Benchmarks

Verified biotech compute benchmarks on AWS.

Every row is a reproducible run on a public reference dataset, with real spot-pricing dollar figures and quality metrics (DockQ, MUE, TM-score, F1, FSC, FDR — the ones reviewers actually care about). Click any row for the full methodology blog post: input data, commands, container digests, output checksums.

Workload
Cost
Dataset · Headline metric
AlphaDIA DIA proteomics cohort
CPU spot (c6i, 16 vCPU) · Mann lab tutorial library
$0.05 per sample · 13 min wall
PXDOrbitrap Astral HeLa 200 ng · 8,591 proteins, 105,414 precursors per sample
OpenFold3 TCR-pMHC ternary complex co-fold
1× A10G (g5.xlarge spot) · 5-chain assembly
$0.13 per complex · 16 min
PDB11 TCR-pMHC reference + post-cutoff · DockQ 0.59 / 0.62 · 0/11 incorrect
OpenFE RBFE relative binding free energy · TYK2 9-edge
1× A10G spot per edge · OpenFE 1.11.1 · 5 ns production, 11 λ-windows
$33 n=1 sanity · ~$100 at n=3
CANONICALTYK2 10-ligand set · MUE 0.38 kcal/mol vs experiment · 9 edges, 18 legs
EGFR hit-discovery funnel · fold → dock → rescore → MD
1× L40s spot · OpenFold3 + Uni-Dock + Boltz-2 + GROMACS 1 ns MD
$1.16 total · ~73 min, 1,000 cpds
ChEMBLEGFR ATP site · CHEMBL203 top 1,000 actives · OpenFold3 TM 0.984 vs crystal
AAV2 capsid-receptor co-fold
1× L40s (g6e.4xlarge spot) · VP3 trimer + AAVR (PKD2)
$0.21 per variant · 13 min
PDB6IHB (AAV2) + 9J6Z (AAV8, post-cutoff) · TM 0.984 / 0.988
30× WGS variant calling · Sarek + Parabricks
1× A10G (g5.8xlarge) + CPU spot · Parabricks BWA + DeepVariant
$2.59 per sample · 2h 32m
GIABNA12878 · SNP F1 0.9966 / INDEL F1 0.9934 vs hap.py truth set
Tn-O-glycoproteomics ADC target discovery
CPU + L4 + L40s spot · O-Pair + Casanovo + Boltz-2 co-fold
$1.28 total · 1h 40m (5-step chain)
PXD011063Vakhrushev SimpleCell · 253 O-glycopeptides · FGFR4 + DLK1 surface as ADC candidates
MAPPs neoantigen discovery · WES + LC-MS/MS chain
Mixed GPU/CPU spot · Comet + Casanovo + Percolator
$0.14 calibration · ~30 min chain
PXD019643HLA-eluted prostate · Casanovo 62 PSMs vs Comet 16 at 1% FDR · 5,367 PSMs on real run
Mel15 neoantigen positive-control rediscovery
16× CPU spot nodes parallel · Comet + Percolator
$7 total · ~50 min, 16 parallel
PXD004894Mel15 HLA-I · 7/8 published ligands recovered · all 3 T-cell-validated recovered
RELION cryo-EM 3D classification
1× A10G (g5.xlarge) · 4× A10G (g5.12xlarge) for 1h 36m alt
$1.65 per run · 3h 18m on 1 GPU
EMPIAR-10028Plasmodium 80S ribosome · 105,247 particles · rlnCurrentResolution 9.65 Å, 6 classes
Cryo-ET subtomogram averaging · immature HIV Gag
4× T4 (g4dn.12xlarge spot) · AreTomo2 + RELION-5
$2.20 total · 1h 40m end-to-end
EMPIAR-10164HIV-1 VLP Gag · 3.99 Å FSC resolution (C6 symmetry) · AreTomo2 3 min on 1× T4
GROMACS T-REMD multi-GPU replica ensemble
4× A10G (g5.12xlarge spot) or 1× A10G single-GPU
$0.01 per microsecond (extrapolated)
REFTrp-cage (8,307 atoms) · 1,547 ns/day single-GPU · 8-replica T-REMD, 53% acceptance

Costs are AWS spot, us-east-1, post-eviction checkpoint-restart enabled, billed to your account — not ours. Numbers are the headline figure from each blog post. Hardware listed is the verified configuration; instance family is selected automatically by Karpenter from the GRES budget you set.

Read the full methodology blogs →