machine.learning.bio

Research

We ask what machine-learned representations can reveal about biology — and where they fail — then use those answers to build more reliable tools for discovery and protein engineering. By doing so, we develop quantitative and machine-learning methods to understand, evaluate, and engineer biology.

Theme 01

Representation learning for biology

Teaching models to read proteins and genomes as languages.

In ProtTrans, we trained large AI models to read protein sequences like a language, similar to how modern AI reads human text. These models learned fundamental principles about how proteins work without being explicitly taught biology, chemistry or physics. Most remarkably, our best model (ProtT5) outperformed the methods available at the time at predicting protein properties, while being dramatically faster.

We contributed Nucleotide Transformer, a deep learning system that learned multiple genomes from humans and other species. It learned the language of DNA by unmasking millions of genetic sequences. Through this process, it discovered patterns in DNA that control how genes work, where genes begin and end, and how DNA variations might affect human health.

PredictProtein has been making protein analysis accessible to scientists worldwide since 1992, predicting structure and function from an amino acid sequence. We contributed the update that added protein language model predictions — the first significant departure from explicit use of evolutionary information.

Work in this theme

Theme 02

Generative and programmable design

Writing new proteins, atom by atom, and testing them at the bench.

We develop generative methods for atomistic protein and binder design, and study how pretraining, inference-time optimisation, and experimental evaluation can improve specificity and usable hit rates.

The line runs through three models. Proteina generates protein backbones, using hierarchical fold-class conditioning and a scaled transformer to produce designable structures up to 800 residues. La-Proteina makes them fully atomistic, modelling coarse backbone explicitly while capturing sequence and side-chain detail in per-residue latents — which sidesteps the awkward problem of side chains that change length during generation. Proteina-Complexa conditions all of that on a target.

Binder design has generally been cast as either conditional generation or sequence optimisation against a structure predictor — hallucination. We argue that is a false dichotomy. Proteina-Complexa unifies both, pretrains on Teddymer, a large synthetic dataset of binder–target pairs built from domain–domain interactions in predicted monomeric structures, and then optimises at inference time. It extends to small-molecule targets and enzyme design as well.

Designs from this work are being validated experimentally with collaborators at Duke and elsewhere. Those results are not published yet, and we will link them here when they are.

Work in this theme

Theme 03

Search and structure at scale

Making biological search and structure prediction fast enough for whole proteomes.

We tackled a computational bottleneck in protein sequence comparison by redesigning the core algorithms to use modern GPU architectures. With our collaborators we proposed GPU-optimised algorithms for both gapless sequence filtering and gapped alignment using protein profiles, reaching speeds up to 100 trillion cell updates per second. Beyond raw performance, we focused on making these search capabilities work even on low-wattage GPUs, so the speed-up is not reserved for people who own a cluster.

That matters because homology search sits in front of nearly every structure predictor. Making it cheap changes what is affordable downstream.

Which is what the second strand does. We build and analyse large-scale predictions of protein complexes to extend structural biology beyond individual chains — interaction interfaces, complex-dependent folds, and conserved structural patterns across the tree of life. We predicted 31 million candidate homo- and heterodimeric complexes across 4,777 proteomes, including model organisms and organisms of global health concern, and established confidence criteria by analysing experimentally determined structures. That yielded 1.81 million high-confidence predictions.

Two findings from that set are worth stating plainly: the top 1% of structural clusters account for roughly 44% of all complexes, and roughly 8.3% of clusters are conserved across multiple domains of life — pointing to a substantial fraction of ancient, universally retained assemblies. Meanwhile 31.3% of high-confidence predictions extend beyond detectable PDB coverage.

Work in this theme

Theme 04

Benchmarks and honest evaluation

Measuring whether these models actually work where it matters.

We develop benchmarks that test whether fitness-prediction models generalise under the distribution shifts that arise in real protein-engineering campaigns — not merely on random train/test splits.

FLIP established the protocols, but was limited to thermostability, binding, and viral capsid viability. FLIP2 adds seven datasets spanning enzymes, protein–protein interactions, and light-sensitive proteins, together with splits that mirror the situations engineers actually face.

The headline result is uncomfortable and we publish it anyway: across these datasets and splits, simpler models often matched or outperformed fine-tuned protein language models, which challenges the utility of existing transfer-learning techniques. Provenance is recorded for every dataset and all data is redistributed CC-BY 4.0.

Work in this theme

Software and data

Everything below is open, and star counts were checked against GitHub in July 2026.