AI Taskforce Update

Feb 5
Kickoff
Feb 14
Project proposal finalized
Mar 17
Overall system design confirmed
Mar 18
8 project teams established
Jul 27
Deliver ≥ 2 project pilots
Sep
Deliver ≥ 5 project pilots
Nov
Launch ≥ 5 projects into production

Project Overview

Search & Evaluation
Target Research
ADMET Prediction
Molecule Differentiation Analysis
Animal Model Translatability Evaluation
Unmet Medical Needs Identification
Clinical Development
Clinical Study QC & QA
Protocol Deviation Analysis
Project Risk & Issue Identification and Management
Medical Monitoring
Protocol Review
GCTO Operation Platform
Frontier Technology
Virtual Cell
Digital Pathology
Genomics Platform
PBPK Prediction
Competitive Intelligence & Scientific Finding Tracking
Scientific Finding Tracking
PV Literature Search
Disease Deep Dive
Clinical Data Benchmarking
Competitive Intelligence Monitoring
IT Infrastructure Excellence Enablement
GPU Resource for AI Model Training and Inference
Data Lake

Current Project Progress

PILOT
MASSIVE USE
Search & Evaluation
Target Research
SEP
NOV
ADMET Prediction
SEP
NOV
Molecule Differentiation Analysis
SEP
NOV
Animal Model Translatability Evaluation
Q1 2027
Q2 2027
Clinical Development
Clinical Study QC & QA Feasibility
Protocol Deviation Analysis
SEP
NOV
Project Risk & Issue Identification and Management Feasibility
NOV
Q1 2027
Frontier Technology
Virtual Cell
Q1 2027
Q2 2027
Digital Pathology Feasibility
Q1 2027
TBD
Competitive Intelligence & Scientific Finding Tracking
Scientific Finding Tracking
SEP
NOV
Disease Deep Dive
SEP
NOV
Clinical Data Benchmarking
SEP
NOV
Competitive Intelligence Monitoring
SEP
NOV
IT Infrastructure Excellence Enablement

Case Study · AI-Assisted Protocol Deviation Analysis

Status Quo

CRA Input

Physician Review

PD

PD Classification

IPD

IPD Determination

Trend Analysis

AI-Assisted

CRA Input

AI Agent

PD CategoryIPD DeterminationTrend Analysis

Physician Review

May 8
Brainstorm
May 11
Prototype completed
May 14
First prototype communication
Jun 12
First round feedback
Jun 26
First round optimization completed
  • After one round of optimization, IPD determination accuracy increased from 67.5% to 80.5%, and PD category classification accuracy increased from 94.4% to 97.3%.
  • AI pre-screening not only improves physician review efficiency, but also enhances consistency across physicians and generates more comprehensive trend analysis and insights, further improving clinical study quality.
  • Human review-based feedback and manual AI tool tuning during development consume significant time (87% of total time spent), becoming a bottleneck for AI application development. How can AI learn to improve accuracy by itself?

Loop Engineering for Self-Evolving Agents

Systematic design of closed feedback loops that enable AI agents to continuously learn, self-correct, and evolve through autonomous iteration.

Loop Engineering · RLAI AgentPOLICY · πTaskENVIRONMENTInference / ActionTRAJECTORY · τReward / FeedbackPOLICY UPDATEOUTCOME EVALREWARD MODEL
Enhancing Agentic Capability
Prompt Engineering
Shape how the agent reasons through instructions, templates, and chain-of-thought triggers.
Context Engineering
Curate the right evidence, memory, and retrieval context so the agent decides from the best inputs.
Harness Engineering
Add tools, validators, guardrails, and feedback channels that keep the agent aligned and on-task.
Loop Engineering
Close the execution-evaluation-update loop so every task outcome becomes a training signal for the agent.

Multi-agent Platform for Drug R&D AI Operating System

Complex problems are rarely solved by a single agent—they require coordinated multi-agent orchestration - turning fragmented into one coherent workflow.

Target Research
ADMET Prediction
Molecule Diff. Analysis
Animal Model Eval.
Unmet Medical Needs
Clinical Study QC/QA
Protocol Deviation
Project Risk & Issue Mgmt
Medical Monitoring
Protocol Review
GCTO Operation Platform
Guidance QA
Virtual Cell
Digital Pathology
Genomics Platform
PBPK Prediction
Scientific Finding Tracking
PV Literature Search
Disease Deep Dive
Clinical Data Benchmark
Competitive Intelligence
Drug R&D AI OSKERNEL

Drug R&D AI Operating System

Scaling Through Applications × Agentic Capabilities

Expand the Application Layer by building specialized tools and scaling trusted data assets
Upgrade Agentic Intelligence by embedding advanced reasoning and planning capabilities
Accelerate Organizational AI Maturity by developing talent while shipping solutions faster
AI AGENTIC CAPABILITY SCALING
Multi-Agent Framework
Loop Engineering
Harness Engineering
Context Engineering
Chain of Thoughts
Prompt Engineering
Target Research
ADMET Prediction
Molecule Differentiation
Animal Model Translatability
Unmet Medical Needs
Protocol Deviation Analysis
Medical Monitoring
Protocol Review
GCTO Operation Platform
Guidance QA
Virtual Cell
Digital Pathology
Genomics Platform
PBPK Prediction
Clinical Data Benchmarking
Competitive Intelligence
Disease Deep Dive
APPLICATION SCALING (TOOL & DATA)

Opportunities Across R&D Value Chain

Data from the MRL GenAI Functional Taskforce
China
China & HQ
HQ
Unexplored
Search & Evaluation
Clinical Development
Regulatory Affairs
Innovation Focus & Ideation
Search & Evaluation
Data Generation & Validation
Study Planning
Execution
Analysis & Reporting
Regulatory Affairs
External Environment Monitoring
Disease Deep-dive
Target Research
Competitive Intelligence Monitoring
Scientific Finding Tracking
Virtual Cell
Unmet Medical Needs Identification
Asset Benchmarking
Animal Model Translatability Eval.
Molecular Differentiation Analysis
ADMET Prediction
Binding Affinity Prediction
Site Selection
Protocol Drafting
ICF Drafting
Protocol Review
IB Drafting
Project Risk & Issue Management
Clinical Study QC/QA
Protocol Deviation
CSR Drafting
Medical Monitoring
IB Updates
Digital Pathology
Patient Recruitment
Operation Management Platform
AE/SAE Detection
Audit Report
AI for RWE
Clinical Data Benchmarking
AI-assisted Drafting
AI-based Virtual Review

Where are the valuable AI applications in different functions or project process?

Where are the opportunities we believe we are more advanced so that we can talk broadly with HQ?