Preview

Physics-Based Binding Affinity Predictions That Actually Work

Published FEP benchmarks typically require weeks of expert preparation. Deep Origin runs end-to-end with zero manual intervention — achieving sub-1 kcal/mol accuracy on half of all tested targets.

60% ΔG error <1 kcal/mol (FEP+ benchmark)

90% ΔG error <2 kcal/mol (FEP+ benchmark)

Zero Manual pose curation

Book a Demo

The Elephant in the Room

What actually goes into "<1 kcal/mol accuracy"?

The Hidden Work Behind Published FEP Benchmarks

Most published FEP benchmark results reflect careful, expert-driven workflows. Achieving those numbers typically requires significant manual intervention:

Careful protein structure selection and refinement
Expert-guided ligand pose curation
Protonation state assignment
Co-crystal structure alignment
Bespoke perturbation network design

The poses fed into published FEP benchmarks are typically the product of considerable human effort — iterative refinement that is difficult to reproduce and impossible to automate.

The Deep Origin Difference

Our benchmark numbers reflect what happens when you point DO Studio at a protein, run docking, and immediately pipe those poses into ABFE — no curation, no manual intervention, no hidden steps.

Published benchmarks = best-case output of expert-driven processes.
Deep Origin's numbers = fully automated, zero-touch pipeline.

The Data

DO ABFE Benchmark Results

Benchmarked against the largest published FEP dataset (Ross et al. 2023, 1,237 compounds across 20 targets). All results reflect a fully automated pipeline — zero manual curation.

20-Target Overview 301 Compounds

Overall performance across the full Ross et al. benchmark set. The experimental reproducibility floor is 0.91 kcal/mol RMSE — results approaching this limit represent the ceiling of what any computational method can achieve.

Overall RMSE 1.31

Overall MAE 0.98

Within 1 kcal/mol 62.9%

Within 2 kcal/mol 91.6%

Sub-1 RMSE Targets 10/20

Half of all targets achieve sub-1 kcal/mol RMSE — approaching the experimental reproducibility floor with a fully automated workflow.

End-to-End Results DO Dock + DO ABFE

Complete Deep Origin pipeline — from raw protein structures through automated docking to final ABFE predictions. No external tools, no manual pose preparation.

TYK2 (n=14)

RMSE: 0.55 MAE: 0.43 R²: 0.78 Within 1 kcal/mol: 93%

Thrombin (n=10)

RMSE: 0.65 MAE: 0.53 R²: 0.62 Within 1 kcal/mol: 80%

JAK2 (Set 1) (n=9)

RMSE: 0.69 MAE: 0.59 R²: 0.59 Within 1 kcal/mol: 89%

Key result: All three end-to-end targets fall well below the published FEP+ benchmark average of 1.25 kcal/mol RMSE and approach the experimental reproducibility floor of 0.91 — with zero manual intervention.

Top 10 Targets (Sub-1 kcal/mol RMSE)

Target	N	RMSE	MAE	R²	% <1 kcal/mol	% <2 kcal/mol
JAK2 (Set 1)	12	0.64	0.57	0.84	91.7%	100%
Thrombin (Core)	7	0.72	0.63	0.46	85.7%	100%
HSP90 (Single Ring)	7	0.84	0.77	0.01	71.4%	100%
Fragment (Liga)	10	0.85	0.78	0.79	60.0%	100%
cMET	11	0.87	0.68	0.59	72.7%	100%
CDK2	13	0.96	0.72	0.25	69.2%	100%
T4 Lysozyme	9	0.89	0.80	0.54	66.7%	100%
HSP90 (2-Ring)	6	0.97	0.70	0.19	66.7%	100%
TYK2	15	0.95	0.79	0.47	66.7%	100%
TAF12	8	1.01	0.89	0.19	62.5%	100%

It's not just about accuracy — it's about what it takes to achieve that accuracy. Deep Origin delivers these results with a fully automated pipeline that any scientist can run in minutes.

What Makes Our FEP Different

FEP calculations are challenging to set up and run correctly, with many pitfalls. We've solved the hard problems so you can focus on drug discovery.

Key Innovation

Separated Topologies (SepTop)

Unlike hybrid topology methods used by most FEP codes, our RBFE implementation treats each ligand as entirely independent. This enables calculations on both congeneric series and structurally diverse molecules, including scaffold hops.

Key Innovation

Novel Restraint Formulation

Industry-standard Boresch restraints require heuristics that can fail and need separate MD simulations. Our restraint formulation is selected automatically, is robust across diverse targets, and requires no additional simulation.

End-to-End Solution

From initial protein and ligand files, through docking and system preparation, to final ABFE and RBFE results — all without manually orchestrating separate tools at each stage.

Active Data & Compute Management

We handle all file parsing and data management for you. No need to wrangle AWS or cloud infrastructure — we take care of compute orchestration automatically.

Three Ways to Access FEP

Whether you prefer a visual interface, programmatic control, or conversational AI — we're building FEP for how you work.

Coming Soon

DO Studio

Our new unified interface for medicinal chemists and drug discovery scientists. Run FEP workflows through an intuitive UI — no command line required.

Try DO Studio Today

Available Now

Deep Origin API

Programmatic access via Python client. Full control for computational scientists who want to integrate FEP into their existing pipelines.

Request Access

Coming Soon

FEP in Balto AI

Run FEP calculations through natural language conversations. Ask questions, get results — no coding or specialized UI knowledge needed.

Coming Soon

Simple, Transparent Pricing

Self-sign up, pay as you go. Try the product without needing to pay large upfront license fees.

Self-Serve

Pay-as-you-go access for individual researchers and small teams

Enterprise

Custom solutions for organizations with high-volume needs

Partnerships

Collaborative drug discovery programs with dedicated support

Detailed pricing available upon joining the waitlist.

Get Started

Request an FEP Demo

Tell us about your project and our team will help guide you to the right FEP solution—whether that's ABFE, RBFE, or a custom workflow tailored to your targets.

We'll reach out within 24 hours to schedule your personalized demo.