Biological Systems Modeling

I’m excited to share a new paper from our AI researchers at Recursion and Valence Labs, published today in Nature Biotechnology. At its core, this work is about a fundamental challenge in drug discovery: understanding, early and with confidence, how cells will respond to a perturbation — and ultimately, to therapeutic intervention. Today, we largely rely on the wet lab to answer that question. We run experiments, generate data, refine hypotheses, and repeat. This approach works, but it is slow and resource-intensive, especially when exploring a space with an almost infinite number of possible perturbations and combinations. At Recursion, we are working toward a more predictive, in silico approach – and building a deeper, more systematic understanding of how transcriptomics can serve as a bridge between perturbational biology and human disease. We’re modelling biological responses computationally, generating stronger hypotheses, and using the lab to validate and refine — rather than explore from scratch. In this paper, we introduce TxPert: a model designed to predict how a cell’s transcriptomic state, or which genes are turned on or off, changes in response to genetic perturbations. Importantly, TxPert can generalize beyond the data it has seen. It can predict responses to perturbations that were not directly observed during training including unseen single-gene perturbations, novel combinations of perturbations, and known perturbations in new cell types. In some settings, its performance begins to approach the reproducibility of experimental measurements. What makes this possible is not just the model itself, but the combination of: ✅ large-scale perturbation datasets ✅ structured biological knowledge, including multiple knowledge graphs ✅ and the ability to generate predictions that can be directly tested and validated in the lab and fed back to continue improving the model That connection matters. It’s how we move from interesting models to results we can trust. Over time, this is how we enable more computationally driven hypothesis generation, faster iteration, and ultimately better outcomes for patients. This is our vision for the Virtual Cell. TxPert is an important step in that direction and we’ve progressed considerably since this work was first submitted. Stay tuned — much more to come! Huge congratulations to the team behind this work — Frederik Wenkel, Wilson Tu , Cassandra Masschelein, Hamed Shirzad, Liam Hodgson, Ihab Bendidi, Cian Eastwood, Shawn Whitfield, Craig T. Russell, Yassir El Mesbahi, Marta Fay, Berton Earnshaw, Emmanuel Noutahi, PhD, and Alisandra Denton. Read the paper in Nature Portfolio here: https://lnkd.in/eNtv6UB3 #VirtualCell #AI #DrugDiscovery #Biology #NatureBiotechnology

Peptides, short chains of amino acids, are fundamental building blocks in biology, showing a remarkable duality of simplicity and versatility. Their sequences dictate their properties and behaviors, enabling them to self-assemble into diverse structures such as fibers, hydrogels, and nanotubes. These assemblies play crucial roles in biological systems and have applications in drug delivery, tissue engineering, and catalysis. However, the factors driving peptide self-assembly—spanning sequence, concentration, pH, and solvent—remain poorly understood due to the fragmented nature of existing data. What governs the self-assembly of peptides into specific structures? Our work addresses this question by integrating literature mining, machine learning, and systematic analysis to uncover the deeper principles behind peptide self-assembly. We curated a dataset of over 1,000 experimental entries from academic literature, capturing peptide sequences, experimental conditions, and resulting assembly phases. We were able to connect the dots between a large set of distinct studies and elucidate principles that cut across individual experimental results, forming a unifying model. We achieved this by: ➡️ Augmented Literature Mining: Fine-tuned a LLM on manually curated data to extract detailed experimental parameters with significantly improved accuracy. This approach accelerated what is typically a time-consuming process while maintaining high fidelity to complex, nuanced data. ➡️ Machine Learning for Phase Prediction: Developed ML models to predict self-assembly phases based on peptide sequences and experimental stimuli. Our model demonstrated over 80% accuracy, providing actionable insights into how variables like peptide concentration and solvent conditions influence assembly outcomes. ➡️ Iterative Improvement: Designed a workflow where newly extracted data can augment the dataset, continuously refining both the ML models and our understanding of peptide self-assembly mechanisms. Broader Implications While peptides are made up of just a few amino acids, their ability to self-assemble into highly ordered and functional structures—such as hydrogels, nanotubes, and even crystals—is astonishing. This phenomenon is driven not by strong chemical bonds, but by weak, non-covalent interactions like hydrogen bonding, van der Waals forces, and hydrophobic effects. By combining human expertise with machine intelligence, we not only accelerate discovery but also promote deeper reflection on the governing rules of complex systems. We believe that many other fields could benefit from such a systematic integration of human insight and machine learning. Code: PeptideMiner, https://lnkd.in/eWBSTDjK Paper: Zhenze Yang, Sarah Yorke, Tuomas Knowles, Markus J. Buehler, Learning the rules of peptide self-assembly through data mining with large language models, https://lnkd.in/ekFQ3giK, 2024

🧬 AI just wrote the code for living organisms and they actually work. Researchers at Arc Institute and Stanford have achieved something unprecedented: using genome language models Evo 1 and Evo 2 to generate 16 viable bacteriophage genomes from scratch the first time AI has designed complete, functional genomes that work in the real world. Think about that for a moment. Not just designing a protein. Not simulating a genome on a computer. Actually creating living viral systems with substantial evolutionary novelty that infect bacteria and replicate successfully. Here's what makes this revolutionary: In 1977, ΦX174 was the first genome ever sequenced. In 2003, it was the first genome chemically synthesized. Now in 2025, it's the template for the first AI-generated genomes. We've gone from reading DNA, to writing it, to designing it. The results? Several AI-generated phages outperformed the wild-type virus, with one variant called EVO-Φ69 being 65x more powerful than natural viruses. One even uses an evolutionarily distant DNA packaging protein that researchers wouldn't have rationally designed. The implications are staggering: → Cocktails of these generated phages rapidly overcome antibiotic-resistant bacteria → Accelerated development of phage therapies for drug-resistant infections → A blueprint for designing synthetic biological systems at genome scale → Foundation for creating useful living systems with entirely novel capabilities This isn't science fiction anymore. AI models trained on 2 million bacteriophage genomes can now propose new genetic codes—and 16 out of 302 designs actually worked MIT Technology Review. We're watching the birth of generative biology in real-time. The ability to design life at the genomic level opens possibilities we're only beginning to imagine. What do you believe the opportunities and risks are of this virus making AI? The conversation is just beginning. 📄 Study: https://lnkd.in/ddP3Fjdp #SyntheticBiology #AI #GenerativeAI #Genomics #Biotechnology #Innovation #Science

Can large language models be used in biotech? The short answer is yes. While LLMs are often associated with chatbots, their capabilities extend beyond that. In biotech, much of the data comes in the form of sequences – like nucleotides in DNA, or amino acids in proteins. Similar to sentences in natural language, these biological sequences have unique semantic meanings based on the arrangement of their components. When input data is fed into an LLM, a transformer converts these sequences into contextual vectors using its attention mechanism. This process allows the model to understand the context and relationships within the data, enabling it to predict subsequent elements. One such use case is prediction of neoantigens that enable targeting tumor cells in personalized cancer immunotherapies. Neoantigens are tumor-specific mutated peptides presented on the surface of tumor cells because they bind to human leukocyte antigen (HLA) molecules. LLMs can predict this binding affinity. This allows the development of personalized therapies that use the patient's own immune system to kill tumor cells without damaging healthy tissues.

Physics-Informed Neural ODEs with Scale-Aware Residuals for Learning Stiff Biophysical Dynamics Kamalpreet Singh Kainth, Prathamesh Dinesh Joshi, Raj Abhijit Dandekar, Rajat Dandekar, Sreedat Panat https://lnkd.in/dtRD_ide This paper tackles a critical challenge in using Neural ODEs for modeling stiff biophysical systems, specifically the difficulty in achieving stable and accurate long-term predictions. The core idea is to improve training stability and accuracy by incorporating physics-informed regularization with scale-aware residual normalization within a Neural ODE framework. The authors propose "PI-NODE-SR," which combines two key elements: 1. Physics-Informed Regularization: They leverage the known governing equations (e.g., Hodgkin-Huxley) to define a physics-informed loss term that penalizes deviations from the expected behavior. This is a standard PINN approach, but the key is how they handle the residual. 2. Scale-Aware Residual Normalization: This is the novel contribution. Stiff systems often involve state variables evolving on vastly different timescales. Directly applying a physics-informed loss can lead to one timescale dominating the training process, hindering convergence and accuracy. To address this, they normalize the residual associated with each state variable by a scaling factor. This factor is chosen to balance the contributions of each variable to the overall loss, preventing the faster dynamics from overwhelming the slower ones. They use a low-order explicit solver (Heun method) to calculate the residual. The authors demonstrate the effectiveness of PI-NODE-SR on the Hodgkin-Huxley equations, showing that it can learn from a single oscillation and extrapolate accurately over longer time horizons. Importantly, they show that the method can recover morphological features in the gating variables that are typically only captured by higher-order solvers. ----- This work directly addresses a significant limitation of Neural ODEs in scientific applications: their struggle with stiff systems. By introducing scale-aware residual normalization, the authors provide a principled way to stabilize training and improve the accuracy of long-term predictions. This is particularly relevant for: - Biophysical modeling: Accurately simulating neuronal dynamics, cardiac electrophysiology, and other complex biological processes. - Chemical kinetics: Modeling reaction networks with widely varying reaction rates. - Multi-physics simulations: Situations where different physical processes evolve on different timescales. The ability to use lower-order solvers with neural correction to achieve accuracy comparable to higher-order methods has significant implications for computational efficiency. While the method's sensitivity to initialization is noted, the overall approach offers a promising direction for developing more robust and efficient Physics-Informed Neural ODEs for a wide range of scientific applications.

Gut bacterium may be helping breast tumors hide from the immune system. And it appears to do it through a metabolite. Researchers found that Enterocloster bolteae, a member of the Lachnospiraceae family, became progressively more abundant as breast tumors developed. Its rise was linked to higher levels of deoxycholic acid, a secondary bile acid produced through microbial metabolism. But the metabolite did not remain confined to the gut. Deoxycholic acid accumulated inside the tumors and activated the farnesoid X receptor, or FXR, in cancer cells. That activation triggered NF-κB signaling and increased production of interleukin-6. IL-6 then recruited immune cells that can suppress antitumor immunity, including granulocytic myeloid-derived suppressor cells and T helper 17 cells. The result was a tumor microenvironment that appeared more capable of protecting the cancer from immune attack. The pathway looked like this: Enterocloster bolteae → deoxycholic acid → tumor FXR activation → NF-κB signaling → IL-6 production → immunosuppressive immune-cell recruitment → breast cancer progression This is more than another study showing that cancer is “associated” with changes in the microbiome. It proposes a specific biological chain connecting a gut organism, a circulating microbial metabolite, a receptor inside the tumor and a measurable immune response. Even more importantly, blocking FXR or IL-6 signaling weakened these effects in the experimental models. That creates several potential intervention points. Not just the bacterium itself, but its metabolic output, the tumor receptor it activates and the downstream inflammatory signal. This does not mean that modifying the microbiome can currently prevent or treat breast cancer. But it strengthens a much bigger idea: The gut microbiome may influence cancer progression from a distance by producing molecules that reach the tumor and reshape its immune environment. The next generation of cancer therapeutics may not focus only on the tumor. It may also target the microbial chemistry helping the tumor survive.

GENERATIVE BIOLOGY AI just wrote genetic instructions that cells actually followed – a breakthrough that turns biology into a programming language. For the first time ever, researchers at the Center for Genomic Regulation created AI-generated DNA sequences that successfully controlled gene expression in healthy mammalian cells. Think of it as writing software, but for living organisms. Why this matters: → The AI can design custom 250-letter DNA fragments with specific instructions like "activate this gene in stem cells becoming red blood cells but not platelets" → These synthetic enhancers worked EXACTLY as predicted when tested in mouse blood cells → Unlike previous efforts focused on cancer cells, this team worked with healthy cells, uncovering subtle mechanisms that shape our immune system → The researchers built a library of 64,000+ synthetic enhancers tested across seven stages of blood cell development Most fascinating was discovering "negative synergy" - where two factors that individually activate genes can completely shut them down when combined. This unlocks precision we never had before. The implications are enormous for gene therapy. Instead of being limited to DNA sequences evolution produced, we can now design ultra-selective gene switches customized to specific cells and tissues - potentially making treatments more effective with fewer side effects. Full paper: https://lnkd.in/en3bGZP9 Follow-up with @EricTopol's post about curing rare diseases with the existing genomic technology stack https://lnkd.in/eGCYMjGJ

Synthetic biology is - quite literally - our future. A goundbreaking new biological foundation model Evo2 achieves state-of-the-art prediction of genetic variation impacts and generates coherent genome sequences, spanning all domains of life. A diverse team from leading research institutions including Arc Institute Stanford University NVIDIA University of California, Berkeley trained the model on 9.3 trillion DNA base pairs and has fully shared all code, parameters, and data. A few highlights from the paper (link in comments) 🔬 Zero-shot prediction achieves state-of-the-art accuracy in genetic variant interpretation. Evo 2 can predict the functional consequences of genetic mutations across all domains of life without specialized training. It surpasses existing models in assessing the pathogenicity of both coding and noncoding variants, including BRCA1 cancer-linked mutations. This generalist capability suggests Evo 2 could revolutionize genetic disease research, reducing reliance on expensive, manually curated datasets. 🛠 Genome-scale generation paves the way for synthetic life design. Evo 2 can generate full-length genome sequences with realistic structure and function, including mitochondrial genomes, bacterial chromosomes, and yeast DNA. Unlike prior models, Evo 2 ensures natural sequence coherence, improving synthetic biology applications like engineered microbes or artificial organelles. This sets the stage for programmable biology at an unprecedented scale. 🧬 Unprecedented long-context understanding revolutionizes genomic analysis. Evo 2 operates with a context window of up to 1 million nucleotides—far beyond the capabilities of previous models—allowing it to analyze genomic features across vast distances. This ability enables it to accurately identify regulatory elements, exon-intron boundaries, and structural components critical for understanding genome function. Its long-context recall is a major breakthrough for interpreting complex biological sequences. 🎛 Inference-time search enables controllable epigenomic design. Evo 2’s generative abilities extend beyond raw DNA sequence to epigenomic features, allowing researchers to design sequences with specific chromatin accessibility patterns. This approach successfully encoded Morse code messages into synthetic epigenomes, demonstrating a new method for controlling gene regulation via AI. This could lead to breakthroughs in gene therapy and epigenetic engineering. 🔮 Future potential: Toward AI-driven biological design and virtual cell modeling. Evo 2 represents a major leap toward AI-powered genomic engineering. Future iterations could integrate additional biological layers—such as transcriptomics and proteomics—to create virtual cell models that simulate complex cellular behaviors. This could revolutionize drug discovery, genetic therapy, and even synthetic life creation.

Can AI help scientists discover new scientific laws? Many of the most important advances in science are expressed as mathematical relationships. But discovering those relationships directly from data remains one of the hardest problems in AI. So how do we solve it? Huge congratulations to Evgeny Saveliev, Samuel Holt and collaborators on their [ICML] Int'l Conference on Machine Learning 2026 paper! Influence-Guided Symbolic Regression: Scientific Discovery via LLM-Driven Equation Search with Granular Feedback Paper link: https://lnkd.in/e89DWwxd Developed through collaborations spanning academia and industry, including AstraZeneca, IGSR helps AI discover interpretable scientific relationships from complex data. Most LLM-based approaches judge an equation as a whole. IGSR instead computes per-term influence scores, quantifying how much error increases when individual terms are removed. Useful terms are retained while irrelevant ones are pruned. The propose-and-prune cycle is embedded within a Monte Carlo Tree Search (MCTS) framework. By balancing exploration of new equation structures with exploitation of promising candidates, IGSR can efficiently navigate a vast search space while avoiding local optima. Across +100 symbolic regression problems and real-world biological datasets, IGSR consistently outperformed existing approaches. > #1 rank on predictive performance > Scales to 263 features  > Strong predictive + structural recovery All while maintaining sparse, interpretable equations. Working with Prof. David Bentley's group, the team applied IGSR to RNA Polymerase II pausing data generated through cutting-edge experimental biology. The framework proposed a novel hypothesis: "DNA methylation suppresses transcriptional pausing in gene bodies." The hypothesis was tested experimentally. The results showed that suppressing DNA methylation with a drug used in the clinic to treat leukemia patients caused a significant increase in RNA Polymerase II pausing, providing strong support for the relationship identified by IGSR. The AI hypothesis held up. Most ML optimises a loss. IGSR optimises for understanding. By combining AI, statistical learning, and cutting-edge experimental biology, this work demonstrates how machine learning can help generate and validate new scientific knowledge. #ICML2026

The tumor microenvironment (TME) is an active participant in cancer progression, and a recent review summarizes something the field is still working to understand. Tumors harbor microbiota, which correlate with immune outcomes. Metabolites mediate many observed effects in mice. For example, L. reuteri converts dietary tryptophan into indole-3-aldehyde, boosting CD8+ T cell immunity and improving ICI efficacy. Where causality is supported: Fusobacterium nucleatum contributes to immune evasion and chemoresistance via several defined mechanisms. It architects the TME by inhibiting NK cells and T cells and recruiting MDSCs and Tregs. It drives chemoresistance in colorectal cancer by activating autophagy and inhibiting apoptosis in tumor cells. Where the picture gets murkier: Associations between intratumoral microbial diversity and patient survival are compelling but difficult to interpret causally. Microbial composition could reflect tumor biology, immune status, or previous antibiotic exposure. Correlation in sequencing studies may not reflect actual contributions to tumor progression. What remains open to debate: Whether deliberately modulating the intratumoral microbiome can reproducibly shift immune outcomes in humans is largely unresolved. The metabolite layer is promising precisely because it offers tractable, measurable intermediaries between microbial community structure and host response. But the field needs better tools to distinguish passenger microbiota from functional contributors. The map is not the territory, but knowing where the map is reliable and where it's still fuzzy can help us better navigate microbiota and TME research. Reference: Yao, Y., Zhu, Y., Chen, K. et al. Microbiota in cancer: current understandings and future perspectives. Sig Transduct Target Ther 11, 39 (2026). doi: 10.1038/s41392-025-02335-3