Abstract

Bispecific antibody drugs provide transformative cures, but they are often refractory to modern bioproduction methods. Here, we will apply large-scale data and AI to integrate manufacturing with early bispecific antibody discovery. We will collect high-throughput manufacturability data and train an AI-based drug design algorithm to enhance drug activity, potency, and product quality. First, we will clone large libraries of bispecific antibody variants into manufacture-ready cell lines and growth conditions. Next, each bispecific antibody-producing cell will be captured emulsion microreactor droplets to analyze functional performance in manufacturing-like conditions. High-throughput sequencing will be used to analyze test results en masse, and AI models will be trained to identify critical features of manufacturable molecules. Finally, we will apply trained AI algorithms to generate new manufacturing-ready bispecific designs and evaluate their improvement related to controls. If successful, we will establish a seamless transition between early discovery and large-scale manufacturing for bispecific antibody drug products.

MIT PI
Brandon DeKosky, Associate Professor, Department of Chemical Engineering

PT PI
Paula Alves, CEO of iBET, Professor at NOVA University of Lisbon, Portugal
Antonio Roldao, Head of Cell-based Vaccines Development Lab, Coordinator of Late-stage R&D and Bioproduction Unit, iBET, Portugal
Jose Escandell, Principal Scientist, Animal Cell Technology Unit iBET, Portugal
Patrícia Alves, Coordinator of Analytical Services Unit, iBET, Portugal

This is a two-year grant

Abstract

AI coding tools can generate effective “one-shot” solutions, but they struggle with the ongoing evolution of real software products. One reason is structural: when software is not modular, AI systems cannot easily determine where changes belong. Another is a limitation of current AI tools themselves. They rely on ever-growing contexts that include all generated code and prior interactions, which gradually reduces the effectiveness of the underlying language model while greatly increasing token costs.

This project addresses both challenges. It introduces a new structuring method, called concept design, that brings full modularity to software development. It also develops a strategy that enables fine-grained, targeted context instead of passing the entire project history to the model, and verification techniques to ensure correctness of generated code. Together, these techniques make AI coding more reliable, flexible, and cost-efficient, while also providing broader benefits for other forms of agentic AI.

MIT PI
Daniel Jackson, Professor, Department of Electrical Engineering & Computer Science

PT PI
Alcino Cunha, Research Coordinator, INESC TEC, Associate Professor at University of Minho
 

Abstract

Bayesian inference and data assimilation are central tools for building reliable digital twins of physical, environmental, and engineering systems. Yet their reliability often depends on priors whose implications for simulated dynamics are difficult to anticipate. This project develops prior-predictive calibration for dynamical systems: a framework for translating expert knowledge about system behavior into principled priors for Bayesian dynamical models. Rather than specifying priors only over system parameters, we will constrain the trajectories they generate, including stability, boundedness, oscillation periods, extinction or coexistence regimes, conservation properties, and admissible responses. We propose to optimize and learn priors whose prior predictive distributions satisfy these behavioral constraints while remaining close to an initial scientific prior. The project will develop differentiable calibration methods, surrogateassisted admissible-set learning, and decision-oriented evaluation for posterior inference and data assimilation. The MIT–Portugal team combines Bayesian machine learning, uncertainty quantification, mechanistic modeling, and sequential decision-making, delivering open-source software, benchmarks, and a realistic dynamical-system case study.

MIT PI
Youssef Marzouk, Professor, Department of Aeronautics and Astronautics

PT PI
Eliezer de Souza da Silva, Assistant Professor, Department of Informatics Engineering, University of Coimbra

Additional collaborator
Matthew E. Levine, Research Scientist, Basis Research Institute; Iñigo Urteaga, Ramón y Cajal and Ikerbasque Research Fellow, Machine Learning Group, BCAM