Japan’s Elix Launches AI Drug Discovery Platform Trained on Confidential Pharma Data via Federated Learning

Tokyo-based Elix (founded 2016) has launched Elix Discovery, an AI drug discovery platform powered by federated learning models trained on proprietary compound data from 16 pharmaceutical companies. The platform is the first commercial deployment to result from Japan’s AMED-backed DAIIA initiative (2020–2025), which brought together 17 pharma companies and institutions including RIKEN, Kyoto University, and Nagoya University.

The core technical framework—kMoL—was developed with Kyoto University’s Department of Biomedical Data Intelligence. It enables federated training: AI models learn from decentralized datasets stored behind firewalls, preserving confidentiality while allowing collective model development. According to Elix, this has resulted in predictive and generative models for tasks such as target prediction, ADMET profiling, and compound generation. The platform’s models were trained on over 1 million compounds and 10 million data points, contributed by DAIIA members.

Source: Elix

This launch is reportedly the first time such models have been commercialized at this scale using federated learning in pharma. Initial adoption is limited to companies that participated in DAIIA, though Elix and LINC plan to open access to additional firms to expand the training pool.

Platform Key Features

Elix Discovery is built for medicinal chemists, with an emphasis on usability without AI expertise. It supports interactive compound design, visualization of chemical space, and score-based molecular optimization. The system incorporates both predictive and generative modeling capabilities.

The predictive layer includes models for activity, ADMET, and physicochemical properties, employing both traditional ML and deep learning. Model selection and training are automated. The generative component supports multiple workflows (de novo molecule generation, lead optimization, linker design) based on ligand- or structure-based principles. It also integrates physics-based docking simulations to complement data-driven approaches, especially where experimental data is sparse.

Use Cases and Empirical Validation

Elix has partnered with over 20 pharmaceutical companies, including Astellas, Shionogi, Eisai, and two global top-30 firms. In a collaboration with PRISM BioLab focused on protein–protein interactions (PPIs), Elix designed six candidate molecules using internal models. Two of these were experimentally confirmed as hits—one validated through two separate assays—representing a novel chemical class.

The workflow emphasized iterative refinement: chemist feedback informed reinforcement learning reward parameters before synthesis. This process deviates from more static generation pipelines and underscores Elix’s hybrid AI-human loop. The generative model used in the project (SmilesFormer) was developed in-house. Multi-parameter optimization factored in synthetic feasibility and pharmacokinetic properties.

In another collaboration with Kaken Pharmaceutical, Elix’s predictive models reportedly outperformed commercial alternatives in ADMET prediction. The models were used to improve bioavailability by a factor of four. Separately, retrosynthesis modeling (Elix Synthesize) has been applied in joint work with Shionogi, using proprietary reaction datasets.

The company claims a differentiating factor is its openness around methodology and modeling details. Additionally, Elix provides dedicated consulting to support onboarding and model adaptation, particularly for chemists without machine learning backgrounds.

Team and Research Track Record

Elix’s team includes researchers from 10 countries, with 78% holding PhDs. Its personnel combine AI expertise with domain knowledge in chemistry, biology, and physics. The company has published research at major AI conferences including ICLR and NeurIPS, and received presentation awards at the CBI Society’s 2022 and 2023 events.

Topics: AI & Digital