Roche | LAB IN A LOOP: Transforming Drug Discovery and Development with Data and Artificial Intelligence

On November 25, 2024, Roche hosted a webinar featuring Aviv Regev, Head of Genentech Research and Early Development, who introduced Roche’s 'Lab in a Loop' drug development paradigm. Post-webinar report link: assets.roche.com/f/1763... Table of contents for the post-webinar report: LAB IN A LOOP in Drug Development I. Target Discovery: Building Pathobiological Maps II. Accelerating the Drug Discovery Cycle III. Enhancing Drug Discovery and Optimizing Therapies IV. Clinical Practice Key content summary: LAB IN A LOOP: Transforming Drug Discovery and Development with Data and Artificial Intelligence LAB IN A LOOP in Drug Development: Embedded across every stage—from target discovery to clinical practice I. Target Discovery: Building Pathobiological Maps Biology is multimodal. Multiple distinct modalities and data types represent a multimodal entity. In multimodal LLMs, text controls image content. The model learns associations between two different data modalities (e.g., text and image). NONA: A multimodal LLM for biology. DNA (text) controls another modality (RNA). Application: Prioritizing variants via effect prediction. By controlling DNA sequences, we can predict the impact of genetic variants in specific cell types. Mapping pathobiological landscapes. Inflammatory Bowel Disease (IBD) Identifying candidate drugs for IBD and COPD. AI foundation models employing collaborative agent-based reasoning. Multiple agents can form distinct 'personas'. Reasoning across all foundation models to solve biological problems. An AI orchestration system that coordinates models and data. Agents integrate specialized tools. II. Accelerating the Drug Discovery Cycle High-content perturbation screening within the loop to shorten timelines. Lab in a Loop in oncology. High-content perturbation screening for target ID. Collaboration with Recursion in neuroscience. Lab in a Loop in neuroscience. The first NeuroMap covers genome-wide CRISPR knockouts and incorporates disease-relevant perturbations. map lib: A multimodal model translating RNA/image relationships—no text or DNA sequences included. Predicting compounds that suppress phenotypes using Cell Oracle. III. Enhancing Drug Discovery and Optimizing Therapies Lab in a Loop for molecular drug discovery. Lab in a Loop for small-molecule drug design. Small-molecule AI modeling. Lab in a Loop for macrocyclic peptides. DeepFitness enables discovery of small macrocycles. DeepFitness improves binding affinity and hit diversity. Lab in a Loop for large-molecule design. Example: DyAb achieves enhanced affinity through iterative mutagenesis. De novo antibody generation. Diffusion models generate realistic samples by learning to denoise corrupted data. Design of regulatory elements for gene therapy.

Roche Just Announced Construction of an Autonomous AI Lab, Setting an 80% AI-Decision-Making KPI

On September 28, Roche announced at its London Pharma Day event that it has begun constructing an autonomous AI laboratory—and aims to achieve 'AI independence' in R&D. Aviv Regev, Head of Genentech Research and Early Development, personally unveiled the initiative. The so-called autonomous AI lab—'Lab in a Loop'—operates as follows: AI proposes hypotheses; automated platforms execute experiments; resulting data flows back to retrain the models—repeating this cycle continuously. Roche’s recently launched Boston Innovation Center has one key mission: embedding AI and machine learning into the drug development workflow. This is an organizational transformation built around 'AI making requests, robots doing the work'—not merely installing off-the-shelf software. AI Is No Longer a Demo—It’s a Measurable KPI Regev revealed that from Q4 2025 through Q2 2026, 40% of Roche’s pipeline decisions will have traceable AI or computational contributions. Roche’s internally developed AI tool, Target Nexus, aims to participate in 80% of R&D portfolio decisions by end-2026. This marks the first time a major pharmaceutical company has quantified AI’s contribution as a trackable, reportable metric. In the past, pharma companies discussed AI in terms of where papers were published or which partners they collaborated with; Roche now focuses on how many decisions were made *with AI involvement*. The former reflects marketing narratives; the latter reflects management accountability. This trajectory has been in the making for five years. Setting AI as a KPI reflects confidence built over years of strategic investment. In 2020, computational biologist and single-cell atlas pioneer Aviv Regev joined Genentech from the Broad Institute to lead its entire early research division (gRED), where she introduced the 'Lab in a Loop' closed-loop vision: AI proposes hypotheses, automated platforms run experiments, and data feeds back to retrain models. Every subsequent step has focused on completing components of this loop. In 2021, Roche partnered with digital pathology AI company PathAI; in 2024, that collaboration expanded to include AI-powered companion diagnostic algorithm development. In November 2023, Genentech entered a multi-year strategic partnership with NVIDIA to accelerate internal model training on the DGX Cloud platform using the BioNeMo framework. The harvest begins in 2026.

Roche outlines plans to move towards autonomous AI labs

1 day ago

Roche announced plans to build autonomous AI labs to accelerate drug discovery, aiming to reallocate $2 billion in R&D savings by 2026. The company targets 20 new molecular entities by 2030 with AI driving most research decisions.

reuters

2026-09-28

Roche outlines plans to move towards autonomous AI labs

Roche outlines plans to move towards autonomous AI labs

FAQ