association between the level of somatic copy-number alteration (SCNA), measured using the weighted genome instability index, and checkpoint inhibitors response.

More context:

Whether the copy alteration landscape of a tumor is predictive of immunotherapy response or not is another controversial topic, with recent papers reaching contradictory conclusions.

Likely predictive value is small, and larger sample size is needed.

https://twitter.com/simocristea/status/1613542247093653512?s=20&t=fLqHHj4NF-4yX3mQDUlj9g

Interestingly, when analyzing biomarkers of immune infiltration, CXCL9 expression was a strong predictor of immunotherapy response.

CXCL9 contributes to regulating immune cell migration, differentiation and activation.

https://www.cancertreatmentreviews.com/article/S0305-7372(17)30199-8/fulltext

Next, the data was analyzed separately per cancer/drug type, with at least 2 cohorts available: melanoma anti-PD-1/L1 and anti-CTLA-4, urothelial carcinoma anti-PD-1/L1 & non-small cell lung cancer anti-PD-1/L1.

The significant biomarkers were similar to the pan-cancer analysis.

Specifically, more than 80% of biomarkers significant in individual histologies were also significant pan-cancer.

This might also be a consequence of the small sample size in separate cohorts.

Effect sizes also varied: e.g. TMB was weaker in melanoma vs. urothelial carcinoma.

Biomarkers within each category correlated highly with eachother (e.g. TMB & clonal TMB, or immune infiltration CD8A & CXCL9).

However, the correlation between separate biomarker categories was generally low,suggesting potential of combining categories into a multivariable test.

Next, the paper quantified the total proportion of variance in checkpoint inhibitors response that could be explained by all biomarkers mentioned above.

For 3/4 studies, ~60% variance was explainable, suggesting that ~40% of the factors determining outcome remain unknown.

The natural next question is whether & how could these biomarkers be combined into a single metric predicting response.

XGBoost was used for this, on the 11 most significant biomarkers from the individual analyses.

The predictive model was first applied on the 4 different cancer types separately. The results were similar, hence a pan-cancer model was trained on all n=1,008 samples, with the goal to predict response to immune checkpoint inhibitors.

The pan-cancer multivariable model outperformed a TMB-only model on predicting checkpoint inhibitor response across three independent test cohorts totaling 350 samples.

A simple two-marker model (clonal TMB and CXCL9 expression) performed reasonably well, also outperforming TMB.

Another interesting avenue are mutational signatures: n = 774 samples were use to calculate the proportion of mutations in a given sample attributable to each signature from the complete set of COSMIC mutational signatures (Alexandrov et al. 2015).

Testing for association between the proportion of mutations fitting each signature & immune checkpoint response found 5/20 significant mutational processes:

signature 1A (aging)
signature 4 (tobacco)
signature 7 (UV)
signature 10 (POLE)
signature 2+13(APOBEC)

These associations remained significant after correcting for total mutation count (TMB), suggesting that not only the amount, but also the type of mutations are important.

All signatures (except aging) were associated with an improved chance of response to checkpoint inhibitors.

Even though this study found that the total burden of SCNAs is not predictive of response, some specific alterations were still significantly associated.

One of these significant alterations was loss of 9q34.

However, this result hasn't been experimentally validated.

Another example is the focal amplification of CCND1: tumors with CCND1 amplification were found to have significantly lower rates of response to immune checkpoint inhibitors.

Lastly, this comprehensive analysis could be the foundation for clinical implementation.

A standardized assay combining exome sequencing & targeted gene expression data could be used on patient tumor samples to compute a multivariable predictive score of immunotherapy response.

For clinical implementation, it is clear though that specific cancer/treatment models will be more accurate than pan-cancer models.

However, available immunotherapy data cohorts are still very low in sample size.

Therefore, for now, a pan-cancer model could still be informative

To conclude, this paper is a must-read if you are working on #cancer #immunotherapy #Datascience

Still, identifying the factors predictive of immunotherapy response is a difficult task,and likely much larger sample sizes will be needed for truly reliable & implementable findings

This huge effort comes from the labs of @kevlitchfield @JL_Reading @NickyMcGranahan @CharlesSwanton in London.

Kudos 👏 to all authors for an impressive work!

The complex genomics behind predicting immunotherapy response

Hands-on weekend learning

Here’s everything you need to know to code your own #GraphNeuralNetworks drug screening predictor from scratch in less than 1 hour 🤯 

Code together with @PetarV_93

It will walk you though the basics &amp; learn you how to think.

https://m.youtube.com/watch?v=e40y0WtRAWM

This live coding session is a great opportunity to learn how to apply Graph Neural Networks to practical problems in #Biology

And learn how to think about designing the optimal network to answer your own custom biological question.

Code your own Graph Neural Network for drug prediction

‼️Why do we say that #methodsmatter?

Here’s a telling example.

Two very recent studies asked the exact same cancer #immunotherapy prediction question, on the exact same dataset.

They reached contradictory conclusions.

Why?

Slight changes in the #Bioinformatics methods used👇 https://twitter.com/NCIEytanRuppin/status/1610202233521438721

The changes are:
1.analyze data per cancer type vs. pan-cancer
2.use cutoff of 0.2 vs. 0.1 for calling a copy number alteration

The contradictory conclusions are:
1.aneuploidy score (AS) is not predictive of immunotherapy response
2.a new metric is predictive &amp; outperforms AS

Here’s links to the two studies, Spurr et al was published in November 2022 &amp; Cheng et al was posted on biorxiv in December 2022

https://www.nature.com/articles/s41588-022-01235-4

https://www.biorxiv.org/content/10.1101/2022.12.28.522101v1

This example is conceptually representative for some of the core principles that define #Bioinformatics #DataScience.

Minor changes in methods or statistical frameworks used can qualitatively change entire narratives.

Choice of thresholds is important

Reproducibility matters‼️

Why do we say that methods matter in Bioinformatics? An immunotherapy example

We all think we're one of a kind.

But sometimes, we come across someone who looks just like us!

A @CellReports study tested the DNA of "fake twins".

Guess what:

They also share 🧬DNA variants related to facial features & behavior 🤯

Surprised or not really? Let’s dig in🧵👇

First, let’s see why this study might NOT surprise you.

Monozygotic twins share almost identical facial traits & the same DNA sequence. Therefore, looking-alike strangers could follow a similar pattern.

Still, looking-alike strangers are not twins! So we can’t know for sure if:

a. they share more of their genome than random people

b. if yes to (a), how much they share & what would be the functional role of the genes on which such SNVs are

c. how about #multiomics similarity, such as DNA methylation or microbiome (different in monozygotic twins)?

Now, let’s see what this paper did.

1️⃣   Identify Human doubles (doppelgängers): pairs of people with very similar facial features, such that one can be confused with the other.

The pairs were selected from the photographic database of the artist Francois Brunelle.

This artist has been collecting artwork on Doppelgängers from around the world since 1999.

For this study, the authors started off with assessing 32 pairs.

http://www.francoisbrunelle.com/webn/e-project.html

2️⃣   Devise a measure of “facial likeness” by combining the results of three #Deeplearning algorithms, each of which provides a score from 0 to 1 as to how similar two faces are: 1 denotes the same face, 0 corresponds to two completely different faces👇

a. the deep CNN Custom-Net algorithm, designed for surveillance (www.hertasecurity.com)

b. MatConvNet algorithm, designed for facial classification (Vedaldi and Lenc,2015)

c. Microsoft Oxford Project face API, designed for generalized facial analysis (https://azure.microsoft.com/es-es/services/cognitive-services/face/)

It’s interesting to understand how such algorithms work. For example, these are the 27 face landmarks computed by the Microsoft algorithm, based on which image similarity is inferred.

Unsurprisingly, 25 out of the 32 look-alike pairs were found to be correlated by at least 2 of the 3 softwares.

The plot below shows the 16 of 32 pairs (50%) which scored consistently high on all algorithms (LALs). For comparison, MZ-twins are monozygotic twin faces and Non-LALs are faces of random people.

Clearly, the selected look-alikes are more similar than random, less than MZ twins.

3️⃣   Genetic analyses performed on the 16 pairs:
- genome: SNP microarrays
- epigenome: methylation microarrays
- microbiome

Based on various clusterings of the SNP profiles, 9 out of the 16 pairs were considered to be “ultra look-alike”.

The rest of the paper focuses on these.

19,277 SNPs (located on 3,730 genes) were shared among the pairs.

Among those, 1,794 genes were “face-related” (as gathered from databases or prior GWAS results).

171 of the shared SNPs caused amino acid changes, affecting 158 genes, most of which related to facial determinants

4️⃣   Different epigenome and microbiome: even though they do share SNPs, the analyses in this paper concluded that Doppelgängers do not also share DNA methylation patterns or microbiome (at least not to a similar extent as SNPs).

5️⃣   Behavior: Behavioral features, extracted from biometric & lifestyle questionnaires, were used to calculate a likeness score of the 32 look-alike pairs.

Doppelgängers had higher likeness scores than the rest & higher similarity on particular behavioral features.

To me, it remains unclear though how robust these results are to changes in e.g. the metric used, the features on the basis of which the metric is computed, or increasing sample size (16 pairs is still low).

Robustly proving this behavioral link needs further analyses.

Let’s recap!

This paper looked at #multiomics of look-alike pairs, i.e. strangers with very similar facial features.

It found that they share more SNPs than expected (e.g. vs. random people) & that most SNPs relate to facial features.

No epigenome or microbiome sharing.

Some limitations of the study:
- small sample size
- more sensitivity analyses needed to increase robustness of results, s.a. varying more the metrics/cutoffs used
- potential selection biases by repeatedly shrinking down the already small sample size to highest signal

The dataset brought forward by this paper is unique, interesting & valuable.

The findings here are in line with the intuition that phenotypic similarities correlate with genomic ones and potentially also with behavior.

Data & custom scripts are openly available & downloadable.

Finally, here is the link to the @CellReports paper.

Congrats to all authors for this paper and for addressing this interesting question! @ManelEsteller @Alfons_Valencia

https://www.cell.com/cell-reports/fulltext/S2211-1247(22)01075-0

The hidden genetics of fake twins (doppelgängers) hints to shared DNA variants

Like it or not, it's happening!

2023 is the year of #DeepLearning in Healthcare.

🚨New editorial @natBME on how Graph Neural Nets & Transformers are shaping Computational Medicine via contextual learning.

Editorial written part by editors, part by #chatGPT🤫

Main takeaways👇

This editorial highlights some of the most recent articles published in @natBME on #DeepLearning from medical images.

It discusses specifically how contextual information can substantially improve model performance and clinical interpretability.

https://www.nature.com/articles/s41551-022-00997-w

Images are very popular in medicine.

They are used all the time for a very wide variety of tasks.

Images are particularly relevant in #oncology, as the most important tool in diagnosing & tracking disease.

Traditionally, humans need to interpret medical images.

But humans are humans, and they sometimes make "human mistakes".

Also, the amount of images a human can process is constant, hence the total amount of processed images scales linearly with the number of folks interpreting.

An #AI algorithm, on the other hand, can scale much, much more efficiently.

Even if it might lack the depth and complexity in interpretation that for example an experienced pathologist has, it can still streamline the entire process by reducing cost & assisting medical experts.

1️⃣ Graph Neural Networks (#GNNs) are particularly relevant in image processing for tasks such as image segmentation, disease classification, and in-depth characterization of tissue organization.

I wrote before on the applications of Graph Neural Networks to #singlecell biology.

https://twitter.com/simocristea/status/1586004178383425536?s=20&t=iK2hkd0Ot6YqmLSwR0PS_g

Here is another example of a #GNN with biotech applications: a tool for dynamic time-dependent data, such as:
- series of images from different progression points in time
- temporal snapshots of protein-protein interaction networks or gene-expression networks.

https://twitter.com/simocristea/status/1597294880027705344?s=20&t=iK2hkd0Ot6YqmLSwR0PS_g

This perspective paper👇 is also mentioned in the editorial.

It's a great overview for folks wanting to understand how so much of biology can be thought of as networks, and therefore modeled using Graph Neural Networks (for example).

https://twitter.com/simocristea/status/1587516737246748673?s=20&t=RrxkTDhwgxaLBKKuxM_obA

On a more molecular level, the editorial briefly discusses a very nice new paper by @james_y_zou, which trains GNNs on spatial protein profiles from multiplexed immunoflorescence.

The GNN can extract clinically relevant features from the tumor microenvironment of cancer patients

When applied on tissues from patients with head-and-neck and colorectal cancers, the model identifies spatial motifs associated with cancer recurrence and patient survival following treatment.

https://www.nature.com/articles/s41551-022-00951-w

Another important study shows how enhancing Graph Neural Network models with histopathological features from the tumor microenvironment extracted from whole-slide images can better predict cancer prognosis in kidney, breast, lung and uterine cancers.

https://www.nature.com/articles/s41551-022-00923-0

Treating histological tissues as graphs, with image patches as nodes, assures that different regions of the same whole-slide image are inter-connected & dependencies can be explicitly modeled (e.g with GNNs).

This opens up lots of opportunities vs. treating patches independently

However, such a strategy is also computationally expensive.

The paper addresses this explicitly by implementing a patch-aggregation strategy with edge-level attention.

A summary of this paper is also presented in this News & Views piece

https://www.nature.com/articles/s41551-022-00924-z


2️⃣Transformer models (e.g. #chatGPT)

Transformers learn the meaning of words and the structure of language via contextual clues.

One of the main advantages of such models is that they are self-supervised & don't need explicit annotations in the training data.

Extrapolating, @pranavrajpurkar & @AndrewYNg use Transformers to identify pathologies in unlabelled chest Xray images.

As contextual clues (text Transformer), a vision Transformer uses learned pathology features from the raw radiology report associated with each Xray image.

The raw radiology reports then act as a natural source of supervision.

The performance of the self-supervised model is comparable to that of radiologists.

That's quite impressive.

https://www.nature.com/articles/s41551-022-00936-9

On a different note, this Review paper highlights something very important: how openly releasing pre-trained models changes the paradigm from model building to model deployment.

I believe this is an important conceptual shift for Computational Medicine.

https://www.nature.com/articles/s41551-022-00898-y

Indeed, such large self-supervised #DeepLearning models need lots of training data.

Getting so much data is usually prohibitive.

Even more, training on so much data is not feasible computationally.

Access to pre-trained models changes this paradigm by increasing accessibility.

To wrap up: self-supervised #DeepLearning models are already very prolific in their applications to Computational Medicine.

Context Learning makes them much better.

They are not perfect, but they need not be.

There’s immense opportunity for improving & extending such models.

In my view, one very exciting, and still under-explored area, is the application of such models to DNA sequencing data.

In particular, decoding broken DNA such as the one from cancer patients, has the potential to be very insightful.

2023 will certainly be big on this!

How contextual learning (Graph Neural Networks & Transformers) is shaping biomedicine

🤔 Science is a deeply creative process.

We literally create knowledge that hasn't previously existed.

Maybe it’s then implicit that creating new knowledge can’t be easy.

To all scientists out there: doing science, in whatever form, is such a unique process.

🧫🧬🔬⚗️🔭🥼🧪🧮 