Molecular AI by Empirical Bio

Frontier AI driven biomolecule design for biotech firms

Tools that predict and optimise biomolecule structures

Developed In Association With

01

AI Research

Building world models at molecular scale

The AI that generates code and images wasn't built for drug discovery. It’s a tailwind as we create and scale the data, architecture, and models for our domain.
02

AI Platform

GEMS: The AI operating system for molecular design

Our researchers run their whole design workflow on the AI platform they shape, powered by the industry's most advanced models.
03

Drug Discovery

From virtual to validation

Our platform is solving the hardest problems in drug discovery to change the landscape of medicine. It powers our own programs and our major pharma partners'. Every cycle on every program leads to new data and AI model capabilities.

AI Research

Building world models at molecular scale

AI research is at our origin. It's how we turn drug discovery into an engineering problem.
01

Next-Gen Foundation Models

Our focus is on scaling AI at the intersection of deep learning and physics.

Frontier AI models don't transfer cleanly to drug discovery: training data is scarce and text is insufficient to describe 3D matter & interactions. We build original models that understand the physics of molecules and augment them with synthetic data. We're the first to demonstrate LLM-like scaling laws with this approach.
02

Core Research Areas

Inventing new architectures for science

Our research is organized around solving fundamental problems in machine learning that are critical for drug discovery. We pioneer novel deep learning architectures and methods with active research and development.

Generative Diffusion Models

Empirical developed Pearl, our leading foundation model for 3D biomolecular structure prediction. Pearl uses a diffusion-based architecture with equivariant neural networks, trained on physics-based synthetic data, to achieve unprecedented precision. Empirical and NVIDIA jointly announced Pearl at GTC 2025.

Pearl

Reinforcement Learning for Goal-Directed Generation

We use advanced reinforcement learning techniques, grounded in physics-based reward functions, to steer our generative models. This allows us to solve complex, multi-objective optimization problems that are intractable with other methods.

Multi-Modal Learning & Prediction

We build multi-modal foundation models that integrate diverse data sources—from 3D structures to bioassay results—to predict properties like potency and selectivity by combining learned structural representations with experimental context.

Geometric & 3D Deep Learning

Our research advances state-of-the-art equivariant networks that inherently understand the physics of molecular systems, ensuring our models respect rotational and translational symmetries.

Agentic Orchestration

We build autonomous workflows to amplify productivity across our organization. Agents can deploy Empirical foundation models and surface top ranked molecules considering multi-parameter optimization objectives.

Generative Diffusion Models
Reinforcement Learning for Goal-Directed Generation
Multi-Modal Learning & Prediction
Geometric & 3D Deep Learning
Geometric & 3D Deep Learning
01

the frontier of Molecular AI

Lorem ipsum dolor sit amet consectetur.

Original research
Engagement with the global research community

Catalyst for Fundamental Research

The virtuous cycle of research and impact

Pearl
Our work is deeply rooted in real world application, creating a powerful feedback loop that fuels our research agenda.
By deploying our most advanced models on active drug discovery programs, we pressure test their current limits and uncover the next set of fundamental research questions. This problem-first approach ensures our fundamental ML research balances immediate impact with solving challenges beyond the state of the art.

AI Platform

GEMS: Empirical Exploration of Molecular Space

The AI operating system for drug discovery, powered by SOTA models

GEMS integrates foundation models, agents, and the exploratory tooling chemists need to generate and triage solutions to challenging chemistry.
01

GEMS: the agentic end-to-end solution

Drug hunters + agents run their full design workflows 24/7 on GEMS

Chemists use GEMS across the full design cycle: generating candidates, predicting their properties, interrogating the predictions, and deciding what to synthesize. Agents orchestrate the routine steps and surface trade-offs so that chemists can focus on design decisions.
Generate
GEMS prompt to test 10 new scaffolds with resulting molecule variants showing empty prediction fields awaiting scoring
GEMS molecule variants populated with potency, ADME, and physicochemical predictions after agent scoring
Predict
GEMS 3D protein viewer alongside molecule workspace, letting chemists compare prediction values while inspecting structures
Interrogate
GEMS agent conversation showing a chemist's prompt to upload a selected molecule for synthesis
Decide
02

Platform Intelligence

GEMS is powered by our industry-leading models for small and medium-size molecule drug discovery

Agents are only as useful as the models they orchestrate. GEMS models are comprehensive and state-of-the-art. GEMS models cover the breadth of predictions required for small and medium-size molecule drug discovery. It’s AI designed from the ground up by elite ML researchers, trained on proprietary (and public) data, and inspired by drug hunters to be maximally useful.
Pearl

Pearl: our industry-leading foundation model for 3D structure prediction

Pearl predicts the 3D structures of protein-ligand complexes at the <1Å RMSD accuracy drug discovery demands. Pearl is trained on physics-based synthetic data proprietary to Empirical, and can be fine-tuned with program-specific data to improve accuracy on individual targets. Chemists can condition Pearl on what they know about their target, steering predictions toward the structures that matter for their program.

Controllable molecular generation

GEMS proposes novel, drug-like, diverse, and synthesizable molecular ideas conditioned on what the chemist is trying to achieve, including ADME, structural, and program-specific constraints. Chemists drive the design strategy, and the platform surfaces candidates worth pursuing.

Potency and selectivity prediction

GEMS integrates structure-based deep learning methods with physical simulation, including molecular dynamics and quantum chemistry, to predict potency and selectivity. This enables Empirical to find drug candidates for challenging targets that lack on-target training data.

ADME property prediction

Using multitask ML models, GEMS predicts 30+ key ADME properties, including solubility, permeability, metabolic stability, and many others. Chemists see signals for drug-likeness on every candidate before deciding what to make.

Controllable molecular generation

Our language models propose novel, drug-like and diverse molecular ideas conditioned on what the chemist is trying to achieve, including ADME, structural, and program-specific constraints. Chemists drive the design strategy, and the platform surfaces candidates worth pursuing.

Potency and selectivity prediction

GEMS integrates structure-based deep learning methods with physical simulation, including molecular dynamics and quantum chemistry, to predict potency and selectivity. This enables Empirical to find drug candidates for challenging targets that lack on-target training data.

ADME prediction

GEMS predicts 30+ key ADME properties using multitask ML models. Chemists see signals for drug-likeness on every candidate before deciding what to make.

03

Platform flywheel

GEMS is actively shaped by the drug hunters across the industry who use it every day.

On collaboration projects, the same chemists and engineers who shape GEMS are deployed alongside partner teams, helping them to realize the value of their data on a platform proven on real drug programs.
The data they generate improves model performance; their needs and judgment shape the tooling itself.

Our Partners & Pipeline

AI-enabled small & medium-size molecule drugs

We deploy our AI in our labs and with large pharma partners to develop highly potent and selective drugs, for chemically complex targets, to treat severe diseases. Wet labs close the loop for model training, prediction, and validation.
01

Our Partners

We partner with industry leaders to amplify the impact of our platform for drug discovery

By combining GEMS with our partners' R&D prowess, we aim to further accelerate impact for patients.

Incyte

Empirical partnered with Incyte to leverage GEMS for Incyte-selected targets. Across the initial 2025 collaboration & expansion in 2026, Empirical has received a total of $150 million in upfront consideration, including a $40M equity investment. The expanded deal in 2026 also involves Incyte sharing significant experimental data for use in training GEMS.
Multi-target collaboration
Started 2025
Expanded 2026

Gilead

In 2024, Empirical partnered with Gilead to deploy GEMS against multiple hard-to-drug targets. Empirical received $35 million in upfront payment to collaborate on 3 initial targets.
Multi-target collaboration
Started 2024
02

AI + Wet Lab Flywheel

GEMS AI drives molecular design

wet lab experimental ground truth

stronger AI

Empirical has a team of forward deployed engineers and drug hunters who work in unison on our partnerships. These same researchers stress test and strengthen GEMS and its predictions on every partnered and internal program. The combination of expertise, technology, and data together sustain a flywheel that is solving the most critical problems in our field.
03

ONCOLOGY PIPELINE

Better options where current therapies fall short.

Using GEMS to generate and optimize first- or best-in-class small molecules for difficult biological targets, which offer potential to advance the treatment of difficult cancers.
Best-in-class pan-mutant PIK3CA
PIK3CA is a common oncogenic driver in breast and colorectal cancers. Empirical is approaching development candidate nomination for highly potent and selective pan-mutant allosteric inhibitors of PIK3CA.
Breast Cancer
Colorectal Cancer
First-in-class cell-death promoter
Empirical is generating small molecules to inhibit an important anti-apoptotic regulator of the extrinsic cell death pathway, which could prevent evasion of apoptosis in cancer cells with this dependency.
04

IMMUNOLOGY PIPELINE

Pursuing autoimmune targets that small molecules haven't reached.

Using GEMS to generate and optimize first-in-class small molecules for well-validated targets, as well as novel immunology targets.

Validated immunology

First-in-class small molecule
Many targets for autoimmune disorders have proven efficacy from approved or clinical stage biologics, but have yet to be drugged with small molecules. Empirical has multiple programs generating small molecules to inhibit such well-validated targets, to offer oral treatment options to patients in need.

Novel immunology

First-in-class autoimmune disorder
Empirical is developing small molecule inhibitors targeting a protein central to several inflammatory signaling pathways, aiming to meet the unmet needs of patients with autoimmune diseases
First-in-class small molecule
Many targets for autoimmune disorders have proven efficacy from approved or clinical stage biologics, but have yet to be drugged with small molecules. Empirical has multiple programs generating small molecules to inhibit such well-validated targets, offering new oral treatment options to patients in need.
First-in-class autoimmune disorder
Empirical is developing small molecule inhibitors targeting proteins central to several inflammatory signaling pathways, aiming to address the unmet needs of patients with autoimmune diseases.