How many compounds will be screened in total during an active learning campaign?

The total number depends on library size, number of cycles, and hit criteria. Typically, we screen a subset of the full library over multiple cycles. The exact scope is defined during project scoping based on your budget and desired confidence.

How do you ensure the initial compound subset is representative of the whole library?

We use chemical fingerprint clustering (e.g., Morgan fingerprints, Tanimoto similarity) to select a diverse set that covers the full chemical space of your library. Statistical sampling methods ensure no major chemotype is missed during the initial screen.

What machine learning algorithms are used to guide compound selection?

We primarily use random forests and gradient-boosted trees for their robustness with small datasets. For larger campaigns, neural-network-based models (e.g., deep neural networks) are implemented. The algorithm is chosen based on data size and target biology.

How are hits validated at the end of the active learning campaign?

Top compounds from the final model are retested in dose-response curves (8–10 point) in the primary assay. Selected hits are further validated in orthogonal assays such as SPR, DSF, or selectivity panels, depending on project scope.

Can active learning be applied to phenotypic screening assays?

Yes, active learning is well suited for phenotypic assays. Our platform supports high-content imaging readouts and uses the same iterative model-guided approach to prioritize compounds that modulate disease-relevant cellular phenotypes.

How do you determine when to stop the active learning loop?

Stopping criteria are defined during project scoping and may include a fixed number of cycles, a budget limit, or a convergence metric such as no significant improvement in model performance. Final hit criteria also determine when the loop ends.

Services

Online Inquiry

Name *

Email *

Subject *

Phone

Your Message

How Did You Hear About Us

High-Throughput Screening with Active Learning Closed Loop

Active Learning HTS

Iteratively combine wet-lab screening with machine learning to identify high-quality hits while dramatically reducing the total number of assays required.

Initial diverse compound subset screened to train predictive models
Uncertainty-guided selection of next compounds in each cycle
Closed-loop iteration until hit criteria are met or budget is exhausted
Final hit lists validated with orthogonal assays for confidence
Rapid, data-driven hit identification for small molecules and fragments

Start your campaign

How does active learning reduce screening effort?

Instead of screening every compound exhaustively, active learning HTS focuses resources on the most informative compounds, reducing total assays while maintaining hit quality. The exact reduction is project-dependent and scoped during consultation.

Uncertainty-guided compound selection maximizes learning per screen
Multi-cycle closed-loop optimization adapts to emerging data
Cost-efficient hit identification without sacrificing hit quality
Scalable from hundreds of thousands to millions of compounds
Orthogonal validation ensures biological relevance

Iterative

Data-driven

Closed-loop

Closed-Loop Workflow

From diverse library to validated hits in iterative cycles

Our active learning HTS service integrates a fully automated screening platform, machine learning model building, and orthogonal hit confirmation to accelerate hit discovery.

Active Learning Workflow

We start with a chemically diverse subset of your compound library, screen it against your target, and use that data to train a predictive model. The model's uncertainty or expected improvement guides selection of the next compounds to test. This closed loop repeats until hit criteria are met.

Select initial diverse compound subset (e.g., based on chemical fingerprint diversity)
Perform high-throughput screening on initial subset (96/384/1536-well plates)
Train machine learning model using screening data (e.g., random forest, neural network)
Acquisition function (e.g., uncertainty sampling) chooses next compounds
Screen newly selected compounds and update model iteratively

Hit Prioritization & Validation

After the final cycle, we deliver a prioritized hit list with predicted activities and confidence scores. Selected compounds are then tested in orthogonal assays (e.g., dose-response, selectivity, biophysical binding) to confirm on-target activity and reduce false positives.

Final hit list ranked by predicted activity and model confidence
Dose-response confirmation for top hits
Selectivity profiling against related targets (scoped per project)
Biophysical validation (e.g., SPR, DSF) available on request
Comprehensive report with data, models, and recommendations

Assay Quality

Rigorous assay optimization and miniaturization ensure robust Z'-factors, low interference, and high reproducibility across plates and cycles.

Statistical Rigor

We apply appropriate multiple-testing corrections and cross-validation to avoid overfitting and ensure that reported hit rates are reliable.

Biological Interpretability

Our reports include cheminformatics analysis, structure–activity relationship insights, and model interpretability to guide downstream optimization.

What you receive

Complete data and reporting package for each active learning campaign

Transparent deliverables from raw assay data to final hit list and model performance metrics.

Data Deliverables

Raw and normalized screening data (plate-level and compound-level)
Machine learning model performance metrics (e.g., ROC-AUC, enrichment factor)
Final prioritized hit list with predicted pIC50 / % inhibition and confidence scores
Dose-response data for confirmed hits (IC50, Hill slope)

Reporting & Support

Comprehensive study report detailing workflow, methods, and results
Cheminformatics analysis (scaffold diversity, physicochemical properties)
Project management with regular updates throughout the campaign
Consultation on hit follow-up strategies (e.g., analog selection, secondary assays)

Service Scope

Customizable parameters scoped per project

Every active learning HTS campaign is configured to your target, library, and budget. The table below outlines typical options discussed at project initiation.

Parameter	Typical Project Scope
Library size	100,000–1,000,000+ compounds (scoped per project)
Assay format	96-, 384-, or 1536-well plates
Detection technology	Fluorescence, luminescence, HTRF, AlphaLISA, high-content imaging – selected based on target
Number of active learning cycles	Determined by budget and hit criteria (scoped per project)
Hit criteria	Defined during consultation (e.g., ≥ threshold inhibition at a given concentration, scoped per project)
Validation assays	Dose-response, selectivity profiling, biophysical binding (scoped per project)
Compound source	Client library, third-party library, or our curated diversity set (as agreed)
Machine learning model type	Random forest, gradient boosting, neural network – selected based on data size

How It Works

Five-stage closed-loop workflow from target to validated hits

A fully integrated process that combines automation, compound management, and active learning to maximize hit discovery efficiency.

Assay Development & Miniaturization

We develop and optimize a robust, miniaturized assay for your target (biochemical or cell-based) with appropriate readout, ensuring high Z'-factor and reproducibility across 384/1536-well plates.

Initial Diverse Subset Selection

Using cheminformatics (e.g., chemical fingerprint clustering), we select a chemically diverse subset of the full library (size scoped per project) to maximize structural coverage for model training.

Primary Screening of Initial Subset

The initial subset is screened against your target in singlicate or duplicate. Raw data are normalized, quality-controlled, and then passed to the machine learning module.

ML Model Training & Uncertainty-Guided Selection

We train a predictive model (e.g., random forest) on the initial screening data. An acquisition function (e.g., uncertainty sampling) then selects the next set of compounds – those with highest predicted activity or greatest model uncertainty – for the next cycle.

Iterative Screening Rounds & Hit Validation

Newly selected compounds are screened and the model is updated. After a predetermined number of cycles or once hit criteria are met, top compounds are validated in dose-response and orthogonal assays to deliver a final hit list.

Applications

Accelerate hit discovery across multiple modalities

Active learning HTS is especially powerful when screening large libraries against difficult targets or when compound supply is limited.

Rapidly identify tractable start points for medicinal chemistry

Our closed-loop approach is ideal for traditional small-molecule libraries, reducing total screens by focusing on the most informative compounds while still covering broad chemical space.

Proven for diverse targets including GPCRs, kinases, and protein–protein interactions
Generates SAR insights from early cycles to guide later compound selection
Final hits are drug-like and suitable for lead optimization

Efficiently screen fragment libraries (100–300 Da) with active learning

Fragment libraries are small in size but require sensitive detection. Active learning can identify weak binders more efficiently than exhaustive screening of all fragments.

Biophysical validation (SPR, DSF, NMR) integrated into the closed loop
Initial subset based on fragment diversity (e.g., scaffold heterogeneity)
Model predicts binding free energies or affinity ranges

Prioritize compounds that modulate disease-relevant cellular phenotypes

Phenotypic assays often have high complexity and cost per well. Active learning reduces the number of wells needed while retaining the ability to discover novel mechanisms.

High-content imaging readouts (e.g., neurite outgrowth, toxicity markers)
Trained model predicts phenotypic response from molecular fingerprints
Iterative selection focuses on compounds with desired phenotypic signatures

Frequently Asked Questions

Active Learning HTS – addressing your key questions

Additional details can be discussed during project scoping.

The total number depends on library size, number of cycles, and hit criteria. Typically, we screen a subset of the full library over multiple cycles. The exact scope is defined during project scoping based on your budget and desired confidence.
We use chemical fingerprint clustering (e.g., Morgan fingerprints, Tanimoto similarity) to select a diverse set that covers the full chemical space of your library. Statistical sampling methods ensure no major chemotype is missed during the initial screen.
We primarily use random forests and gradient-boosted trees for their robustness with small datasets. For larger campaigns, neural-network-based models (e.g., deep neural networks) are implemented. The algorithm is chosen based on data size and target biology.
Top compounds from the final model are retested in dose-response curves (8–10 point) in the primary assay. Selected hits are further validated in orthogonal assays such as SPR, DSF, or selectivity panels, depending on project scope.
Yes, active learning is well suited for phenotypic assays. Our platform supports high-content imaging readouts and uses the same iterative model-guided approach to prioritize compounds that modulate disease-relevant cellular phenotypes.
Stopping criteria are defined during project scoping and may include a fixed number of cycles, a budget limit, or a convergence metric such as no significant improvement in model performance. Final hit criteria also determine when the loop ends.

Scientific Literature

References

Chen J, Hou J, Wong KC. Categorical Matrix Completion With Active Learning for High-Throughput Screening. IEEE/ACM transactions on computational biology and bioinformatics. 2021;18(6):2261-2270. View on PubMed
Abranches DO, Dean W, Muñoz M, et al. Combining High-Throughput Experiments and Active Learning to Characterize Deep Eutectic Solvents. ACS sustainable chemistry & engineering. 2024;12(38):14218-14229. View on PubMed
Liu X, Ren X, Li P, et al. Research progress of artificial intelligence in high-throughput drug screening. Frontiers in pharmacology. 2026;17:1847031. View on PubMed

Please kindly note that our products and services are for research use only.