Four core research domains, plus the occasional detour elsewhere in AI and science.
Sequence models, variant-effect prediction, and the work to read non-coding DNA.
An evidence-first review of genomic foundation models across clinical molecular diagnostic workflows. Compares DNABERT-2, Nucleotide Transformer, HyenaDNA, Caduceus, Evo 2, and AlphaGenome against established clinical tools across eight diagnostic tasks and post-calling variant effect prediction.
In its developers' comparisons, SpliceAI2 scored above SpliceAI, Pangolin and AlphaGenome on every benchmark printed, including cryptic splicing in Genomics England blood RNA-seq. Every value is developer-reported, most test variants lie in genes the model was trained on, no uncertainty is printed, and motif scoring, the baseline for both use cases, was not compared.
Eight published somatic SNV and indel caller comparisons, read side by side: rankings change between studies, indel accuracy is lower, and no study covers FFPE, ctDNA or panels.
GPU and DRAGEN pipelines cut a whole-genome run from 17 to 39 hours on CPUs to between about 25 minutes and two and a half hours, and the one study that compared their calls with the CPU pipeline found close agreement. No study reruns a pipeline to test reproducibility, and FP8 gave a genomic foundation model at most 19.3% more throughput.
SpliceAI and Pangolin lead the DNA-based predictors on variants tested in patient RNA, and SpliceAI beats motif scoring on the same data, but no predictor replaces the RNA test. RNA-seq outlier callers recover between none and most known events and flag outliers in nearly every negative sample.
Two H100s on different clouds gave bit-identical Evo 2 scores for 3,893 BRCA1 variants. An L40S running the same container reshuffled two thirds of the ranking. Running the model in FP32 brought the two GPUs back into agreement, at three to six times the compute.
If you can test 100 variants in a splicing reporter, which score should rank them? On 8,297 held-out MFASS variants, a simple trained baseline put 61 splice-disrupting variants in its first 100, and SpliceAI and Pangolin put 63 to 66. The difference was not established. Pangolin ranked the whole list better in exploratory analyses.
In this exploratory study, with annotation, reference genome and variants matched, SpliceAI and Pangolin put 63 to 66 assay-confirmed splice disruptions in their top 100 of the same 8,297 held-out MFASS variants. Masking raised average precision slightly for both tools; no top-100 difference was established.
Genomics MCP is a local MCP server that gives AI agents bounded, provenance-tracked access to real genomic data: discover public files, then read bounded regions from bigWig, BAM, CRAM and VCF. Includes a case study where Claude used only its tools to compare BCL11A enhancer signal across 26 public ENCODE files.
The baseline, SpliceAI and Pangolin found 61, 64 and 65 disruptions in their top 100 MFASS variants. The differences do not establish a reliable winner.
RNA models share an alphabet, not an estimand: choose them by molecular object, output, information inputs, evidence class, and artifact contract.
A reported genomic-model comparison is auditable only when its tokenisation, strand rule, biological target, evaluation split, and artifact terms are explicit.
Protein generators propose different biological objects. A defensible design starts with the assay, records the complete computational stack, and preserves every experimental denominator.
A practical guide to choosing, extracting, adapting and validating protein language-model representations without mistaking a system score for biological generalisation.
A task-first guide to antibody representations, structure prediction, CDR design, humanisation and developability, with the evidence and reproducibility checks that model scores leave out.
Suppose somebody hands you a FASTA file and asks for “the structure.” First decide whether the target is one chain, a protein assembly or a mixed complex containing ligands or nucleic acids.
One pretrained genomic model can look dominant on a leaderboard and lose to a linear baseline on the next task. Two evidence ledgers explain why.
Carbon-3B matches Evo2-7B on sequence recovery, variant-effect prediction, and motif-perturbation discrimination while generating DNA over 150 times faster.
A 1-billion-parameter model conditioned on RNA chemical-probing reactivity folds a held-out transcript to an F1 of 0.987 against an experimentally guided reference.
Google DeepMind's AlphaGenome reads 1 million base pairs of DNA and predicts thousands of regulatory functions at single-nucleotide resolution, beating 25 of 26 specialized models.
AlphaGenome processes 1 million DNA base pairs to predict variant effects across 7,000+ genomic tracks in one second, outperforming specialized models on 25 of 26 VEP benchmarks.
A technical look at AlphaGenome's architecture, its 2D pairwise embeddings for splicing prediction, and what the model means for clinical variant interpretation.
Basecamp Research's EDEN model trains on proprietary environmental metagenomics to design gene-insertion enzymes, antimicrobial peptides, and synthetic microbiomes -- all validated in the wet lab.
A practical guide to selecting genomic foundation models for bioinformatics tasks. Covers ESM-2, DNABERT-2, HyenaDNA, Nucleotide Transformer, scGPT, and Evo with scoped comparisons for DNA, proteins and single-cell analysis, with corrected references and clear distinctions between frozen representations and trained predictors.
How TTT-E2E achieves constant inference latency regardless of context length by treating long context as a learning problem rather than an architecture problem.
How the same transformer architecture powering GPT learned to predict molecular properties by treating chemistry as a language problem
How researchers adapted BERT for molecular property prediction, turning SMILES strings into drug discovery insights
Machine learning models now outperform FDA-approved biomarkers in predicting treatment response, but the best-performing models often resist explanation. Here's how precision oncology is navigating the trade-off between performance and interpretability.
Pedro Domingos proposes that neural networks and symbolic AI are the same mathematical operation - a logical rule can be equivalently written as a tensor equation in Einstein summation notation. If true, we've been building separate tools for problems that share identical structure.
GPT-4's 128K context window? It only uses about 10% effectively. Google's TITANS architecture introduces test-time memory learning that outperforms GPT-4 on long-context tasks with 70x fewer parameters.
Nested Learning: The Illusion of Deep Learning Architectures - A comprehensive guide to the arXiv paper revealing how neural networks learn at multiple timescales through hierarchical optimization.
How gene length, trait specificity, and luck systematically distort which genes we think are important for disease, and what researchers should do about it
Why does it feel like our tools weren't designed by pathologists? Billions poured into AI models that compress whole slide images into tiny vectors, ignoring how pathologists actually examine tissue. The evidence reveals why scaling won't fix this disconnect.
Traditional readability formulas miss the mark. Modern embedding models can capture semantic nuance and syntactic structure, but do they actually predict complexity better?
Single-sequence folding, protein language models, and structure without alignment.
A model can put proteins in a useful order while barely improving its predictions. Six rhodopsin models show why error in nanometres matters alongside rank correlation.
A small CNN cut prediction error by 28% against cheap features on unseen sequence families. For choosing the top 1% of sequences, the two were indistinguishable, and frozen UTR-LM embeddings added nothing.
ESMFold predicts atomic-level structure from a single sequence. No multiple sequence alignment, no database search, no Evoformer churning over homologs.
A technical deep dive into DeepSeek's Engram architecture, which introduces conditional memory as a new axis of sparsity for large language models.
VL-JEPA changes the prediction target and separates visual inference from text decoding. Here is how its objective differs from token prediction and latent diffusion, and what its efficiency results establish.
Recent research shows 1024-layer networks achieve 2x to 50x improvements in goal-conditioned RL. Here's why extreme depth works now, and when you should consider it for your own agents.
Contrastive screening, protein–ligand interaction, and design that survives the wet lab.
The best-supported RNA metagenomic result is 93.6% sensitivity for one UCSF assay, reported by its developers. A new Oxford nanopore validation found 51%. No study compares workflows prospectively against an adjudicated reference.
Seven published comparisons of DNA sequencing workflows for pathogen identification, from mock communities to patient specimens. None is an independent clinical validation of competing workflows.
Structure-based and ligand-based drug design evolved as separate fields solving the same problem. ConGLUDe, a contrastive geometric learning model, unifies both approaches and outperforms specialist methods on realistic benchmarks without requiring pre-defined binding pockets.
Why biological systems offer the ideal training ground for reinforcement learning: automated verification through physics, not human judgment. From protein design with AlphaFold to RNA folding with ViennaRNA, biology provides the verifiable inverse problems that RL needs at scale.
MIT researchers used generative AI to create novel antibiotics from scratch, not finding them, but designing them. The breakthrough matters less for what it creates than for what it proves.
What models encode, where they fail, and how evaluation and oversight hold up.
A recall reversal and two constant-prediction controls show why benchmark comparisons need populations, selection rules and reference scores.
Reinforcement learning works when you can check the answer. A chess engine wins or loses. A code-generation model passes or fails the test suite.
Remaining tags across the archive.
