Predictive Genetic Modeling

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.

Computational modeling of gene regulatory networks has become increasingly important for understanding the biological complexity underlying disease progression in diverse cell types and for the identification of potential therapeutic targets in drug discovery. To address this inherent complexity, a team led by Christina Theodoris and Patrick Ellinor developed Geneformer, a context-aware deep learning model pretrained on large-scale transcriptomic data to enable predictions in network biology with limited data. I included the link to the full paper and brief summary below. Transfer learning enables predictions in network biology. https://lnkd.in/dXiKxTga The model and data for training are available on Hugging Face at https://lnkd.in/d3Esd6eK.  https://lnkd.in/dGxB2QaZ. Methods overview: The authors assembled a large-scale pretraining corpus called Genecorpus-30M, comprising 29.9 million human single-cell transcriptomes from various tissues. They developed a rank value encoding method to represent the transcriptome of each single cell, ranking genes by their expression within that cell normalized by their expression across the entire corpus. The researchers designed Geneformer's architecture with six transformer encoder units, each composed of a self-attention layer and feed-forward neural network layer. They implemented a masked learning objective for pretraining, where 15% of genes within each transcriptome were masked, and the model was trained to predict the masked genes. The authors optimized the pretraining process using dynamic length-grouped padding and distributed GPU training to handle the large-scale dataset efficiently. Results overview The authors showed that Geneformer boosted cell-type predictions compared to alternative methods, especially in complex multiclass prediction applications. The researchers fine-tuned Geneformer to predict gene dosage sensitivity, achieving high accuracy with limited data and generalizing well to newly reported disease genes. They applied Geneformer to predict chromatin dynamics, including bivalent domains and transcription factor regulatory range, outperforming alternative methods. The authors used Geneformer to predict network hierarchy and distinguish central versus peripheral factors within gene networks. They developed an in silico deletion approach to model gene network connections and identify dosage-sensitive genes. The researchers applied Geneformer to disease modeling of cardiomyopathy, identifying candidate therapeutic targets that were experimentally validated in an iPSC-based model of the disease.

A new DNA model can now predict how genetic variants affect splicing, gene expression, and chromatin, with single-base precision and across tissues. 1️⃣ AlphaGenome looks at 1 million letters of DNA at a time and predicts over 5,900 functional features, like gene activity, splicing, and chromatin structure. 2️⃣ It works across 11 types of genomic signals, including RNA expression, transcription start sites, splicing patterns, DNA accessibility, and 3D genome contacts. 3️⃣ Most models either look at small regions in high detail or large regions with blurry output. AlphaGenome does both: wide context and sharp, base-level predictions. 4️⃣ It identifies how variants change gene regulation across tissues and data types in under 1 second, making it useful for large-scale variant screening. 5️⃣ It is especially strong for splicing: it predicts not just splice sites, but also which sites get used and how exons are joined. It outperforms current tools like SpliceAI and Pangolin. 6️⃣ For gene expression, it more accurately predicts how variants affect RNA levels, even those far from the gene, better than other models like Enformer or Borzoi. 7️⃣ It explains known disease variants, such as non-coding mutations near the TAL1 oncogene, showing how they disrupt multiple layers of regulation. 8️⃣ Tests show its best performance comes from combining wide DNA context, fine resolution, and multiple regulatory signals in one model. 9️⃣ AlphaGenome is freely available, with tools to predict gene regulation and variant effects directly from DNA sequence. 🔟 Clinicians and researchers can use it to explore how non-coding variants may contribute to disease, without needing multiple separate tools. ✍🏻 Žiga Avsec, Natasha Latysheva, Jun Cheng, Guido Novati, Kyle Taylor, Tom Ward, Clare Bycroft, Lauren Nicolaisen, Eirini Arvaniti, Joshua Pan, Raina Thomas, Vincent Dutordoir, Matteo Perino, PhD, Soham De, Alexander Karollus, Adam Gayoso, Toby Sargeant, Anne Mottram, Lai Hong Wong, Pavol Drotár, Adam Kosiorek, Andrew Senior, Richard Tanburn, Taylor Applebaum, Souradeep Basu, Demis Hassabis, Pushmeet Kohli. Advancing regulatory variant effect prediction with AlphaGenome. Nature. 2026. DOI: 10.1038/s41586-025-10014-0

AI for predicting gene expression with experimental accuracy 🧬 I've been digging into the details of this new GET (General Expression Transformer) since it was published earlier this month. The model accepts a matrix of two inputs (Figure 1): 1. A peak from ATAC-seq data. This is a region of the genome that should be accesible to be bound by transcriptional regulators like transcription factors (TFs). 2. TF motifs. These are the stretches of DNA sequence that these regulators bind to. By masking regions of this matrix, the model learns regulatory syntax. Matching the accesibility data with RNA sequencing data, the regulatory syntax can then be used to predict gene expression—even in cell types that weren't included in training. It's an elegant formulation that showed strong performance and generalization across multiple tasks (Hence the use of "foundation model"). It seems to outperform other expression prediction models (Enformer shown) for predicting experimental data (Figure 2). One interesting part of this study: only a hand-picked set of ~1M cells from specific data sets were used for training—rather than training on the bulk ENCODE data set. This is where BioML can differ from ML more broadly. Carefully curated data + specific biological priors can deliver great performance. Super interesting. Link to study: https://lnkd.in/gtsMJmPi

Is "AlphaGenome" the key to understanding the mechanisms of human ageing? "Deep learning models that predict functional genomic measurements from DNA sequence are powerful tools for deciphering the genetic regulatory code. Existing methods trade off between input sequence length and prediction resolution, thereby limiting their modality scope and performance. We present AlphaGenome, which takes as input 1 megabase of DNA sequence and predicts thousands of functional genomic tracks up to single base pair resolution across diverse modalities – including gene expression, transcription initiation, chromatin accessibility, histone modifications, transcription factor binding, chromatin contact maps, splice site usage, and splice junction coordinates and strength. Trained on human and mouse genomes, AlphaGenome matches or exceeds the strongest respective available external models on 24 out of 26 evaluations on variant effect prediction. AlphaGenome’s ability to simultaneously score variant effects across all modalities accurately recapitulates the mechanisms of clinically-relevant variants near the TAL1 oncogene. To facilitate broader use, we provide tools for making genome track and variant effect predictions from sequence." "AlphaGenome advances efforts to decipher the genome’s regulatory code, offering a unified deep learning model that simultaneously predicts diverse functional genomic signals from megabase-scale DNA sequences. It matches or surpasses specialized SOTA models in their respective domains, which highlights the model’s robust grasp of fundamental regulatory principles encoded in DNA and its utility for mechanistic interpretation of non-coding variation. A core strength is AlphaGenome’s efficient multimodal variant effect prediction, which simultaneously scores variant impacts across all predicted modalities in a single inference pass. This integrated capability is crucial for understanding variants with complex mechanisms, as illustrated by the recapitulation of oncogenic TAL1 variant effects, and could power large scale analyses that dissect regulatory sequence elements genome-wide. Furthermore, AlphaGenome’s novel capability to directly model splice junctions enables a more holistic view of splice-disrupting variants." https://lnkd.in/erZJ2nM7

🧪 New Machine Learning Research: FREEFORM — A Breakthrough in Genotype Feature Engineering At the intersection of genetics and AI, researchers Joseph Lee, Xiaoxi Liu, and Tianlong Chen from the University of Pennsylvania, RIKEN, and University of North Carolina at Chapel Hill developed FREEFORM, a knowledge-driven framework leveraging large language models (LLMs) for genotype feature selection and engineering. - Research Goal: The study aimed to address the challenges of analyzing genotype data, including high dimensionality and limited samples, by designing a scalable LLM-driven framework to enhance feature selection and predictive modeling. - Research Methodology: The FREEFORM framework uses chain-of-thought reasoning and ensembling principles to select relevant genetic variants and engineer new features. It was tested on two real-world datasets: Genomic Ancestry and Hereditary Hearing Loss. - Key Findings: FREEFORM achieved state-of-the-art results in few-shot settings. For instance, it matched LASSO's performance on the ancestry dataset with only 10 samples, compared to LASSO's 80. On the hearing loss dataset, FREEFORM delivered higher AUC scores in low-shot scenarios. - Practical Implications: The framework's open-source implementation can be adopted to improve predictions in personalized medicine and genomic research, saving time and computational resources while maintaining accuracy. It provides a robust tool for extracting insights from complex genetic datasets. #LabelYourData #MachineLearning #Innovation #AIResearch #MLResearch #Genomics #DataScience

Hot off our recent transformer paper, we're excited to share another AI model for precision medicine! Biological data collected from patients has exploded in recent years, presenting a challenge: how do we decipher that data to understand which patients will benefit most from specific therapies?  We in the Applied Data Science team at AstraZeneca are thrilled to share our paper in Cancer Cell called "AI-Driven Predictive Biomarker Discovery with Contrastive Learning to Improve Clinical Trial Outcomes." Here, we introduce the *Predictive Biomarker Modeling Framework (PBMF)*, a neural network-powered contrastive learning process that: 🔍 Explores vast multimodal datasets to uncover predictive biomarkers in an automated, systematic, and unbiased manner  🧠 Distinguishes predictive biomarkers (which indicate a likely benefit from a specific therapy) from prognostic biomarkers (which indicate general disease outlook)  💡 Distills its outputs into an interpretable decision tree, showing what drives treatment response In our studies, the PBMF:  📊 Surpassed existing methods in finding predictive biomarkers for immunotherapy success across various cancers in clinical trial and real-world data  📈 Discovered a predictive biomarker in an early-stage trial that boosted efficacy by 15% when retrospectively applied to the corresponding phase 3 clinical trial  📈 Discovered predictive biomarkers in single-arm early phase trial data with synthetic control arms, retrospectively improving the efficacy of the corresponding phase 3 trials by at least 10% We believe the PBMF has the potential to improve the way we design clinical trials and match patients to the right therapies. It can integrate with other models like our Clinical Transformer, creating exciting possibilities to someday discover biomarkers of adverse events, dosing strategies, and even to back-translate new drug targets. Read the full paper here: https://lnkd.in/eveAnVRY   Thanks to all the co-authors: Gustavo Arango, Damian Bikiel, Gerald Sun, Elly Kipkogei, Kaitlin Smith, Sebastian Carrasco Pro, Elizabeth Choe #PrecisionMedicine #ClinicalTrials #AIinHealthcare #Biomarkers #Immunotherapy

Another interesting paper from Arc Institute. This one combines protein language models with directed evolution to rapidly engineer proteins. Here is how it works, step-by-step: 1. Select the protein you want to engineer. Give the protein sequence to four different protein language models which, together, "score" the likely fitness of each amino acid mutation. Get a ranked list of 50-100 protein mutations that these models *predict* might improve function. (This is all done on the computer.) 2. Take the top 15 predicted "hits" and then synthesize the pairwise combination for each of them. For example, if a mutation at amino acid #100 boosted activity, but so did mutations at positions #120, #135, and so on, then you'd make a protein with each of these "double" mutations. If you took the top 15 hits, then this is only 105 total proteins to synthesize. 3. Make and test all the double mutants in the wet-lab. Measure the activity of each double mutant. So, for example, if you wanted to engineer a protein to be "brighter," then you would put each double mutant in a microplate well and measure this directly. (This step basically captures epistatic relationships; it helps the models figure out which mutations are beneficial or damaging.) Feed the single + double mutant activity data to a neural network, called MULTI-evolve. The model extrapolates these data to infer *additional* mutations that might be synergistic, like combinations of 5-7 amino acid swaps. 4. Take the top three predictions for proteins with 5-, 6-, or 7 amino acid mutations, based on the neural network. Synthesize these proteins using a new DNA assembly approach, also reported in this paper, called MULTI-assembly. (The gist is that you anneal together a bunch of short oligos, each carrying one of the mutations, in a tube to reconstruct each of the full genes. This yields correctly built sequences 40-70% of the time.) 5. Finally, express the proteins in cells, measure their activities, and benchmark them against the wild-type protein. The researchers used this method for various proteins. For one protein, called dCasRx (a Crispr protein that targets RNA instead of DNA), they used it to create a variant with 9.8-fold better activity, and they validated this across three different human genes. You can also optimize proteins for two different properties at the same time. The authors used their method, for example, to engineer an antibody targeting CD122 for both binding affinity AND its expression yield in cells. TL;DR This is a new way to speed up directed evolution. Instead of using random mutations to search through a huge amount of "biological space" (remember that a single protein, with just 100 amino acids, has 20^100 possible combinations), these researchers use AI models to navigate this search space for proteins more quickly. Can we use the same approach to make entire gene circuits?

This past week, the folks at Google DeepMind published their latest genomic model, AlphaGenome , and it made quite a splash on LinkedIn (links in comments). So for this week’s AI ∩ Bio post, let’s dig into what it means, and just as importantly, what it doesn’t. At a high level: AlphaGenome learns to predict thousands of experimental “tracks” (quantitative signals like chromatin marks, accessibility, and splicing) directly from DNA sequence, sometimes across megabase-scale contexts. But let’s be clear: high benchmark scores don’t always mean the model has learned biology. They can also reflect memorization of statistical patterns within known assay data. 🔍 What AlphaGenome Achieves Well ➖ Predicts expected assay readouts from sequence alone with strong accuracy. ➖ Learns sequence motifs, long-range dependencies, and multi-assay relationships. ➖ Acts as a statistical prior for genomic activity: essentially, a learned baseline expectation of what’s typically observed across many cell types and conditions. (In plain English: it predicts what an experiment is likely to show, given the sequence, not why it shows it.) ⚙️ What AlphaGenome Doesn’t (Yet) Do ➖ Does not model causal mechanisms of gene regulation. ➖ Generalizes well within training-like data but may falter under new biological conditions or perturbations. ➖ High base-pair precision ≠ deep biological explanation. 📊 Why AlphaGenome Matters AlphaGenome is a step forward for sequence-driven genomics. However, it doesn't explain why biology acts the way it does; it captures how biology usually behaves. The best way to think of it is as a phenomenology engine: a system that expresses what the genome tends to do on average, not the underlying reasons or mechanism. 🧭 Leadership Lens: Lessons for Diagnostics & Biomarker Strategy For biotech and diagnostics leaders, one insight from AlphaGenome is how to think about AI models as decision tools, not oracles. Two quick heuristics worth considering: ▪️ Prediction ≠ Causation: Treat high model scores as triage signals, not mechanistic truths. Use them to prioritize experiments, not replace them. ▪️ Evaluate by Utility, Not Novelty: Validate AI outputs the same way you validate biomarkers. ▶️ Takeaway: The edge in diagnostics and discovery lies in AI-augmented hypothesis generation, not AI-only pipelines. Leadership in this space will belong to those who build trust, verification, and responsible translation into their AI roadmaps. 🌱 Mentorship Lens: Guidance for Early-Career Scientists ▪️ Learn across boundaries. ▪️ Stay humble about prediction. Knowing what a model cannot say is just as valuable as knowing what it can. ▪️ Build interpretability muscles. Don’t just train, interrogate. ▶️ Takeaway: This era belongs to bilingual thinkers (or, as I like to say, amphibians): scientists fluent in both biology and machine learning, grounded in skepticism, and driven by curiosity.

New review out in Trends in Plant Science from our group, tackling one of the most persistent problems in plant breeding: moving complex traits into elite varieties is hard, slow, and expensive, and breeders have long lacked the tools to plan it rigorously. Led by Seema Yaadav, the paper frames trait introgression not as a fixed backcross protocol but as a constrained multi-objective optimisation problem, balancing success probability, elite genome recovery, time, and cost. It maps the full toolkit to tackle this: predictive cross metrics, pan-genomes and introgressiomics libraries, speed breeding, doubled haploids, and recombination engineering. The vision is 'designed diversity'. Pipelines where the genotype is planned computationally before a single cross is made, then refined iteratively as data accumulates. The building blocks are largely here. The challenge now is integrating them into breeder-facing tools that work in practice. Great to see Seema lead this one. Congratulations also to co-authors Meredith McNeil (CSIRO), Peter Dodds (CSIRO), and Ben Hayes (UQ) and thanks to the Grains Research and Development Corporation for supporting our R&D in this space. 📄 https://lnkd.in/gz6tCkAT