Danny Reidenbach

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.

A scalable and accurate transformer model unifies diffusion and flow-matching approaches to generate realistic 3D molecules.

While deep learning has transformed structure prediction for ordered proteins, intrinsically disordered proteins remain poorly predicted due to systematic underrepresentation in training data, despite constituting approximately 30% of eukaryotic proteomes. We introduce PeptoneBench, the first benchmark to enable systematic assessment of ensemble generators for both ordered and disordered proteins, integrating diverse experimental observables. Our analysis reveals that existing evaluation metrics exhibit systematic bias toward the structured spectrum of the proteome. Assessment of popular predictors (AlphaFold2, ESMFlow, Boltz2) confirms high accuracy on ordered proteins but shows performance degradation with increasing disorder. We further present PepTron, a flow-matching ensemble generator trained on data augmented with synthetic disordered protein ensembles. On our benchmark PepTron matches BioEmu on disordered regions while maintaining competitive accuracy on ordered protein benchmarks. Our data augmentation approach demonstrates that targeted training strategies can approach the performance of computationally expensive simulation-based methods, establishing a generalizable framework applicable to other protein generative models. All datasets, models, and code are openly available.

De novo 3D molecule generation is a pivotal task in drug discovery. However, many recent geometric generative models struggle to produce high-quality 3D structures, even if they maintain 2D validity and topological stability. To tackle this issue and enhance the learning of effective molecular generation dynamics, we present Megalodon-a family of scalable transformer models. These models are enhanced with basic equivariant layers and trained using a joint continuous and discrete denoising co-design objective. We assess Megalodon's performance on established molecule generation benchmarks and introduce new 3D structure benchmarks that evaluate a model's capability to generate realistic molecular structures, particularly focusing on energetics. We show that Megalodon achieves state-of-the-art results in 3D molecule generation, conditional structure generation, and structure energy benchmarks using diffusion and flow matching. Furthermore, doubling the number of parameters in Megalodon to 40M significantly enhances its performance, generating up to 49x more valid large molecules and achieving energy levels that are 2-10x lower than those of the best prior generative models.

Molecular conformer generation (MCG) is an important task in cheminformatics and drug discovery. The ability to efficiently generate low-energy 3D structures can avoid expensive quantum mechanical simulations, leading to accelerated virtual screenings and enhanced structural exploration. Several generative models have been developed for MCG, but many struggle to consistently produce high-quality conformers for meaningful downstream applications. To address these issues, we introduce CoarsenConf, which coarse-grains molecular graphs based on torsional angles and integrates them into an SE(3)-equivariant hierarchical variational autoencoder. Through equivariant coarse-graining, we aggregate the fine-grained atomic coordinates of subgraphs connected via rotatable bonds, creating a variable-length coarse-grained latent representation. Our model uses a novel aggregated attention mechanism to restore fine-grained coordinates from the coarse-grained latent representation, enabling efficient generation of accurate conformers. Furthermore, we evaluate the chemical and biochemical quality of our generated conformers on multiple downstream applications, including property prediction and large-scale oracle-based protein docking. Overall, CoarsenConf generates more accurate conformer ensembles compared to prior generative models.

Molecular conformer generation (MCG) is an important task in cheminformatics and drug discovery. The ability to efficiently generate low-energy 3D structures can avoid expensive quantum mechanical simulations, leading to accelerated virtual screenings and enhanced structural exploration. Several generative models have been developed for MCG, but many struggle to consistently produce high-quality conformers. To address these issues, we introduce CoarsenConf, which coarse-grains molecular graphs based on torsional angles and integrates them into an SE(3)-equivariant hierarchical variational autoencoder. Through equivariant coarse-graining, we aggregate the fine-grained atomic coordinates of subgraphs connected via rotatable bonds, creating a variable-length coarse-grained latent representation. Our model uses a novel aggregated attention mechanism to restore fine-grained coordinates from the coarse-grained latent representation, enabling efficient generation of accurate conformers. Furthermore, we evaluate the chemical and biochemical quality of our generated conformers on multiple downstream applications, including property prediction and oracle-based protein docking. Overall, CoarsenConf generates more accurate conformer ensembles compared to prior generative models.

Motivation: Many computational methods aim to identify genetic variants associated with diseases and complex traits. Due to the absence of ground truth data, simulated genotype and phenotype data is needed to benchmark these methods. However, phenotypes are frequently simulated as an additive function of randomly selected variants, neglecting biological complexity such as non-random occurrence of causal SNPs, epistatic effects, heritability and dominance. Including such features would improve benchmarking studies and accelerate the development of methods for genetic analysis. Results: Here, we describe GEPSi ( G WAS E pistatic P henotype Si mulator), a user-friendly python package to simulate phenotype data based on user-supplied genotype data for a population. GEPSi incorporates diverse biological parameters such as heritability, dominance, population stratification and epistatic interactions between SNPs. We demonstrate the use of this package to compare machine learning methods for GWAS analysis. Availability and Implementation: GEPSi is freely available under an Apache 2.0 license, and can be downloaded from https://github.com/clara-parabricks/GEPSi . Supplementary information: Supplementary data are available online.

A bstract: Single-cell ATAC sequencing (scATAC-seq) is a powerful and increasingly popular technique to explore the regulatory landscape of heterogeneous cellular populations. However, the high noise levels, degree of sparsity, and scale of the generated data make its analysis challenging. Here we present PeakVI, a probabilistic framework that leverages deep neural networks to analyze scATAC-seq data. PeakVI fits an informative latent space that preserves biological heterogeneity while correcting batch effects and accounting for technical effects such as library size and region-specific biases. Additionally, PeakVI provides a technique for identifying differential accessibility at a single region resolution, which can be used for cell-type annotation as well as identification of key cis-regulatory elements. We use public datasets to demonstrate that PeakVI is scalable, stable, robust to low-quality data, and outperforms current analysis methods on a range of critical analysis tasks. PeakVI is publicly available and implemented in the scvi-tools framework: https://docs.scvi-tools.org/ .