Skip to main content
AiDDA·Benchmarks

From target to optimised portfolio — weeks to hours

Transparent validation data across 14 DUD-E targets, 7 protein families, and 42 real campaigns. Every number is reproducible from the same natural-language input.

0
DUD-E targets validated
0
Protein families tested
0
Real campaigns measured
0
Hours average runtime
The real timeline problem

A campaign isn't one step.
It's a cycle.

Here's what changes when the entire design-make-test-analyse loop runs autonomously.

Traditional6–8 weeks
AiDDA1–3 hours

~40× time compression

Traditional workflow
Wk 1

Target research, PDB selection, protein prep, compound library assembly

Wk 2

Docking, scoring, manual hit triage, ADMET profiling

Wk 3

Medicinal chemistry reviews SAR, proposes modifications

Wk 4

Second-round library, re-dock, re-score, re-triage

Wk 5–6

MM-GBSA or FEP on selected compounds. Waiting for compute.

Wk 6–8

Final candidate selection, report, handoff to synthesis

6–8 weeks · no guarantee of convergence
AiDDA
Hr 1

Target research → protein prep → library → docking → ADMET → validation → SAR → tiered portfolio

Hr 2–3

Autonomous evaluation. Generates new compounds if needed. Repeats until convergence.

Hours · same rigour · automated iterations
Iterative optimisation

The hard part, automated

The first screen is easy. The real work starts when you ask: “How do I make these hits better?”

Traditional MPO

Medicinal chemist proposes changes
Computational chemist re-docks
ADMET scientist flags liabilities
Team debates in a meeting
Repeat for each design idea
One iteration per week, if fast
ScreenEvaluateDiagnoseGenerateRe-screen

Each iteration: minutes, not weeks

01

Evaluate portfolio

Qualified hits, scaffold diversity, ADMET pass rate

02

Diagnose bottleneck

Binding? ADMET? Diversity? Auto-detected.

03

Generate compounds

Scaffold hops, bioisosteric swaps, R-group enumeration

04

Full pipeline screen

Dock → score → qualify → ADMET → validate

05

Merge & compare

SAR accumulates. Tracks what improved.

06

Stop or continue

Converged → stop. Improving → another round.

Balanced scoring

What gets optimised

The platform balances 8 objectives simultaneously — the same trade-offs a medicinal chemist navigates manually, resolved in every iteration.

Improving binding often breaks ADMET. Fixing ADMET drops binding. This whack-a-mole takes weeks manually. The platform's multi-parameter scoring resolves these trade-offs in every iteration, tracking what improved and what regressed.

Binding strengthPose qualityDrug-likenessSynthetic accessibilityADMETNovelty3D characterScaffold diversity
Figure 1

DUD-E benchmark results

Complete automated workflow per target. AUC-ROC and Precision@20 on initial screening pass.

Nuclear ReceptorKinasePolymeraseProteaseGPCRChaperoneHydrolase
ERαNuclear Receptor
0.95
20/20
100%
SRCKinase
0.95
20/20
100%
VEGFR2Kinase
0.92
20/20
100%
HIV-RTPolymerase
0.88
15/20
75%
PPARγNuclear Receptor
0.88
10/20
50%
BACE1Protease
0.87
14/20
70%
CDK2Kinase
0.87
16/20
80%
A2AGPCR
0.84
9/20
45%
HSP90Chaperone
0.82
9/20
45%
EGFRKinase
0.79
10/20
50%
ACEProtease
0.72
11/20
55%
ABL1Kinase
0.66
2/20
10%
AmpCHydrolase
0.61
4/20
20%
DRD3GPCR
0.56
3/20
15%
Median0.8510.5/2053%

Initial screening pass. In real campaigns, iterative optimisation generates additional compounds targeting weaknesses — improving coverage in subsequent rounds.

Deliverables

What a campaign produces

After the final iteration, a complete actionable package.

Tiered compound portfolio

Typical distribution
Tier 1: Top candidates

Strong binding, ADMET-clean, physics-validated, diverse scaffolds

25%
Tier 2: Promising with flags

Good binding, specific liabilities identified

40%
Tier 3: Backup / SAR context

Moderate binding, valuable for structure-activity understanding

35%

Iteration history

  • What was tried and results
  • Per-round metrics & trends
  • Generation strategy rationale

SAR analysis

  • Matched molecular pairs
  • Scaffold-activity relationships
  • What helps, what hurts

Per-compound profiles

  • Binding mode & residue contacts
  • ADMET predictions
  • Synthetic accessibility
CSVScores, SMILES, scaffolds, ADMETSDF3D docked posesPDFStructured report
Transparency

Where we're honest

The pipeline underperforms on 3 of 14 targets. We show these results because transparency is how trust gets built.

Virtual screening is not solved. No tool gets every target right. What matters is automating the 90% that doesn't require expert judgment — and being transparent about the 10% that still does.

ABL1
0.66

Non-canonical DFG-out binding conformation. CNN-based pose scoring was trained on canonical modes and misjudges binding quality here.

AmpC
0.61

Fragment-like actives (avg MW ~295) fall below the molecular weight range where CNN scoring is most reliable.

DRD3
0.56

Deep transmembrane GPCR pocket. Structure-based virtual screening remains challenging for this target class generally.

The platform detects scoring unreliability from score distributions and flags results rather than presenting false-confident rankings.

Methodology

DUD-E (Database of Useful Decoys — Enhanced), 14 targets, ~500 compounds per target. Full automated pipeline: literature review, PDB retrieval, protein preparation, compound library construction, 3D conformer generation, CNN-based docking, multi-signal scoring, adaptive qualification, scaffold-diverse selection, ADMET screening, physics-based binding validation. AUC-ROC and Precision@20 on initial screening pass. Campaign times across 42 real campaigns. All benchmarks reproducible from same natural-language input. March 2026.

See these results on your target

AiDDA runs the full virtual screening pipeline — from target to ranked portfolio — in hours, not months. Run a campaign on your protein target and compare against your current workflow.

No credit card required