# [[Durian Labs MOC]] Durian Labs is a [[TechBio 101|TechBio]] company building Morpheus: an AI-native, physics-driven platform for design and optimisation of small molecule and biologics therapeutics. The core bet: pre-clinical [[Drug Discovery]] is too slow, too expensive, and locked behind computational skills that most wet-lab researchers don't have. Morpheus democratises access. ## The Problem: Three Bottlenecks The drug discovery pipeline moves through Biology → Chemistry → Execution. Each phase has a structural bottleneck. 1. **Understanding disease biology.** Target discovery and validation. Knowing what to hit. ($20-40M over 3-5 years for early discovery through lead optimisation.) 2. **Compound design.** Hit identification, hit-to-lead, lead optimisation. Designing the molecule that hits it. This is where computational chemistry and [[Molecular Modelling x Accelerating Conformer Search]] live. 3. **90% failure rate in the clinic.** Pre-clinical and clinical trials burn $0.3-1.8B over 4-8 years. Most candidates fail. Better computational predictions upstream compress waste downstream. The direction arrow: shift risk and cost left. Spend more compute early to spend less capital late. Every improvement in in-silico prediction quality reduces the number of candidates that fail in vivo. [[Wright's Law]] applies: cumulative simulation cycles should drive down cost-per-viable-candidate over time. The task hand-off problem compounds these bottlenecks. Today's pharma pipeline is a patchwork of disconnected tools from different vendors, each generating different data formats, each requiring different expertise. A single hit expansion workflow touches cheminformatics tools, computational chemistry software, procurement systems, generative AI, retrosynthesis planners, and assay analysis platforms. Individual scientists rarely know how to operate all of them. Every hand-off between tools, teams, and data formats creates downtime. The integration layer is the real unlock. ## What Morpheus Does Four pillars, each attacking a specific access barrier: 1. **No-code agentic AI platform.** Bridges the skills gap. Researchers with non-computational backgrounds can run advanced ML and physics-driven algorithms using natural language. No code required. This is [[Agent Skills as Codified Domain Expertise]] applied to drug design: package the computational chemist's workflow into something a biologist can operate. 2. **Computational heavy lifting.** GPU and CPU orchestration, parallelisation, secure cloud management. The researcher focuses on science, not infrastructure. Same pattern as cloud labs in the [[Digitalisation of Biology]]: abstract the hardware, sell the capability. 3. **Cost-optimised hardware.** Pre-built automated pipelines that cut development cycles by up to 90% and compute costs by 60%+. Makes advanced methods accessible without enterprise budgets. 4. **Curated chemical and biological datasets.** Updated, domain-specific data. Data quality is the bottleneck for model quality. Proprietary, curated datasets become a compounding moat. The long-term vision: scientific superintelligence for therapeutic development. Specialised AI agents for chemistry, biology, and clinical prediction. Not general purpose AI doing drug discovery on the side. Domain-specific, physics-grounded intelligence. This aligns with the broader industry trajectory toward what Zhavoronkov et al. (2026) call "Pharmaceutical Superintelligence" (PSI): a prompt-to-drug pipeline where a plain-language request initiates autonomous target identification, compound design, synthesis, validation, and clinical trial planning. The concept: type "Design a drug for idiopathic pulmonary fibrosis" and the system handles the rest. Durian's Morpheus is building toward this from the computational chemistry layer up. ## Roadmap - Phase 1: [[Small Molecule Drugs]] design and optimisation platform. - Phase 2: [[Biologics]] focus, specifically peptide and antibody therapeutics (parallel development). - Phase 3: Advanced chemical and biological AI agents specialised for end-to-end therapeutic development. ## First Principles Analysis - **[[Consultancy-to-Platform Transition]] risk.** The classic deep tech failure mode. Durian needs to avoid becoming a computational chemistry services shop. The no-code platform is the right move: productise the workflow, not the expertise. Watch deployment hours per customer. If they're not falling with each new user, the platform claim is hollow. - **[[Technical Moat Assessment Framework]] applied.** Individual components (ML models, physics simulations, GPU orchestration) are standard. The potential moat lives in three places: (1) the integration layer, end-to-end pipeline from target to optimised lead without manual engineering, (2) curated proprietary datasets that improve with usage, and (3) the no-code interface that creates switching costs through workflow lock-in. - **[[Bottleneck Business]] logic.** Durian sits at the bottleneck between computational capability and researcher accessibility. The value accrues where the constraint is tightest. Right now, open-source tools exist but require computational skills and hardware. Proprietary tools exist but cost enterprise budgets. Durian occupies the gap. - **Commoditisation risk.** General-purpose AI labs (DeepMind, OpenAI) are building protein folding and molecular design tools. The defence: depth over breadth. Specialised, physics-driven models for specific therapeutic modalities will outperform general models on precision tasks. Same logic as [[domain specific sense-making]]: the generalist gets you 60%, the specialist gets you the last 40% that matters in regulated environments. - **Why LLMs alone are insufficient.** Pattern-recognition LLMs lack deep understanding of biochemical principles underlying structure-dependent properties. Tokenising SMILES or SELFIES molecular representations loses critical information like stereochemical properties. Generative models also struggle to explore chemical space outside their training sets. The solution: hybrid architectures that combine LLM-based planning and orchestration with physics-based simulations (molecular dynamics, quantum mechanical calculations, docking engines). Durian's emphasis on being "physics-driven" alongside AI-native positions it correctly here. Pure AI plays will hit accuracy ceilings that physics-grounded approaches avoid. - **Design-Make-Test-Analyse (DMTA) closed loops.** The winning architecture for AI drug discovery is iterative: generate candidates computationally, synthesise, test in vitro/in vivo, feed results back to refine the model. Each cycle compresses the next. Insilico Medicine's GENTRL model demonstrated this by discovering potent DDR1 kinase inhibitors in 21 days, with synthesis and validation in 27 more. Their AI-designed TNIK inhibitor rentosertib has reached Phase 2a clinical trials for pulmonary fibrosis, the first real proof point that prompt-to-drug can produce clinical-grade candidates. Durian's automated pipelines need to enable these rapid DMTA cycles for their users. ## Direction Arrow of Progress > Wet lab → in-silico simulation → AI-guided design → autonomous AI agents. Each step compresses cycle time and cost. The ultimate state: AI agents that can autonomously design, simulate, and optimise therapeutic candidates, with humans validating and steering rather than executing. [[Digitalisation of Biology]] traces the full arc from capture to simulate to automate. The system-of-systems architecture matters. The approach gaining traction is central AI orchestration coordinating independently operating subsystems (target discovery, molecular design, biological validation) via APIs. Each subsystem is self-contained, optimised for its domain. The central controller handles interfacing, error detection, and data flow. Multi-agent frameworks like DORA, ChemCrow, and Coscientist are early prototypes. Durian's Morpheus needs to become this orchestration layer for pre-clinical computational work. ## Key Links **Drug Discovery Fundamentals:** - [[Drug Discovery]] - [[Drug Types]] - [[Small Molecule Drugs]] - [[Biologics]] - [[Molecular Modelling x Accelerating Conformer Search]] - [[Quantum Chemistry]] - [[Enhancing Drug Discovery Efficiency]] **TechBio Ecosystem:** - [[TechBio MoC]] - [[TechBio 101]] - [[TechBio Myths]] - [[Digitalisation of Biology]] **Business Model and Strategy:** - [[Consultancy-to-Platform Transition]] - [[Technical Moat Assessment Framework]] - [[Bottleneck Business]] - [[Agent Skills as Codified Domain Expertise]] - [[Wright’s Law]] - [[Deployment Velocity]] **Venture and Investment:** - [[Venture Building Manifesto]] - [[First Principles and Mental Models MoC]] - [[domain specific sense-making]] **Landscape and Comparables:** - Insilico Medicine (Chemistry42, PandaOmics, rentosertib Phase 2a) - Recursion, Generate Biomedicines, Absci - ChemCrow, Coscientist, DORA (agentic AI research assistants) - Terra Quantum (TQ Chemistry) per [[Enhancing Drug Discovery Efficiency]] **Reference:** Zhavoronkov, Gennert & Shi (2026). "From Prompt to Drug: Toward Pharmaceutical Superintelligence." ACS Central Science. DOI: 10.1021/acscentsci.5c01473 --- Tags: #deeptech #kp #wip