“We gave Nextflow Frontier one of the hardest problems in our field: predicting how mutations change protein binding affinity. On our own held-out data, the model it built outperformed every published method we tested, including recent deep learning approaches.”
BioML-bench · 22 tasks
Breakthrough benchmark performance
Nextflow Frontier outscored Rosalind and Claude Science, ranking highest in all four areas by finding the best answer, not the first one.
| Entry | Best PR-AUC | Run |
|---|---|---|
| Nextflow Frontier | 0.887 | 820 attempts, 80 h |
| MolGPS entry | 0.866 | Reference entry |
| Rosalind | 0.864 | 51 attempts, last at 2 min |
| Claude Science | 0.821 | 2 attempts, last at 1 min |
One task up close: ranking 144 molecules by how likely they are to bind EGFR, a protein behind some lung cancers. Rosalind stopped at 0.864 after 51 attempts, and Claude Science at 0.821 after 2. Nextflow Frontier kept experimenting, changed how it described each molecule, and reached 0.887, surpassing the previous leaderboard best.
Read the results#1 Benchmark
Drug Discovery
We built two simple, readable chemistry models and combined them to rank molecules by how likely they are to bind EGFR, a protein involved in cancer growth.
Read the results#2 Benchmark
Biomedical Imaging
We combined CT measurements with clinical information to forecast lung function and prediction uncertainty in pulmonary fibrosis.
Read the results#3 Benchmark
Protein Engineering
We adapted a protein language model to rank insertions and deletions in a mouse CSN4 domain by experimentally inferred folding stability.
Read the results#4 Benchmark
Single-Cell Omics
We trained boosted decision trees on the genes that vary most between cells to name cell types in kidney single-nucleus RNA profiles.
Nextflow Frontier best | Competing method best | |||
|---|---|---|---|---|
 Naming kidney cell types | Single-cell omics | 0.977 weighted F1 | 0.963, XGBoost | Ahead, beyond margin of error |
| Predicting protein levels from RNA | Single-cell omics | 0.604 RMSE (lower is better) | 0.724, KNNR (Py) | Ahead, beyond margin of error |
| Forecasting lung-function decline | Biomedical imaging | −6.74 log-likelihood (higher is better) | −7.10, retrained competition-winning method, same patients | Ahead, beyond margin of error |
| Ranking compounds for EGFR activity | Drug discovery | 0.887 PR-AUC | 0.866, 1B_MPNN_MolGPS-ens_LargeMix | Within margin of error |
| Predicting how fast the liver clears a drug | Drug discovery | 0.796 Pearson r | 0.787, AceProp1 | Within margin of error |
How it works
A Recursive, Closed-loop Discovery System
Scientific evaluation at every stage
Adaptive Discovery Model: Intelligent optimization & orchestration of every step
Hypothesis
- Planning and Design: Design experiment, select data, models, and tools. Craft novel methods
- Execution: Run thousands of pipelines in parallel on your compute
- Validation: Evaluate outcomes and assess confidence
- Learning: Analyze results to inform the next round of experiments
Lab
Scientists guide the loop with full visibility, control and auditability
Customer Control
Discovery, on your terms
Unlock the power of your data. Zero egress or training on your data.
Guide the loop with full visibility. Every hypothesis, test, and result, traceable and interpretable in minutes.
Integrate all your complex systems across organizations with 160+ proprietary connectors.
Run inside your environment and security perimeter. SOC 2 Type II, 21 CFR Part 11-aligned audit logging.
Work within your organization’s governance patterns. Every action fully auditable.
Own the custom models and methods we deliver.
The Stack
The technology behind the intelligence
Data
- Internal data
- Public data
- Program traces
Tools and models
- 2000+ tool registry
- SOTA models, MCPs
- Proprietary and open LLMs
Connectors:Proprietary connectors integrate into any custom environment
Control Panel
Advanced steering and observability system for scientists
Research Harness
Technical orchestration and context engine
Adaptive Discovery Model
Underlying, self-learning intelligence layer optimizing and orchestrating the discovery loop
Proprietary Agentic Nextflow
Custom Nextflow dialect designed for massively parallel agentic runs
Nextflow Platformby
- Compute
- Governance
- Security
- Auditability
- Provenance
Connectors:Proprietary connectors integrate into any custom environment
We deliver novel models and methods
Designed to be problem-agnostic, Nextflow Frontier generates solutions in any computational area where success can be measured.
Target Discovery & Validation
Structure Prediction
Ensemble Sampling
Epitope Modeling
De novo Design
Hit2Lead Optimization
Binding Affinity
Virtual Cell
ADME & Toxicity
PK & PD Modeling
