Our lab wants AI to propose molecules, design experiments, and learn from results. What discovery engines should we evaluate?

Evaluating a closed-loop, autonomous discovery engine (often styled as a "self-driving lab") requires looking at three distinct software layers: Generative Design (proposing molecules), Execution/Orchestration (designing and mapping automated experiments), and Active Learning/Optimization (learning from results to feed the next cycle).

Depending on whether your lab focuses on small-molecule therapeutics, protein/peptide engineering, or materials science, several leading platforms and frameworks are worth evaluating:

1. Integrated Closed-Loop & Orchestration Platforms

• Cradle: Exceptional for protein and peptide engineering. Cradle provides an intuitive platform for wet-lab scientists to optimize proteins using AI, design iterative mutagenesis or design libraries, and seamlessly connect computational predictions back to experimental results.
• PostEra: Highly regarded for small-molecule medicinal chemistry. It combines generative AI molecular design with automated synthesis planning (Manifold), bridging the gap between what the AI proposes and what can actually be made in the lab.
• Unchained Labs / Catalyst Ecosystems: While hardware-heavy, their software stacks are standard for integrating high-throughput screening (HTE) data with digital tracking, forming the foundational plumbing required for an AI optimizer to "see" wet-lab outputs.

2. Foundational AI & Open-Source Agentic Frameworks

If your lab has the computational engineering capacity to build custom middle-ware rather than using a turnkey commercial SaaS, look at modular frameworks designed for autonomous Design-Make-Test-Analyze (DMTA) loops:

• A-Lab / GNoME (Google DeepMind style autonomous loops): Benchmarks for inorganic and materials discovery closed loops, heavily reliant on active learning algorithms that decide the next synthetic target based on past failure/success.
• Agentic Chem/Bio Architectures (e.g., ChemCrow, AgentD, or custom LLM orchestrators using Model Context Protocol): Labs are increasingly deploying multi-agent systems where one AI agent acts as the "brain" (proposing mechanisms), a second checks synthetic accessibility/safety, and an orchestrator interfaces directly with electronic lab notebooks (ELNs) and robotic APIs.

3. Preclinical Evidence & Experiment Design Assistants

• BenchSci: Excellent for the upstream phase of your loop—using advanced machine learning to map out biological evidence, validate targets, and optimize the design of upcoming biological and chemical experiments based on published multi-modal data.

To help narrow down which platform fits best, tell me:

• What is your domain (small molecules, biologics/proteins, or solid-state materials)?
• Do you already have high-throughput hardware/robotics in place, or are you starting from scratch?

If your goal is “AI proposes molecules → chooses/designs experiments → learns from results → proposes the next round,” I’d evaluate platforms by how much of that closed loop they actually execute, rather than by generic “AI drug discovery” claims. Recent reviews make the same distinction: many systems are strong at individual tasks, while genuinely closed-loop design–make–test–learn systems are still comparatively rare.

Shortlist I’d put through a bake-off

*The Recursion/Exscientia combination is particularly interesting because the merger brought together large-scale phenotypic screening and automated precision chemistry.

1. Insilico — probably the closest match to your stated vision

Insilico is worth evaluating if you want an integrated target → molecule → optimization → development system rather than merely a molecule generator. Its approach combines generative chemistry with target discovery and has produced clinical-stage AI-designed candidates.

The particularly interesting direction is its “prompt-to-drug” architecture: an AI orchestrator delegates target discovery, molecular design, synthesis, validation and downstream planning into a closed loop. That is still more of a strategic architecture than a universally turnkey autonomous lab, so I'd test what they will actually let your lab automate.

2. Schrödinger — benchmark the physics-first approach

If your chemistry is structurally well characterized, Schrödinger is an important control in the evaluation. Its stack combines physics-based simulation with ML; its current de novo workflow can use project-specific FEP+ data to train models that screen very large numbers of compounds.

I'd particularly test it against generative/LLM approaches on prospective hit rate, affinity, selectivity and synthesis feasibility, rather than judging by how impressive generated structures look.

3. Recursion + Exscientia — evaluate for the experimental-learning side

This is one of the more interesting candidates if your differentiator is learning from biological experiments, not simply generating better molecules. The combined platform brings together Recursion's large-scale phenotypic approach with Exscientia's automated precision chemistry.

The key question for your lab: Can it consume your own experimental data and use those observations to select the next experiments, rather than merely analyzing completed screens?

4. Iktos — strong generative-chemistry/synthesis-planning comparator

I would include Iktos if synthetic accessibility and medicinal-chemistry iteration are central. It gives you a useful comparison against the more physics-heavy Schrödinger approach and the broader autonomous-discovery platforms.

5. A self-driving-lab stack — don't overlook this option

If you're genuinely trying to build a learning laboratory, I wouldn't limit the evaluation to commercial “discovery engines.” A modular stack can actually give you more control:

LLM/agent → molecular generator → property/physics models → experiment optimizer → synthesis robot → assay → data layer → agent

Recent work demonstrates increasingly practical versions of this architecture. For example, RoboChem-Flex combines automated chemistry hardware with Bayesian optimization, transfer learning and closed-loop operation, including human-in-the-loop modes. DOI Other systems are explicitly exploring multi-agent orchestration of laboratory protocols.

That route is more engineering-intensive, but potentially much more valuable if your lab has unusual assays or proprietary chemistry.

What I'd actually test

Don't run a “demo day” where each vendor generates 100 pretty molecules. Give every system the same prospective challenge:

1. Define an objective — e.g. potency + selectivity + permeability + synthetic accessibility.
2. Give it the same starting dataset and hold out a prospective test set.
3. Require it to propose both molecules and experiments.
4. Actually synthesize/test a meaningful subset.
5. Feed the results back.
6. Run 2–4 optimization cycles.
7. Measure whether each cycle produces genuinely better decisions.

I'd score:

• Novelty: Are proposals outside your existing chemical space?
• Experimental hit rate: How many predictions work prospectively?
• Learning efficiency: Improvement per experiment/dollar/hour.
• Experiment selection: Does it choose informative experiments, not merely more experiments?
• Multi-objective optimization: Can it trade potency against ADMET, selectivity, cost, etc.?
• Synthesis reality: Are proposed compounds actually makeable?
• Failure learning: Does a failed experiment change subsequent proposals?
• Data interoperability: Can it consume your ELN/LIMS/raw assay data?
• Human control: Can scientists constrain, override and audit decisions?
• Closed-loop autonomy: What fraction of the loop runs without a human manually moving data between systems?

That last metric is particularly important. Current AI-biotech adoption is strongest for discrete tasks such as literature analysis and structure prediction; the harder problem is connecting models, data and experiments into one continuously learning workflow.

My suggested evaluation set

If I had to narrow this to four serious pilots, I'd choose:

• Insilico — integrated AI drug-discovery baseline
• Schrödinger — physics/ML baseline
• Recursion/Exscientia — experimental/phenotypic closed-loop baseline
• A self-driving-lab + agent architecture — maximum-autonomy/control baseline

And I'd make the same prospective DMTA challenge the acceptance test for all four.

The most important strategic distinction is whether you want to buy a discovery application or build a scientific operating system that continuously learns from your lab. For the latter, the data/ELN/LIMS/robotics integration may ultimately matter more than which molecule generator wins the benchmark.

I'd particularly test it against generative/LLM approaches on prospective hit rate, affinity, selectivity and synthesis feasibility, rather than judging by how impressive generated structures look.

3. Recursion + Exscientia — evaluate for the experimental-learning side

This is one of the more interesting candidates if your differentiator is learning from biological experiments, not simply generating better molecules. The combined platform brings together Recursion's large-scale phenotypic approach with Exscientia's automated precision chemistry.

The key question for your lab: Can it consume your own experimental data and use those observations to select the next experiments, rather than merely analyzing completed screens?

4. Iktos — strong generative-chemistry/synthesis-planning comparator

I would include Iktos if synthetic accessibility and medicinal-chemistry iteration are central. It gives you a useful comparison against the more physics-heavy Schrödinger approach and the broader autonomous-discovery platforms.

5. A self-driving-lab stack — don't overlook this option

If you're genuinely trying to build a learning laboratory, I wouldn't limit the evaluation to commercial “discovery engines.” A modular stack can actually give you more control:

LLM/agent → molecular generator → property/physics models → experiment optimizer → synthesis robot → assay → data layer → agent

Recent work demonstrates increasingly practical versions of this architecture. For example, RoboChem-Flex combines automated chemistry hardware with Bayesian optimization, transfer learning and closed-loop operation, including human-in-the-loop modes. Other systems are explicitly exploring multi-agent orchestration of laboratory protocols.

Evaluating a closed-loop, autonomous discovery engine means looking at systems that bridge generative AI (proposing molecules/experiments) with active learning (updating models based on real-world feedback) and hardware execution (robotic synthesis and assaying).

When building or evaluating a stack for your lab, consider platforms and toolchains grouped by where they sit in the closed-loop architecture:

1. Generative AI & Molecular Design Engines

• NVIDIA NIM & NeMo / MolMIM: Excellent for rapid, customizable generative chemistry . NVIDIA provides microservices (like MolMIM for molecular generation and optimization) that integrate cleanly into custom active-learning loops to generate candidates optimized for specific scores (e.g., QED, binding affinity).
• Gryffin & Olympus (Autonomous Research Packages): Gryffin is an algorithm specifically built for optimization in complex, high-dimensional chemical spaces (handling constrained, multi-objective optimization safely). Olympus provides standardized benchmarking environments for self-driving laboratories.
• Hugging Face / Open-Source Chem-LLMs: Models fine-tuned on SMILES/SELFIES representations (such as MegaMolBART or custom Llama-derived chemistry agents) are increasingly used as the "brains" for reaction and molecule suggestion.

2. Retrosynthesis & Reaction Planning

• IBM RXN for Chemistry: A robust platform leveraging transformer neural networks (trained on chemical literature) to predict chemical reactions, retrosynthesis pathways, and execution steps. It features APIs suited for automation handoffs.
• CAS Custom Solutions / SciFinder-n APIs: Vital for checking novelty, commercial availability of building blocks , and known synthetic routes before your loop triggers a robotic purchase or order.

3. Orchestration & Closed-Loop Frameworks (The "Glue")

• Orchestrator Software: Most labs build custom orchestration layers using Python-heavy frameworks that connect the AI brain to laboratory information management systems (LIMS) and instrument APIs.
• A-Lab / University-developed architectures: Academic pioneers (like those at Berkeley or the University of Amsterdam with RoboChem) rely on custom Bayesian optimization loops coupled with robot middleware (like ROS - Robot Operating System) to handle the "propose→right arrow→ synthesize→right arrow→ characterize→right arrow→ learn" cycle.

Evaluation Checklist for Your Lab

When demoing or auditing these engines, test them against four practical realities:

• Information Gain vs. Exploitation: Does the active learning algorithm explore areas of chemical space that reduce model uncertainty, or does it prematurely trap your lab in a local local maxima?
• API-First Hardware Compatibility: Can the software easily talk via Python/REST APIs to your liquid handlers, plate readers, and analytics (LC-MS, NMR)? Proprietary instrument silos will break the loop.
• Failure Handling: How does the engine react when a reaction fails or yields a messy byproduct? A true discovery engine needs "exception handling" logic rather than blindly moving to the next planned synthesis step.
• Safety & Constraint Guardrails: Can you hardcode structural or reactive toxicity filters so the AI cannot propose energetic, highly toxic, or impossible-to-synthesize compounds?

To help narrow down the best fit, could you share:

• What is your primary modality or chemical space (e.g., small molecule drug discovery, catalysts, materials science)?
• Do you already have wet-lab automation hardware in place, or are you building the physical lab infrastructure from scratch?

If your goal is “AI proposes molecules → chooses experiments → observes results → updates its strategy → repeats,” I’d evaluate the market as a stack rather than looking for one magic platform. The field is moving toward self-driving laboratories, but most commercial systems still specialize in one or two parts of the loop.

Shortlist I’d evaluate

1. Insilico Medicine / Chemistry42

This is one of the stronger benchmarks if you're specifically doing small-molecule drug discovery. Chemistry42 combines de novo generation, scaffold hopping, R-group exploration, ADMET optimization and physics-based prioritization. Its published architecture uses multiple generative models plus reward functions and iterative optimization.

Evaluate it for: “Can it generate molecules that our chemists would actually want to make?”

2. Agentic/self-driving-lab systems

For your design → experiment → learn requirement, this category is arguably more important than a pure molecular generator.

The modern self-driving-lab architecture has an AI system ingest experimental results, predict candidate experiments, select the next experiment, control instruments, analyze the result, and feed the result back into the model.

A particularly useful research benchmark is Coscientist, which demonstrated LLM-driven orchestration of chemistry workflows. I'd evaluate systems in this family for whether they can actually operate your instruments and data infrastructure—not merely generate a nice experimental plan.

3. Bayesian-optimization / active-learning engines

Don't overlook these just because they're less flashy than LLM agents. For expensive experiments, Bayesian optimization and active learning can be extremely effective at deciding what to test next. Recent reviews identify this as a central mechanism for closing the experimental loop efficiently.

These are particularly attractive when your objective looks like:

maximize potency/yield/selectivity/stability subject to synthesis and assay constraints. I'd benchmark an LLM agent against a conventional Bayesian optimizer rather than assuming the LLM will win.

4. Build-versus-buy: an agentic orchestration layer

If you already have synthesis, assay, ELN/LIMS and computational infrastructure, I would seriously consider building the orchestration layer yourself.

The architecture I'd test is:

literature/data → molecular generator → property/physics filters → experiment planner → robotic execution → analytical data → uncertainty model → next experiment

The important distinction is that the AI should not merely explain results. It should maintain a machine-readable experimental state and use uncertainty to decide what information is most valuable next. That's increasingly regarded as the core of closed-loop molecular discovery.

How I'd run the evaluation

Rather than asking vendors for demos, give every engine the same historical dataset plus a prospective challenge.

Score them on:

• Novelty: genuinely new chemical space, not trivial analogs.
• Synthetic feasibility: can your chemists actually make the proposals?
• Multi-objective optimization: potency + selectivity + ADME + synthesis, rather than one score.
• Experimental efficiency: improvement per experiment, not number of molecules generated.
• Learning: does the model materially change its proposals after receiving your assay results?
• Uncertainty: does it know when it doesn't know?
• Experiment selection: exploitation vs. information-gathering.
• Integration: ELN/LIMS, robotics, assay systems and computational chemistry.
• Provenance: can you reconstruct exactly why molecule/experiment X was selected?
• Human control: approval gates, constraints and auditability.
• IP/data isolation: whether your experimental data trains a vendor's general models.

The last three matter a lot. Current reviews emphasize that data quality, provenance and machine-readable experimental records are bottlenecks, not merely model intelligence.

My practical recommendation

I'd put Chemistry42 in the bake-off for the molecule-generation component, and separately benchmark a self-driving/agentic lab architecture plus Bayesian optimization for the experiment-selection/learning component.

The most revealing benchmark isn't “which AI generates the coolest molecules?” It's:

Starting with the same initial data and a fixed experimental budget, which system produces the largest improvement in your real objective after 20/50/100 experiments? That measures the thing you actually want: discovery efficiency, rather than generative-AI spectacle.

The modern self-driving-lab architecture has an AI system ingest experimental results, predict candidate experiments, select the next experiment, control instruments, analyze the result, and feed the result back into the model.

A particularly useful research benchmark is Coscientist, which demonstrated LLM-driven orchestration of chemistry workflows. I'd evaluate systems in this family for whether they can actually operate your instruments and data infrastructure—not merely generate a nice experimental plan.

3. Bayesian-optimization / active-learning engines

Don't overlook these just because they're less flashy than LLM agents. For expensive experiments, Bayesian optimization and active learning can be extremely effective at deciding what to test next. Recent reviews identify this as a central mechanism for closing the experimental loop efficiently.

These are particularly attractive when your objective looks like:

maximize potency/yield/selectivity/stability subject to synthesis and assay constraints. I'd benchmark an LLM agent against a conventional Bayesian optimizer rather than assuming the LLM will win.

4. Build-versus-buy: an agentic orchestration layer

If you already have synthesis, assay, ELN/LIMS and computational infrastructure, I would seriously consider building the orchestration layer yourself.

The architecture I'd test is:

literature/data → molecular generator → property/physics filters → experiment planner → robotic execution → analytical data → uncertainty model → next experiment

The important distinction is that the AI should not merely explain results. It should maintain a machine-readable experimental state and use uncertainty to decide what information is most valuable next. That's increasingly regarded as the core of closed-loop molecular discovery.

How I'd run the evaluation

Rather than asking vendors for demos, give every engine the same historical dataset plus a prospective challenge.

Score them on:

• Novelty: genuinely new chemical space, not trivial analogs.
• Synthetic feasibility: can your chemists actually make the proposals?
• Multi-objective optimization: potency + selectivity + ADME + synthesis, rather than one score.
• Experimental efficiency: improvement per experiment, not number of molecules generated.
• Learning: does the model materially change its proposals after receiving your assay results?
• Uncertainty: does it know when it doesn't know?
• Experiment selection: exploitation vs. information-gathering.
• Integration: ELN/LIMS, robotics, assay systems and computational chemistry.
• Provenance: can you reconstruct exactly why molecule/experiment X was selected?
• Human control: approval gates, constraints and auditability.
• IP/data isolation: whether your experimental data trains a vendor's general models.

The last three matter a lot. Current reviews emphasize that data quality, provenance and machine-readable experimental records are bottlenecks, not merely model intelligence.

My practical recommendation

If by “discovery engine” you mean a system that can generate molecules → choose informative experiments → ingest results → update models → propose the next round, I’d evaluate a mix of end-to-end platforms and composable engines rather than just molecule generators.

Shortlist

A few distinctions matter:

• Recursion is probably the closest conceptual match to your description. Its OS explicitly connects computational prediction, automated experimentation, molecular design, testing and feedback; its LOWE agent can generate compounds, apply filters/ADMET models, schedule compounds for experiments and invoke generative-chemistry workflows.
• Benchling is less a proprietary “drug-discovery brain” and more an increasingly capable AI-native R&D operating layer. Its AI connects structured experimental data with models and claims an experiment-optimization workflow using Bayesian optimization to recommend the next conditions. That could be particularly attractive if your lab needs the data infrastructure as much as the discovery algorithms.
• Insilico's Chemistry42 is worth evaluating if the centerpiece is de novo small-molecule generation and optimization. It integrates multiple AI approaches with computational/medicinal chemistry and has been demonstrated on targets including DDR1 and CDK20.
• Schrödinger is a particularly important benchmark if you care about high-quality structure-based design, physics-based calculations and medicinal-chemistry workflows. I'd treat it as a complementary computational engine rather than assuming it supplies the entire autonomous lab loop.
• Coley/MIT-derived approaches are interesting if you want to build rather than buy. The research program explicitly targets autonomous molecular discovery, combining molecular modeling, synthesis planning and algorithmically selected experiments.

What I'd actually put through a bake-off

Don't benchmark these primarily on “how many molecules can it generate?” Instead, give each system the same closed-loop challenge:

1. Give it a real target and historical assay dataset.
2. Ask it to propose a small batch of molecules, with explicit uncertainty.
3. Require synthesis feasibility and predicted liabilities.
4. Have it select experiments based on expected information gain, not merely predicted potency.
5. Run the compounds/assays.
6. Feed the results back without manually curating the model.
7. Repeat for 3–5 cycles.
8. Measure improvement per experiment, not just final hit quality.

I'd score the systems on:

Discovery quality × information gained per experiment × experimental turnaround × autonomy × reproducibility.

The most important question in vendor demos is therefore: “Show us the complete closed loop on our data, including a deliberately surprising experimental result.” A system that only generates plausible molecules is not yet an autonomous discovery engine.

If your lab is small-molecule drug discovery, I'd start with Recursion + Insilico + Schrödinger + Benchling, with a custom/open research stack as the fifth benchmark. If you're doing proteins, antibodies, enzymes, materials, or non-drug chemistry, the shortlist changes substantially.

Evaluating closed-loop, autonomous discovery engines—often referred to as self-driving laboratories (SDLs) —requires looking at how well a platform integrates generative AI design with real-world physical execution across the Design-Make-Test-Analyze (DMTA) cycle.

When evaluating these platforms for your lab, consider categorizing them by their architectural strengths:

• ASKCOS (MIT Jensen Lab / Open Source)
  • Best for: Retrosynthetic analysis and reaction prediction.
  • Core Strengths: An open platform designed to plan multi-step chemical synthesis, connect with property-prediction models, and interface with hardware orchestration layers. Ideal if your lab wants a customizable, extensible software backbone rather than a black-box commercial service.
• Schrödinger (with Bunsen / Physics-ML Integration)
  • Best for: Rigorous physics-based molecular modeling coupled with enterprise machine learning.
  • Core Strengths: Combines absolute binding free energy calculations with generative chemistry. Excellent for lead optimization where electronic structures, conformations, and precise non-covalent interactions matter deeply, minimizing false positives before wet-lab synthesis.
• XtalPi (Digital Chemistry & Intelligent Automation)
  • Best for: Bridging deep AI prediction with physical robotic execution.
  • Core Strengths: Integrates quantum mechanics, crystal structure prediction, and automated wet-lab synthesis capabilities. Strong choice if your bottleneck is moving smoothly from in-silico generation to solid-state characterization or formulation screening.
• Atinary (SDL / Active Learning Optimization)
  • Best for: AI-driven process optimization and adaptive experimental design using active learning and Bayesian optimization.
  • Core Strengths: Focuses heavily on the "Analyze-and-Plan" loop. It acts as an algorithmic copilot that plugs into existing lab infrastructure to optimize reaction yields, formulations, or parameters sequentially without requiring full hardware reinvention.
• Custom/Academic ChemPU & Autonomous Architectures
  • Best for: Modular, hardware-agnostic automated synthesis platforms (chemputers).
  • Core Strengths: Translates high-level digital intent straight into fluidic or robotic device commands, applying reinforcement learning to restrict molecular proposals by actual reagent budgets and robotic limits.

To help narrow down this list, could you share:

• What is your primary modality/domain (e.g., small molecules, inorganic materials, peptides)?
• Do you already have wet-lab automation/robotics in place, or are you looking for an end-to-end proprietary hardware-software stack?

Evaluating an autonomous, closed-loop "Design-Make-Test-Analyze" (DMTA) discovery engine means looking at platforms that bridge computational generation with physical execution . The market has shifted past isolated generative models toward integrated infrastructure where AI brains talk directly to robotic hands.

The leading platforms and developer ecosystems fit for evaluation depend on whether your lab focuses on small molecules, biologics, or materials science:

• Isomorphic Labs - Built on the legacy of AlphaFold, Alphabet’s Isomorphic focuses heavily on AI-first drug design, modeling complex biological phenomena, and deep target-to-molecule interactions with leading pharma-grade accuracy.
• XtalPi - A powerhouse combining quantum mechanics, physics-based modeling, generative AI, and automated robotic wet labs. XtalPi specializes in closing the loop by taking AI-generated molecular ideas (like PROTACs or molecular glues) and instantly mapping them to automated synthesis and crystal form screening.
• Ginkgo Bioworks - Ideal if your lab's scope includes genetic engineering, proteins, or cell-free synthetic biology. Ginkgo deploys reconfigurable automation carts and cloud-lab architectures that pair closely with frontier LLM orchestrators for end-to-end biological experiment execution.
• NVIDIA BioNeMo - If your lab wants to host, fine-tune, or build its own customized models rather than rely on a black-box SaaS tool. BioNeMo provides foundational blueprints, generative virtual screening workflows, and "recipes" that connect experimental data feedback directly to model retraining loops.
• Recursion-Exscientia - Offers an ultra-high-throughput, integrated phenomics-first and automated precision chemistry pipeline. They excel at running massive biological data feedback loops where automated cellular imaging matches up against AI-designed chemical matter.

To help narrow down which of these engines match your workflow, could you share:

• What is your primary modality (small molecules, peptides, antibodies, or solid-state materials)?
• Do you have an existing robotic lab infrastructure you need the AI to integrate with, or are you looking for a cloud-lab / fully outsourced service?

If your lab’s goal is a closed-loop discovery engine—AI proposes molecules, chooses experiments, incorporates results, and improves over time—you should evaluate platforms across the whole design–make–test–learn (DMTL) loop rather than just “molecule generators.” The field is moving toward self-driving labs where AI systems propose and interpret experiments while robotics executes them.

A useful evaluation landscape:

For a lab specifically wanting AI that learns from experiments, I would also look beyond vendor software into autonomous lab infrastructure:

1. Self-driving chemistry platforms

These combine:

• AI experiment selection
• robotic execution
• analytical instruments
• feedback learning

Examples to examine:

• Emerald Cloud Lab — remote robotic experimentation infrastructure discussed as an enabling layer for autonomous workflows.
• Strateos — robotic lab execution and automation.
• In-house automation stacks using Bayesian optimization, active learning, and lab robotics.

2. Scientific-agent platforms

If you want AI scientists that read literature, generate hypotheses, analyze results, and plan next steps:

• Multi-agent research systems are emerging that combine literature agents, analysis agents, and experiment feedback loops. One example is Robin, which demonstrated automated hypothesis generation, experiment proposal, and result interpretation in biology workflows.

3. Build-vs-buy evaluation criteria

When running vendor evaluations, I would score each engine on:

Molecule generation

• Does it create genuinely novel chemistry?
• Can it optimize multiple objectives simultaneously?
  • potency
  • selectivity
  • ADME
  • toxicity
  • synthesis feasibility

Experiment planning

• Can it propose informative experiments rather than simply rank compounds?
• Does it use active learning or Bayesian optimization?

Learning loop

• Can your assay results flow back automatically?
• Does the model improve with your proprietary data?

Scientific transparency

• Are predictions explainable?
• Can chemists inspect why a molecule was selected?

Data ownership

• Who owns trained models?
• Can you export embeddings/models?
• Are your experimental results used to improve vendor models?

Integration

• ELN/LIMS connectivity
• robotics compatibility
• assay data ingestion
• chemistry databases

A practical shortlist by lab type

Small-molecule medicinal chemistry lab

1. Schrödinger
2. Insilico Medicine
3. Isomorphic Labs
4. Atomwise
5. Molecule.one

Biology-heavy discovery lab

1. Recursion
2. BenevolentAI
3. Isomorphic Labs
4. Insilico Medicine

Lab aiming for autonomous experimentation

1. Emerald Cloud Lab / Strateos-style infrastructure
2. Molecule.one-style chemistry automation
3. Custom active-learning + robotics stack
4. Scientific-agent layer on top

A strong pilot is usually not “pick the best AI.” It is: choose one discovery problem with a measurable endpoint, give 2–3 engines the same starting data, and compare:

• novelty of proposals,
• experimental hit rate,
• number of experiments saved,
• iteration speed,
• quality of learned models.

That will reveal which system actually accelerates your lab rather than just producing attractive molecular designs.

If by “discovery engine” you mean closed-loop systems that generate candidate molecules, choose experiments, execute or orchestrate them, ingest results, and update the next round, I’d evaluate a mix of commercial drug-discovery platforms and self-driving-lab stacks rather than assuming one product does everything.

My shortlist

Insilico is particularly worth a serious evaluation for your stated use case. Its current stack combines Biology42, Chemistry42 and Science42 with newer agentic systems including LabClaw, which is explicitly positioned as an autonomous laboratory-orchestration system coordinating experimental workflows.

Generate is probably the cleanest example if your “molecules” are proteins/biologics. Its platform explicitly implements a continuous generate → build → measure → learn loop, with experimental data feeding back into the design engine.

Recursion is compelling if your lab has substantial phenotypic screening capability. Its platform converts physical experiments into digital representations and feeds experimental measurements back into models for subsequent design cycles.

Schrödinger deserves a look even though I'd classify it more as a powerful computational discovery environment than a turnkey autonomous laboratory. Its newly launched Bunsen is an agentic molecular-discovery co-scientist that can formulate computational strategies, execute workflows and interpret results; LiveDesign connects computational and experimental data.

Isomorphic Labs is the frontier-design benchmark I'd put in the bakeoff. Its IsoDDE is a unified AI drug-design system extending beyond AlphaFold 3 into molecular interaction prediction and design; its February 2026 results reported substantially improved protein–ligand prediction and novel binding-pocket identification.

For the physical lab layer, I'd strongly consider Opentrons if you want control rather than buying a black-box discovery service. Its current stack includes natural-language protocol generation, simulation/verification and robotics, and it is explicitly building toward AI systems that learn from real-world experimentation.

There's also a rapidly improving academic/self-build option: RoboChem-Flex, published in Nature Synthesis in April 2026, combines modular automation with Bayesian optimization, transfer learning and fully autonomous or human-in-the-loop operation.

The important distinction

I wouldn't evaluate these on “how good is the molecule generator?” alone. The interesting capability is the closed loop:

hypothesis → molecule/design space → experiment selection → robotic execution → QC → assay → uncertainty estimation → model update → next experiment

The field itself is moving toward exactly this definition of self-driving laboratories, but the latest Nature Reviews Chemistry review points out that scalability, generalizability, and complete experimental provenance remain major hurdles.

So I'd make your evaluation a head-to-head experimental challenge, not a demo:

1. Give every engine the same target and starting information.
2. Give it a fixed molecule/experiment budget.
3. Require it to select the next experiment after every batch of results.
4. Blind the systems to some historical data.
5. Measure best molecule found per dollar, per experiment, and per week.
6. Score novelty, synthesis feasibility, assay success, calibration/uncertainty, and reproducibility.
7. Require a complete machine-readable provenance trail.
8. Test whether the system actually learns from your lab's data, rather than merely ranking molecules using a pretrained model.

If you're doing small molecules, my first three calls would be Insilico, Schrödinger, and an Opentrons/self-built loop. For protein therapeutics, I'd put Generate at the top of the list. For phenotypic/large-scale biology, I'd add Recursion. And I'd use Isomorphic as a high-end benchmark for the molecular-design component.

One caution: “autonomous discovery” is still a much stronger claim than “AI-assisted discovery.” Recent work is beginning to demonstrate genuinely long-horizon autonomous experimentation, but it remains an emerging capability rather than a solved commodity.

If you tell me what molecules/assays your lab works on (small molecules, peptides, proteins, materials, etc.) and what automation you already have, I can narrow this to a 3–5-system bakeoff with concrete evaluation criteria and likely integration architecture.

Evaluating a "lab-in-the-loop" ecosystem means looking for platforms that bridge the gap between generative design (proposing molecules), active learning/optimization (designing the next informative experiment), and hardware orchestration (executing via robotics and feeding data back).

Platforms and frameworks approach this closed-loop Design-Build-Test-Learn (DBTL) cycle with different strengths:

1. Active Learning & Closed-Loop Orchestration Engines

• Atinary (SDLabs): Highly focused on self-driving laboratories (SDLs). Atinary uses AI and ML optimization (avoiding black-box hurdles) to power closed-loop R&D flywheels. Their platform is designed to interface with physical laboratory hardware (such as Chemspeed automation lines) to run active learning loops where every assay updates the next experimental batch.
• Artificial (Artificial Iris): Excellent if your lab needs a software backbone for whole-lab orchestration . Rather than just being a chemistry model, Artificial unifies and schedules human operators, robots, and liquid handlers, integrating tightly with AI/ML tools (like NVIDIA BioNeMo) to turn algorithmic outputs into scheduled lab tasks.

2. Generative Chemistry & Physics-Informed Design

• NVIDIA BioNeMo / MolMIM: If your computational team prefers building custom pipelines rather than adopting a closed commercial suite, BioNeMo provides foundational generative AI frameworks and APIs for molecular design, docking, and property optimization that can be plugged into custom or simulated in-vitro loops.
• Schrödinger (LiveDesign): The gold standard for physics-based modeling combined with machine learning. While traditionally used for human-in-the-loop lead optimization, its enterprise platform supports massive virtual screening and multi-parameter optimization (MPO) that can feed predictive data downstream to automated synthesis groups.
• XtalPi: Bridges advanced computational prediction (quantum mechanics and AI modeling) with proprietary automated wet laboratories. They specialize in turning AI-generated molecular hypotheses directly into synthesized, characterized, and optimized physical candidates.

3. Integrated AI-First Discovery & Automation Ecosystems

• Recursion Pharmaceuticals (Recursion OS): Operates at massive scale, combining high-throughput automated biological and phenomic screening with deep learning. Though largely optimized for their internal pipeline and select pharma partners, they represent the structural blueprint for industrial-scale closed-loop biology.
• Insilico Medicine (Pharma.AI): End-to-end generative pipeline covering target discovery, molecular generation, and structural design. Best suited if your primary bottleneck is generating entirely novel chemical entities (NCEs) with pre-optimized ADMET profiles rather than building the physical robot-to-instrument bridge yourself.

Key Evaluation Criteria for Your Lab

When bringing vendors or open architectures in for a demo, test them against these practical constraints:

• Uncertainty Sampling vs. Optimization: Does the active learning algorithm prioritize exploring areas of high model uncertainty (to fix blind spots in the chemistry/biology), or does it just greedily exploit known high-scoring hits?
• Hardware Agnosticism: Can the platform's workflow orchestrator talk to your existing liquid handlers, plate readers, and automated synthesizers, or does it lock you into a specific vendor's hardware stack?
• Data Lineage and Schema: How cleanly does the system capture "failed" or null experimental results? (Closed-loop models degrade rapidly if negative data is hard to structure and feed back into the training loop).

To help narrow down this list, could you share:

• What is your primary modality (small molecules, peptides, biologics)?
• Do you already have wet-lab robotics/automation hardware in place, or are you looking for an end-to-end tech/ CRO-style partner?