Kieran Didi

The catalytic and other sophisticated functions of biomolecules arise from the placement of many atoms in three-dimensional space. Here, we describe a generative artificial intelligence method for biopolymers, RFdiffusion3 (RFD3), which employs free atom diffusion to satisfy complex sets of atomic constraints more rapidly and accurately than previous methods across several challenges from enzyme design to RNA structure design. We demonstrate experimentally the power of the RFD3 pipeline for enzyme design by designing de novo phosphotriesterases with a complex active site architecture incorporating a carbamylated lysine coordinating two metal ions and zinc proteases specific for the transthyretin protein involved in amyloid diseases; in both cases we obtain active catalysts in zero shot 96 construct design campaigns.

De novo protein design has advanced rapidly, yet designing binders to polar, solvent-exposed epitopes and small, flexible ligands remains challenging. Such hydrated surfaces and flexible molecules, including carbohydrates, provide few of the hydrophobic contacts favoured by current methods and have largely resisted de novo binders. To address this challenge, here we introduce latent generative search for binder design, a novel framework that uses reward-guided search at inference time to steer the Proteína-Complexa generative model. The model codesigns sequence and structure — generating them together in a continuous latent space — and thereby removes the inverse-folding step on which current methods rely. In a screen of more than one million designs by multiplexed phage display, latent generative search produced more validated binders than every other method tested, its codesigned sequences surpassing post hoc redesign. It delivered high-affinity binders across therapeutic receptors, a viral attachment protein and intracellular signalling targets. Our approach also accessed previously untapped biology, generating the first de novo proteins that bind a free carbohydrate, including one that discriminates between blood-group antigens — a polar, flexible target class beyond the reach of current design methods.

Structure prediction has reached RNA, where generating deep alignments is now the bottleneck. We developed Riboseek, a search and alignment tool for RNA and DNA that represents sequences as overlapping di-mers. Riboseek is more sensitive for homology detection than BLASTN and nhmmer, and generates alignments 250- and 376-fold faster than nhmmer and rMSA. With structure-aware realignment, it approaches rMSA’s base-pair recovery, while remaining 29-fold faster. We provide a web server and 1.7 million precomputed alignments for putatively structured RNAs. Riboseek is freely available at https://github.com/steineggerlab/riboseek .

We introduce accelerations for deep learning inference with OpenFold and TensorRT that, combined with MMseqs2-GPU on an x86 system with one NVIDIA RTX PRO 6000 Blackwell Server Edition GPU, reach up to 131× faster inference compared to AlphaFold2. ARM-optimizations enable homology search on power efficient DGX Spark, and allow search beyond available GPU RAM on NVIDIA Grace Hopper Superchip at comparable overall folding speed. These accelerations enable high-throughput protein structure inference at no accuracy cost using familiar tools.

Machine learning methods that predict protein fitness from sequence remain sensitive to changes in data distributions, limiting generalization across common conditions encountered in protein engineering. Practically, protein engineers are thus left wondering about the effective utility of ML tools. The FLIP benchmark established protocols for testing generalization under some domain shifts, but it was limited to measurements of thermostability, binding, and viral capsid viability. We introduce FLIP2, a protein fitness benchmark spanning seven new datasets, including enzymes, protein-protein interactions, and light-sensitive proteins, as well as splits that measure generalization relevant to real-world protein engineering campaigns. Evaluating a suite of benchmark models across these datasets and splits reveals that simpler models often matched or outperformed fine-tuned protein language models on FLIP2, challenging the utility of existing transfer learning techniques. Provenance for all datasets has been recorded and we redistribute all data CC-BY 4.0 to facilitate continued progress.

Deep learning has accelerated protein design, but most existing methods are restricted to generating protein backbone coordinates and often neglect interactions with other biomolecules. We present RFdiffusion3 (RFD3), a diffusion model that generates protein structures in the context of ligands, nucleic acids and other non-protein constellations of atoms. Because all polymer atoms are modeled explicitly, conditioning the model on complex sets of atom-level constraints for enzyme design and other challenges is both simpler and more effective than previous approaches. RFD3 achieves improved performance compared to prior approaches on a range of in silico benchmarks with one tenth the computational cost. Finally, we demonstrate the broad applicability of RFD3 by designing and experimentally characterizing DNA binding proteins and cysteine hydrolases. The ability to rapidly generate protein structures guided by complex sets of atom-level constraints in the context of arbitrary non-protein atoms should further expand the range of functions attainable through protein design.

Structure-based drug design (SBDD) aims to design small-molecule ligands that bind with high affinity and specificity to pre-determined protein targets. Generative SBDD methods leverage structural data of drugs with their protein targets to propose new drug candidates. However, most existing methods focus exclusively on bottom-up de novo design of compounds or tackle other drug development challenges with task-specific models. The latter requires curation of suitable datasets, careful engineering of the models and retraining from scratch for each task. Here we show how a single pretrained diffusion model can be applied to a broader range of problems, such as off-the-shelf property optimization, explicit negative design and partial molecular design with inpainting. We formulate SBDD as a three-dimensional conditional generation problem and present DiffSBDD, an SE(3)-equivariant diffusion model that generates novel ligands conditioned on protein pockets. Furthermore, we show how additional constraints can be used to improve the generated drug candidates according to a variety of computational metrics.

Rapidly growing protein databases demand faster sensitive sequence similarity detection. We present GPU-accelerated search utilizing intra-query parallelization delivering 6x faster single-protein searches compared to state-of-the-art CPU methods on 2×64 cores—speeds previously requiring large protein batches. It is most cost effective, including in large-batches at 0.45x MMseqs2-CPU speed (8 GPUs delivering 2.4x). It accelerates ColabFold structure prediction 31.8x compared to AlphaFold2 and Foldseek search 4-27x. MMseqs2-GPU is open-source at mmseqs.com .