Alejandro Tejada Lapuerta

Wow this is trained on 21 patients data in 1 cancer. 😬 For context we are training these types of models on almost 4,000 patients across all modalities paired. “Our training data comprises of data collected from 21 patients across different stages of lung adenocarcinoma.”

labs will publish details on arch, optim, objectives, scaling, kernels, literally everything except data and academia will be astounded for the hundredth time, wondering to itself where the secret sauce is

Me defending my O(n^3) solution to the coding interviewer.

Across many labs, SCALING single cell foundation models has had mixed success. We think the key is CONTEXT. *Spatial* single cell RNA data preserves the natural biological context of gene expression within animal tissue — in our case, tissue from human patients. When we train models on large, diverse spatial datasets (100M cells across a dozen cancer types) we see BIG benefits from bigger models and longer context (effectively how much patient data the models see at once.) Interestingly, the bigger the model, the better it gets with longer context. Maybe only larger models can capture complex spatial gene expression patterns across large regions of tissue. We think that scaling SPATIAL single cell models is the way — maybe the only way — to discover new, therapeutically actionable biology across patients and solve the CLINICAL TRANSLATION problem that plagues drug development.

Big day for @NOETIK_ai ! We're licensing our OCTO-VirtualCell models in lung and colorectal cancer to @GSK, one of pharma's best at using AI to help patients. @Ronalfa is not being licensed at this time.

We're excited to release tcellMIL, an attention-based multiple instance learning model for predicting patient outcomes after CAR T cell therapy for lymphoma and nominating cell design strategies in #neurips2025 AI4D3! ai4d3.github.io/2025/papers/16…

🧬 Excited to share Nicheformer out now in Nature Methods! A transformer foundation model linking single-cell & spatial omics, learning spatial context from gene expression to map tissue organization. Led by Ale Tejada & Anna Schaar 👏 👉 nature.com/articles/s4159…

@latkins This code is extremely dangerous. Here, I improved it.

GRPO is so far behind the frontier we use GDPR in europe fucking open-source peasants

Want your biology AI model to learn spatial relationships? Add images as a modality! scPortrait enables fast, standardized generation and use of single-cell image datasets, powering AI/ML-based discovery. GitHub ⭐️: github.com/MannLabs/scPor… Preprint 📚: biorxiv.org/content/10.110…

🚀 Excited to share scPortrait! Led by Sophia Mädler & Niklas Schmacke w/ the Mann lab — a new @scverse tool for standardized single-cell image data. Enables ML-ready extraction, >1B cell processing, cross-omics, & cancer macrophage insights. 🔗 biorxiv.org/content/10.110…

CAR T cells showcase the enormous potential of cell therapies, but often fail due to lack of evolutionary optimization. Today in @Nature, we use #CELLFIE to engineer cell therapies at scale and share the largest resource of CRISPR screens in CAR T cells. nature.com/articles/s4158…

Great company and exciting times! Very proud of the collaboration, the @NOETIK_ai team, and the OCTO-Virtual Cell work highlighted here — but it feels like an eternity ago! On to simulating patients and clinical trials. Stay tuned.

🛡️How do macrophages tailor their defenses to different pathogens? Our new paper in @CellSystems combines dense multi-omics time series with high‐content CRISPR screens (CROP-seq) to map the regulatory landscape underlying macrophage immune responses. #Immunity #Screening (1/9)

What xAI researchers see when they run “nvidia-smi”

Excited to share that a scaled up version of Gemini DeepThink achieves gold-medal standard at the International Mathematical Olympiad. This result is official, and certified by the IMO organizers. Watch out this space, more to come soon! deepmind.google/discover/blog/…

started my new job! i will be doing a combination of ml engineering and writing at @NOETIK_ai. picture very related tldr: we are building foundation models of tumor microenvironments to: 1. stratify patients into responders/non responders for cancer therapy trials 2. inform which oncologic drugs are worth licensing it is very rare to come across an ai biotech that have clearly profitable and specific use-cases that lack equivalent non-ai solutions. even rarer that their approach is unlike anyone else’s. rarest of all is having early results that demonstrate the success of the approach noetik fits the bill for all of these. very excited to be a part of it! longer post coming soon also: dm me if you have a perplexing cancer therapy! we’re seeking partners!

I'm honoured to join @CRUK_CI and @Cambridge_Uni as a Junior Group Leader of AI for Cancer. My group will focus on reasoning, multimodality, hypothesis-making, and more to decode disease and health. A few positions are opened — consider applying to outsmart cancer together!