AdaptiveFlow: AI Cuts Ultra-Large Virtual Screening Costs by 1,000x
Ultra-large virtual screening (ULVS) has a fundamental scaling problem: as chemical libraries grow from millions to billions of compounds, exhaustive molecular docking becomes increasingly expensive in both compute time and infrastructure.
Researchers from St. Jude Children’s Research Hospital, the University of Pavia, Dana-Farber Cancer Institute, and Harvard Medical School have developed AdaptiveFlow, an open-source AI platform designed to address that bottleneck.
Rather than docking every molecule with equal computational effort, AdaptiveFlow uses Adaptive Target-Guided Virtual Screening (ATG-VS) to identify promising regions of chemical space and progressively concentrate compute resources where they are most likely to produce useful hits.
The platform was demonstrated against the 69-billion-compound Enamine REAL Space library and achieved approximately a 1,000x reduction in computational cost compared with brute-force ultra-large screening. Its cloud implementation also demonstrated near-linear scaling to 5.6 million virtual CPUs (vCPUs).
Experimental validation against PARP1 and the more challenging FSP1 target produced nanomolar inhibitors, with researchers also resolving co-crystal structures of FSP1 bound to small-molecule inhibitors.
🧬 From Brute-Force Docking to Adaptive Screening #
Traditional virtual screening treats an enormous compound library largely as a flat search space.
If the library contains tens of billions of molecules, screening the entire collection requires enormous amounts of docking computation. Increasing the library size therefore tends to increase computational requirements approximately proportionally.
AdaptiveFlow changes the strategy by using chemical-space organization and machine learning to determine which molecules deserve expensive docking calculations.
The high-level workflow can be summarized as:
[69 Billion Compounds]
│
▼
[18D Physicochemical Grid]
│
▼
[Diverse Prescreening]
│
▼
[ML Classification]
│
▼
[High-Affinity Chemical Regions]
│
▼
[Targeted Deep Screening]
│
▼
[Experimental Validation]
The key idea is not to make every docking calculation faster.
Instead, AdaptiveFlow tries to avoid unnecessary docking calculations altogether.
⚙️ AdaptiveFlow’s Core Architecture #
The platform combines large-scale chemical-library preprocessing, structured chemical-space sampling, machine learning, and heterogeneous docking methods.
AdaptiveFlow Ligand Preparation #
The AdaptiveFlow Ligand Preparation (AFLP) component prepares enormous molecular libraries for downstream screening.
Its responsibilities include:
- Enumerating stereoisomers
- Enumerating tautomers
- Validating three-dimensional conformers
- Calculating molecular descriptors
- Preparing compounds for large-scale docking workflows
The system calculates 28 molecular descriptors for each prepared compound.
At this scale, preprocessing itself becomes an engineering challenge. The objective is to transform an enormous raw chemical collection into a representation that can be efficiently searched, partitioned, and processed across distributed infrastructure.
18-Dimensional Physicochemical Grid #
One of AdaptiveFlow’s central ideas is its 18-dimensional physicochemical grid.
Instead of treating 69 billion molecules as an unstructured list, the platform maps compounds according to properties such as:
- Molecular weight
- Polar surface area
- Lipophilicity
- Other physicochemical descriptors
The resulting chemical space is divided into more than 12 million discrete tranches.
This representation provides a structured way to sample chemical diversity.
Rather than selecting molecules randomly from the entire library, AdaptiveFlow can identify representative compounds from different regions of the grid and use those molecules to learn which regions are more promising for the target under investigation.
🤖 Machine Learning and Active Learning #
The machine-learning component is what turns chemical-space organization into an adaptive screening process.
The initial screening stage evaluates representative compounds from diverse regions of the physicochemical grid.
Those results become training data for a target-specific classifier.
Initial Diverse Sampling
│
▼
Experimental / Docking
Prescreen Results
│
▼
ML Classification Model
│
▼
Identify Promising Regions
│
▼
Allocate More Compute
│
▼
Deeper Screening
│
└──────> Update Model
│
└──> Repeat
This creates an active-learning loop.
The system does not need to know the ideal chemical characteristics of a target in advance. Instead, it learns from the results generated during the screening process and progressively focuses computational resources on chemical regions associated with stronger predicted binding.
That adaptive behavior is the foundation of the claimed cost reduction.
🔬 Multi-Protocol Docking with AFVS #
AdaptiveFlow’s virtual-screening engine, AFVS, supports more than 1,500 distinct docking protocols.
This is important because molecular docking is not a single universal algorithm. Different targets and ligand classes can benefit from different representations, scoring functions, sampling strategies, and computational implementations.
The platform combines:
- Classical physics-based docking
- GPU-accelerated methods
- Machine-learning-based approaches
- Diffusion-based docking
- Protein-ligand screening
- DNA and RNA screening
- Peptide screening
Machine-learning methods such as DiffDock and TANKBind can therefore participate alongside more traditional physics-based approaches.
The result is a heterogeneous screening framework rather than a workflow tied to a single docking engine.
☁️ Scaling Ultra-Large Screening on Cloud Infrastructure #
Screening tens of billions of compounds is not only a chemistry problem. It is also a distributed-computing problem.
AdaptiveFlow was designed to execute across large cloud infrastructures and demonstrated near-linear scaling to 5.6 million vCPUs on AWS.
A simplified execution model looks like:
AdaptiveFlow Controller
│
┌──────────────┼──────────────┐
▼ ▼ ▼
Worker Pool Worker Pool Worker Pool
│ │ │
Docking Jobs Docking Jobs Docking Jobs
│ │ │
└──────────────┼──────────────┘
▼
Results Aggregation
│
▼
ML Re-ranking
At this scale, infrastructure efficiency becomes as important as algorithmic efficiency.
AdaptiveFlow combines adaptive sampling with distributed execution and fault-tolerant cloud infrastructure. The result is intended to make ultra-large screening practical without requiring every research group to operate a dedicated supercomputing facility.
Cost Comparison #
| Metric | Traditional Brute-Force ULVS | AdaptiveFlow |
|---|---|---|
| Library Size | 69 billion compounds | 69 billion compounds |
| Search Strategy | Exhaustive docking | Grid-guided adaptive sampling |
| Relative Compute Cost | 100% baseline | ~0.1% |
| Reported Cost Reduction | — | ~1,000x |
| Scaling | Increasing communication and compute overhead | Near-linear to 5.6M vCPUs |
| Infrastructure | Large HPC resources | Cloud infrastructure / Spot Instances |
| Docking Methods | Typically limited by workflow | 1,500+ protocols |
The important distinction is that AdaptiveFlow does not eliminate the 69-billion-compound search space.
It changes how that search space is explored.
💰 Turning a Multi-Million-Dollar Workflow into an Accessible Pipeline #
The economics of ultra-large virtual screening have historically limited who can perform these experiments.
An exhaustive screen across tens of billions of compounds can require enormous compute resources. Even when cloud infrastructure is available, paying for every docking calculation remains impractical for many academic laboratories.
AdaptiveFlow instead attempts to make the workflow selective:
69 Billion Molecules
│
▼
Chemical-Space Organization
│
▼
Representative Sampling
│
▼
Initial Screening
│
▼
Machine-Learning Model
│
▼
Promising Sub-Libraries
│
▼
Expensive Docking
│
▼
Experimental Testing
By allocating expensive computation selectively, the platform reportedly reduces computational requirements by approximately three orders of magnitude.
This is the central engineering contribution: the platform treats screening strategy itself as an optimization problem.
🧪 Experimental Validation with PARP1 and FSP1 #
A major concern with aggressive computational filtering is that the algorithm could eliminate genuine active molecules before they ever reach experimental testing.
The researchers therefore validated AdaptiveFlow against two oncology-related targets with different levels of computational difficulty.
PARP1 #
Poly(ADP-ribose) Polymerase 1 (PARP1) is a well-characterized cancer target with substantial existing structural and pharmacological knowledge.
AdaptiveFlow identified nanomolar-potency inhibitors against PARP1, demonstrating that the adaptive screening strategy can recover high-affinity compounds without requiring exhaustive docking of the entire library.
FSP1 #
The more challenging validation target was Ferroptosis Suppressor Protein 1 (FSP1).
FSP1 has attracted interest because of its role in cancer-cell survival and ferroptosis regulation, but its binding environment creates additional computational challenges.
The binding pocket involves cofactors and associated molecular components, including:
- NAD+
- FAD
- Coenzyme Q10
These additional structural factors make conventional computational screening more difficult.
AdaptiveFlow nevertheless identified high-affinity FSP1 inhibitors.
More significantly, the researchers resolved co-crystal structures of FSP1 bound to small-molecule inhibitors, providing structural evidence that the computationally identified compounds interact with the intended target.
🧩 Why FSP1 Is an Important Demonstration #
The FSP1 results matter beyond a single target.
Ultra-large virtual screening is most valuable when it can be applied to targets where conventional drug-discovery approaches struggle.
A platform that works only on well-characterized binding pockets would have limited impact. Challenging targets such as FSP1 provide a stronger test of whether adaptive computational screening can discover chemically useful molecules in difficult structural environments.
The combination of:
Ultra-Large Chemical Library
+
Adaptive ML Screening
+
Multiple Docking Strategies
+
Experimental Structural Validation
provides a more complete demonstration than computational ranking alone.
🌍 Open Source and Democratization of Drug Discovery #
AdaptiveFlow has been released as a free, fully open-source platform.
The release includes supporting materials intended to make large-scale screening easier to reproduce, including:
- Step-by-step workflows
- Preprocessed compound libraries
- Cloud deployment scripts
- AWS-oriented execution infrastructure
- Screening and analysis components
This open approach is significant because the barrier to ultra-large screening is not purely scientific.
Researchers need access to:
- Large compound libraries
- Molecular preprocessing pipelines
- High-throughput docking software
- Distributed compute infrastructure
- Machine-learning models
- Workflow orchestration
- Methods for analyzing enormous result sets
Providing these pieces through a single open platform can substantially reduce the engineering effort required to perform large-scale virtual screening.
📈 From Billion-Scale Screening to Routine Discovery #
The broader implication of AdaptiveFlow is that ultra-large virtual screening may no longer need to mean brute-force computation.
Traditional scaling assumes:
More Compounds
│
▼
More Docking
│
▼
More Compute
│
▼
Higher Cost
AdaptiveFlow instead attempts to introduce intelligence into the scaling relationship:
More Compounds
│
▼
Better Chemical-Space Representation
│
▼
Adaptive Sampling
│
▼
Target-Specific Learning
│
▼
Selective Deep Screening
│
▼
Controlled Compute Growth
That distinction becomes increasingly important as commercial and academic compound libraries continue to expand.
The challenge is no longer simply whether researchers can generate enough candidate molecules. It is whether they can search enormous chemical spaces efficiently enough to identify useful candidates.
🏁 Conclusion #
AdaptiveFlow represents a shift from exhaustive molecular docking toward AI-guided, adaptive ultra-large virtual screening.
By combining an 18-dimensional chemical-space representation, diverse prescreening, active learning, machine-learning classification, and more than 1,500 docking protocols, the platform can selectively direct expensive computation toward the most promising regions of a 69-billion-compound library.
Its reported results are particularly notable for three reasons:
- Approximately 1,000x lower computational cost than brute-force ULVS
- Near-linear scaling to 5.6 million vCPUs
- Experimental identification and structural validation of inhibitors targeting PARP1 and FSP1
The open-source release further lowers the infrastructure barrier for researchers who want to explore ultra-large chemical libraries.
The fundamental idea is simple but powerful:
Don't Dock Everything.
Learn Where to Look.
Then Spend Compute Where It Matters.
If the approach generalizes across targets and chemical spaces, AdaptiveFlow could make billion- and tens-of-billions-scale virtual screening a practical component of routine computational drug discovery rather than a workflow reserved for organizations with access to exceptional computing resources.