Source-linked AI summary
Autonomous discovery in the chemical sciences part II: Outlook
Connor W. Coley, Natalie S. Eyke, Klavs F. Jensen
TL;DR
The paper asks whether key workflow components have closed the gap toward autonomous chemical discovery and examines models, validation, and evaluation challenges. Reported workflows can reduce experiments by modest factors or by 2–3 orders of magnitude, while datasets and sequential workflows remain constrained.
Problem
The paper asks whether progress has closed the gap toward key workflow components for autonomous discovery.
Method
The paper examines workflow models, including a collector model that treats domain knowledge as elements to be identified.
Results
Experiments required to find high-performing candidates can decrease by factors of 2–10 or by 2–3 orders of magnitude.
Takeaways & Limitations
Evaluation should prioritize metrics that assess identifying items on which models were not trained.
Takeaways & Limitations
Sequential workflow matching remains challenging because workflows may require plans for parking.
Abstract
from arXiv · showhide
This two-part review examines how automation has contributed to different aspects of discovery in the chemical sciences. In this second part, we reflect on a selection of exemplary studies. It is increasingly important to articulate what the role of automation and computation has been in the scientific process and how that has or has not accelerated discovery. One can argue that even the best automated systems have yet to ``discover'' despite being incredibly useful as laboratory assistants. We must carefully consider how they have been and can be applied to future problems of chemical discovery in order to effectively design and interact with future autonomous platforms. The majority of this article defines a large set of open research directions, including improving our ability to work with complex data, build empirical models, automate both physical and computational experiments for validation, select experiments, and evaluate whether we are making progress toward the ultimate goal of autonomous discovery. Addressing these practical and methodological challenges will greatly advance the extent to which autonomous systems can make meaningful discoveries.
2 Reflection on case studies
The case studies show substantial progress in automating discovery workflows, especially experiment selection and validation, but most systems still pursue narrowly defined objectives within constrained spaces. Computer-assisted selection can reduce experiments needed to find high-performing candidates, while broader autonomous discovery and generalizable conclusions remain open goals.
- Assessment framework: The review assesses case studies using goal breadth, search-space constraints, experiment selection, validation, interpretation, and contribution to scientific knowledge.These questions frame how far workflows have progressed toward autonomous discovery.
- Discovery goals: Most studies optimize a specific scalar performance objective rather than identify a therapeutic, material, reaction, or explanatory model.The authors characterize discovery goals as narrow in most studies.
- Search spaces: Search spaces are constrained for automated validation, convenience, reduced dimensionality, smaller candidate libraries, or assumptions of accurate validation across the full space.Human operators implicitly or explicitly assume that good solutions exist in the restricted spaces.
- Experiment selection: Formal experiment-selection frameworks quantify performance or information value, but uncertainty-aware active learning is less common than ad hoc diversity or greedy selection.Model-free genetic algorithms remain prevalent and require domain experts to define allowable mutations.
- Validation and automation: Automated validation, computational pipelines, and programmatic process control have advanced, but some workflows retain manual operations or manually performed validation.Iterative validation distinguishes autonomous workflows from merely automated calculations.
- Autonomous discovery: The review identifies high-level, human-interpretable generalized conclusions beyond identifying a molecule, material, device, process, or black-box model as a remaining goal.Existing systems have nevertheless offloaded much of the manual and mental burden of navigating design spaces.
3 Challenges and trends
Autonomous discovery capabilities are converging rapidly, but major challenges remain in handling complex data, automating validation and feedback, selecting experiments, and evaluating progress.
- Challenges and trends: Key remaining challenges concern complex data, empirical models, automated validation and feedback, experiment selection, and evaluation.The section presents these as directions requiring additional effort.
3.1 Working with complex data
Chemical discovery increasingly depends on complex datasets, computational methods, and accessible databases, but progress is constrained by inconsistent data quality, representation challenges, and limited interpretability. The section identifies open challenges in standardizing, curating, modeling, and extracting knowledge from chemical data.
- Working with complex data: Complex data can be large, unstructured, or relationally difficult enough that algorithms identify relationships humans cannot readily recover.The review distinguishes complex data from merely large datasets by emphasizing content and inferential difficulty.
- Working with complex data: Supervised learning uses input-output datasets to build surrogate models, while unsupervised methods infer dataset structure and support generative proposals for new matter.The review presents these methods as complementary ways to model existing chemical data and propose new materials.
- Creating and maintaining datasets: Open databases can have outsized discovery impact, but creating and maintaining them requires data capture, standardized representations, privacy-preserving sharing, and searchable experimental records.Standardization is particularly difficult for polymeric materials with stochastic structures and process-dependent attributes.
- Working with complex data: Information extraction, natural language processing, ontologies, embeddings, knowledge graphs, and causal models convert chemical literature into analyzable structure.These approaches support mining published text and organizing chemical knowledge for discovery.
- Creating and maintaining datasets: Reliable modeling requires curation because datasets may be underspecified, incomplete, misentered, or insufficiently validated before use.The review contrasts independently validated PDB data with PubChem data lacking prerequisite curation and notes errors in databases including PDB and Reaxys.
3.2 Automated validation and feedback
Automated validation enables iterative hypothesis testing by rapidly generating data, falsifying false positives, and exploring broad design spaces. Its expansion remains constrained by experimental hardware, workflow integration, synthesis planning, and the limits of computational and physical validation.
- Automated validation and feedback: Iterating between hypothesis generation and validation can refine or falsify hypotheses, while automation can quickly identify false positives.Automation supports a “fast to fail” workflow when new data are collected to test hypotheses.
- Automated validation and feedback: High-throughput and parallelized experimentation have transformed validation by simplifying screening across broad design spaces.Split-and-pool and other combinatorial methods allow multiple candidates to be studied simultaneously.
- Experimental validation: Experimental automation is routine for well-plate operations, but complex procedures, milligram-scale solid dosing, and full synthesis remain difficult or require custom platforms.A universal chemical synthesizer remains elusive, and available hardware can constrain design spaces beyond intentional prior-knowledge narrowing.
- Experimental validation: Integrated platforms must coordinate independent hardware and software for scheduling, error handling, analysis, and throughput matching, but nonstandard interfaces and data formats increase complexity and cost.Sequential workflows also need plans for parking or stabilizing samples when downstream bottlenecks occur.
- Experimental validation: Automated synthesis requires implementable multistep route planning, reaction conditions, quantities, temperatures, times, and intermediate and final purification.Robotic platforms can execute precisely planned steps, but current CASP tools cannot provide recommendations with sufficient implementation precision.
3.3 Selection of experiments for validation and feedback
The review identifies experiment selection as central to autonomous discovery, emphasizing uncertainty, information gain, experimental cost, and higher-order discovery goals. It argues that current approaches remain limited by unreliable uncertainty estimates, cost-insensitive selection, and prescribed hypotheses.
- Uncertainty and applicability: Accurate uncertainty quantification can direct attention toward underexplored regions and help triage experiments.Statistical and probabilistic frameworks support experiment selection, but meaningful uncertainty estimates remain necessary.
- Uncertainty and applicability: QSAR/QSPR models struggle with epistemic uncertainty and activity cliffs, where small structural changes can cause sharp affinity changes.These limitations challenge direct connections between structural and functional similarity.
- Cost and information gain: Experiment selection should quantify the tradeoff between information gain and experimental difficulty through cost-sensitive active learning.Current methods rarely account for experiment cost quantitatively, although synthesis and simulation can differ substantially in cost and failure likelihood.
- Cost and information gain: For batched optimization, synthesis cost depends on shared precursors and reaction conditions rather than simply summing individual molecule costs.This makes cost modeling more complex for parallel libraries.
- Higher-level discovery goals: Autonomous discovery requires broader objectives and more open-ended hypothesis generation, but optimization strategies ultimately require a reducible scalar objective.Knowledge discovery, unsupervised learning, novelty detection, and reinforcement learning are identified as possible routes toward greater flexibility.
- Higher-level discovery goals: Validation experiments should pursue higher-order discovery goals rather than only maximizing performance within a designed library.Early information-gathering experiments may be valuable because they support eventual identification of the best candidate.
3.4 Evaluation
The review frames evaluation as a major barrier to autonomous discovery: predictive models must demonstrate extrapolation, while closed-loop systems need realistic tests beyond narrow proof-of-concept studies. It proposes benchmark problems, with computational chemistry as a practical testing ground, while cautioning against inaccessible compute and data requirements.
- Predictive-model evaluation: Predictive models should be evaluated for extrapolative power: generalization to chemical design spaces absent from training.The review prioritizes this capability over brute-force experimentation in known spaces.
- Closed-loop evaluation: Closed-loop design, synthesis, and testing studies remain proof-of-concepts restricted to narrow chemical spaces, limiting practical utility.Automated validation constraints prevent many interesting questions and optimization objectives from being addressed.
- Benchmark problems: Autonomous-discovery research lacks realistic benchmarks for hypothesis generation, experiment selection, and related tasks.Existing evaluations such as the coupon-collector model represent limited problem formulations.
- Benchmark problems: Computational chemistry may provide a practical playground for advancing autonomous-discovery methods because physical validation presents substantial obstacles.A benchmark sandbox should preserve some complexity of real domain applications without eliminating it entirely.
- Benchmark problems: Standard evaluations are also absent for interpretability, uncertainty quantification, and generalizability in empirical QSAR/QSPR modeling.The field needs shared problem formulations that describe discovery tasks of interest to domain experts.
- Benchmark design: Benchmarking can disadvantage research groups when competitive performance requires inaccessible compute or restricted training data.The review cautions that quantitative evaluation should control what training data and resources are permissible.
4 Conclusion
Automation and computation have become ubiquitous and productive in chemical science, but few systems have achieved high autonomy or can yet claim discovery itself. The outlook therefore emphasizes improving data, models, validation, experiment selection, and evaluation while recognizing distinct human and machine capabilities.
- Automation reduces manual effort and enables new discovery approaches at unprecedented throughput, but its direct contribution to discovery remains an open question.The authors distinguish useful laboratory assistance from autonomous discovery itself.
- Few studies achieve high autonomy because humans still specify objectives, define search spaces, and identify the relevant regions to explore.Subject-matter experts remain necessary for constructing platforms and associated workflows.
- Discovery standards and confidence requirements vary by domain, especially as computational validation and proxy assays become more accurate.The review intentionally avoids a precise confidence threshold and notes that some predicted properties still require direct experimental observation.
- Autonomous systems may discover through capabilities distinct from human work, including searching larger design spaces, generating more data, and designing experiments using extensive existing information.The authors argue that platforms should excel at tasks humans cannot easily or safely perform rather than imitate human discovery exactly.
- The outlook prioritizes high-quality datasets, interpretable and generalizable models, broader automated experiments, faster computational validation, cost-sensitive active learning, open-ended hypothesis generation, and realistic benchmarks.Benchmarks should test extrapolation into underexplored design spaces that are expensive or intractable to explore.