iiinova — where your digital self keeps evolving

Guillermo Prol-Castelo @gprolcastelo.bsky.social · 04/05/2026

I would like to appreciate the work from my supervisors and co-authors, Davide Cirillo and @alfonsovalencia.bsky.social . I would also like to acknowledge the support from our home institution, the @bsc-cns.bsky.social

011

Guillermo Prol-Castelo @gprolcastelo.bsky.social · 04/05/2026

our systematic literature reviewper on deep learning for modeling cancer over time just published open access in Briefings in Bioinformatics! shared a preprint summary here last year—peer review really improved it; if you're into cancer AI or temporal modeling, read it here: doi.org/10.1093/bib/...

doi.org

Deep representation learning for temporal inference in cancer omics: a systematic literature review

Abstract. Deep learning methods, including deep representation learning (DRL) approaches such as variational autoencoders (VAEs), have been widely applied

140

Reposted by Guillermo Prol-Castelo

BSC-CNS @bsc-cns.bsky.social · 27/11/2025

🤖 Torna la #BitsxlaMarató! I ja van 7 edicions col·laborant amb #LaMarató3CAT 💻 El BSC és un dels organitzadors d’aquesta hackathon, que aquest any compta amb la col·laboració de l'Institut Català d'Oncologia.

131

Reposted by Guillermo Prol-Castelo

More Perfect Union @moreperfectunion.bsky.social · 04/11/2025

OpenAI pirated large numbers of books and used them to train models. OpenAI then deleted the dataset with the pirated books, and employees sent each other messages about doing so. A lawsuit could now force the company to pay $150,000 per book, adding up to billions in damages.

18772062204

Reposted by Guillermo Prol-Castelo

EvenFlow Project @evenflowproject.bsky.social · 19/06/2025

🎉 Congratulations to EVENFLOW partner @gprolcastelo.bsky.social on the publication of his second PhD pre-print! 📄 A contribution to understanding VAEs in cancer progression research, supported by the @evenflowproject.bsky.social project. Learn more 👇 #CancerResearch #DeepLearning

031

Guillermo Prol-Castelo @gprolcastelo.bsky.social · 12/06/2025

The full study is already available www.biorxiv.org/content/10.1101/2025.05.29.656750

biorxiv.org

011

Guillermo Prol-Castelo @gprolcastelo.bsky.social · 12/06/2025

I would like to thank Davide Cirillo and @alfonsovalencia.bsky.social for their supervision

111

Guillermo Prol-Castelo @gprolcastelo.bsky.social · 12/06/2025

Studying cancer progression with DRL is challenging due to limited longitudinal data and the absence of methods for real-time tracking. Still, solving these problems could lead to major breakthroughs in personalized treatment

121

Guillermo Prol-Castelo @gprolcastelo.bsky.social · 12/06/2025

Single cell omics data is commonly used to infer pseudo-time trajectories of cancer cells. However, these trajectories lack actual temporal information from either real time or stages.

111

Guillermo Prol-Castelo @gprolcastelo.bsky.social · 12/06/2025

It is also not clear how to study cancer in a temporal manner, given cancer progression differs from patient to patient. From the literature, we see that stages may be used as a proxy time-unit to study cancer’s time dimension.

111

Guillermo Prol-Castelo @gprolcastelo.bsky.social · 12/06/2025

Given the difficulties of performing a longitudinal study in human patients, there is a lack of follow-up data, especially in cancer.

111

Guillermo Prol-Castelo @gprolcastelo.bsky.social · 12/06/2025

We found that the most common uses of VAEs to study cancer include diagnosis, prognosis, and subtyping.

Word cloud of exclusion reasons. Most publications use deep and representation learning applied to cancer for subtyping, diagnosis, and studying prognosis and survival.

111

Guillermo Prol-Castelo @gprolcastelo.bsky.social · 12/06/2025

Cancer is a highly complex and dynamic disease, making it well-suited for analysis using DRL methods and VAEs. We wanted to elucidate the most common uses of DRL and the VAE in the study of cancer, paying special attention to the temporal component of cancer, which remains understudied.

101

Guillermo Prol-Castelo @gprolcastelo.bsky.social · 12/06/2025

DRL methods are used to learn a low-dimensional embedding from data. Specifically, the VAE can learn said representation, the latent space, keeping non-linear relationships in the original data. Moreover, it is also a generative method, as it can create new, synthetic data from the original data.

Summary figure of the most common methodologies involving representation learning and cancer studies.
On the left side of the figure are shown the most common types of data used to train VAEs particularly—and DRL methods more generally—, in cancer studies. After being passed to the encoder, data is embeded into a lower-dimensional representation, the latent space. The latent space has been commonly used for subtyping, survival anlyses (commonly as part of prognostic analyses), and, when it comes to single-cell data, also for pseudo-time trajectories inference. Possible further applications, shown on the right-hand side of the figure, include leveraging the decoder’s generative capabilities to reconstruct noisy data, and align and infer out-of-sample data, which may be used to study cancer progression in time. However, such applications of the decoder remain underexplored in biomedical research.

121

Guillermo Prol-Castelo @gprolcastelo.bsky.social · 12/06/2025

New research alert 📝❗❗❗ In our latest pre-print, 2nd of my PhD, we performed a Systematic Literature Review on the use of Deep Representational Learning (DRL), especially the Variational Autoencoder (VAE), in cancer progression research. This thread explains our main findings. (1 minute read)

253

Reposted by Guillermo Prol-Castelo

bioRxivpreprint @biorxivpreprint.bsky.social · 05/06/2025

10 Years of Variational Autoencoder: Insights from Cancer Temporal Progression Studies, a Systematic Literature Review www.biorxiv.org/content/10.1101/202…

021

Reposted by Guillermo Prol-Castelo

AI x Bio Discovery @aixbiobot.bsky.social · 05/06/2025

10 Years of Variational Autoencoder: Insights from Cancer Temporal Progression Studies, a Systematic Literature Review [new] VAE cancer omics analysis reveals temporal modeling gap (limited data). Proposes VAEs for cancer staging.

011

Reposted by Guillermo Prol-Castelo

bioRxiv Bioinfo @biorxiv-bioinfo.bsky.social · 05/06/2025

10 Years of Variational Autoencoder: Insights from Cancer Temporal Progression Studies, a Systematic Literature Review www.biorxiv.org/content/10.1101/202…

021

Reposted by Guillermo Prol-Castelo

Alfonso Valencia @alfonsovalencia.bsky.social · 13/01/2025

Happy to be part of the amazing new world of synthetic data by the hand of @gprolcastelo.bsky.social in Davide Cirillo’s group!!

042

Guillermo Prol-Castelo @gprolcastelo.bsky.social · 13/01/2025

References [1] link.springer.com/article/10.1... [2] doi.org/10.3171/2019... [3] academic.oup.com/neuro-oncolo... [4] doi.org/10.1016/j.ce... [5] jhoonline.biomedcentral.com/articles/10.... [6] linkinghub.elsevier.com/retrieve/pii... [7] doi.org/10.1016/j.cc... [8] arxiv.org/abs/1312.6114

000

Guillermo Prol-Castelo @gprolcastelo.bsky.social · 13/01/2025

Shout out to all the authors: Alejandro Tejada-Lapuerta, Beatriz Urda-García, Iker Núñez-Carpintero, @alfonsovalencia.bsky.social, and Davide Cirillo. Thanks to the #Evenflow project, and my institution @bsc-cns.bsky.social

111

Guillermo Prol-Castelo @gprolcastelo.bsky.social · 13/01/2025

5. How is this helpful for MB research? We believe our contributions will help develop better treatments for MB: labeling patients’ subgroups leads to different treatment strategies, so elucidating the most adequate is essential for an optimal recovery.

100

Guillermo Prol-Castelo @gprolcastelo.bsky.social · 13/01/2025

4.2. We have seen there are about 2,500 genes’ expressions that are unique to the G3-G4 subgroup, some of which are commonly mutated in MB: KMT2C, MYC, SNCAIP, SYNCRIP, and TP53.

100

Guillermo Prol-Castelo @gprolcastelo.bsky.social · 13/01/2025

4. What do we find? 4.1. By identifying and augmenting the patients in the G3-G4 subgroup, we achieved high classification performance, reinforcing that this intermediate group displays distinct features in comparison to G3 and G4.

Boxplot of classification results for G3, G4, and G3-G4, augmented with synthetic data.

100

Guillermo Prol-Castelo @gprolcastelo.bsky.social · 13/01/2025

3. How can we study a rare occurrence of a rare disease? We have obtained the data from the largest repository available on MB [7] and used the VAE's [8] generative ability to amplify the G3-G4 subgroup. This means we can learn from real patient data to generate new, synthetic patients.

100

Guillermo Prol-Castelo @gprolcastelo.bsky.social · 13/01/2025

2.2. Research has suggested the possibility to deem G3 and G4 as a continuum [4] but also the existence of an additional subgroup between G3 and G4 [5, 6] sometimes referred to as G3-G4. However, the limited number of patients in this intermediate case have made gaining relevant insights a challenge

110

Guillermo Prol-Castelo @gprolcastelo.bsky.social · 13/01/2025

2. Why study Medulloblastoma Subgroups? 2.1. G3 and G4 subgroups tend to be closely clustered. This tight relationship is reflected in the latest consensus classification of MB, dividing the disease into WNT, SHH, and non-WNT/non-SHH subgroups [3].

110

Guillermo Prol-Castelo @gprolcastelo.bsky.social · 13/01/2025

1. What is Medulloblastoma? Medulloblastoma (MB) is a childhood brain tumor that is classically divided into four molecular subgroups [1]: Wingless (WNT), Sonic Hedgehog (SHH), Group 3 (G3), and Group 4 (G4). It is a rare disease, with 5 cases per million in the pediatric population [2].

100

Guillermo Prol-Castelo @gprolcastelo.bsky.social · 13/01/2025

New Insights into Medulloblastoma! 🧠 I am very happy to announce the first paper in my Ph.D. thesis journey: Exploring the Boundaries of Medulloblastoma Subgroups with synthetic Data Generation -> www.biorxiv.org/content/10.1... Let’s dive in into our findings with this thread! 🧵⤵️

biorxiv.org

Exploring the Boundaries of Medulloblastoma Subgroups with Synthetic Data Generation

Medulloblastoma is a childhood brain tumor traditionally classified into four molecular subgroups. Recent evidence suggests that Groups 3 and 4 represent a biological continuum rather than distinct en...

131

Reposted by Guillermo Prol-Castelo

AI x Bio Discovery @aixbiobot.bsky.social · 02/01/2025

Exploring the Boundaries of Medulloblastoma Subgroups with Synthetic Data Generation [new] Explores subgroups by generating synthetic transcriptomics data with a VAE, supporting a continuum between groups 3 and 4.

022

Reposted by Guillermo Prol-Castelo

bioRxivpreprint @biorxivpreprint.bsky.social · 30/12/2024

Exploring the Boundaries of Medulloblastoma Subgroups with Synthetic Data Generation www.biorxiv.org/content/10.1101/202…

011