Research
At the center of an interconnected network of innovators, Pioneering Intelligence generates capabilities that compound, not just within the ecosystem, but across the field. Our research advances the models, methods, and infrastructure that move AI forward.
Designing proteins that bind to specific targets is essential for developing new therapeutics, but current AI methods often produce limited diversity in their outputs. This work introduces a sampling approach that combines protein structure prediction with sequence generation to explore a broader range of viable designs. The method produces five times more designable protein structures with two to three times greater structural diversity, expanding the pool of candidates for drug discovery.
Modern AI models for predicting protein and RNA structures rely on a computationally expensive attention mechanism called Invariant Point Attention. This work reimplements that core operation to run faster and use less memory, enabling researchers to work with larger biomolecular complexes on standard hardware. The optimized code is open-source and can accelerate structure prediction workflows across the field.
AI models that predict how proteins fold and interact have achieved remarkable accuracy, but their internal reasoning remains opaque. This work applies interpretability techniques to uncover what these models learn about protein-protein interactions, revealing features that correspond to known biological phenomena like binding interfaces and structural contacts. Understanding what the models have learned can help researchers build more reliable tools for drug design.
Inverse folding鈥攄esigning an amino acid sequence that folds into a target 3D structure鈥攊s a key step in protein engineering. This work combines diffusion models with reinforcement learning to generate sequences that are more likely to fold correctly while maintaining diversity. The approach achieves 29% foldable diversity compared to 23% from prior methods, giving researchers more viable candidates to test experimentally
Large language models often settle on a single solution path too quickly, missing alternative approaches that might be better. This work introduces a reasoning framework that explicitly explores multiple options before committing to an answer. On data science and therapeutic chemistry tasks鈥攄omains where considering diverse hypotheses matters鈥攖he method improves performance by 37-69% over standard prompting approaches.
AI models for biology increasingly need to represent molecules at atomic detail, but encoding every atom in large complexes is computationally prohibitive. This work introduces a method to compress all-atom biomolecular structures into compact tokens that can be efficiently processed by modern architectures. The approach scales to structures with approximately 100,000 atoms while maintaining sub-angstrom reconstruction accuracy, enabling new applications in molecular modeling.