Target-adaptive compute allocation for de novo binder design

From one-shot generation to an evidence-guided search campaign.

T-ReX reallocates a matched worker-GPU-hour budget across generator families, rescue operations and exploration routes. This page lets you inspect throughput, diversity, candidate quality and the evidence-linked decisions behind the campaign.

Explore the benchmarkOpen a generated complex
Publication snapshot: 21 July 2026. Endpoint tables use the audited six-target TM0.6 benchmark; replay curves cover all six targets and the five archived method arms in the reconciled publication curve table.Not a live scheduler monitor

What the benchmark shows

The primary endpoint is structure-unique strict success (SU) under a common canonical criterion. Sequence diversity and independent confidence scores are shown separately so the dimensions are not conflated.

6benchmark targets
144worker-GPU-h per arm
1,222T-ReX TM0.6 SU
3,799T-ReX unique sequences
1.22×T-ReX / PUCT total SU

Compare methods and targets

Choose a target and endpoint. Counts are evaluated at the nominal 144 worker-GPU-hour endpoint; ipTM-qualified counts are counted over all scored SU representatives.

CD45: Foldseek SUT-ReX 301 · PUCT 283 · Complexa-only 45 · BindCraft-only 2 · BoltzGen-only 5

Target summary

T-ReX SU
301
PUCT SU
283
Best official arm
45
T-ReX sequence clusters
808
T-ReX SU/GPU-h
2.090

Default static view: CD45. Use the controls to filter targets and endpoints.

MethodStrict successesFoldseek SUUnique sequencesipTM ≥ 0.7ipTM ≥ 0.8SU/GPU-h
T-ReX1,9363018082971532.090
PUCT1,8512837922801481.965
Complexa-only6474513544200.312
BindCraft-only1322210.014
BoltzGen-only22522520.035

Campaign replay

Replay reconciled TM0.6 cumulative-SU trajectories by target and method. Curves are downsampled from append-only campaign records; the x-axis is worker-GPU-hours and excludes controller/LLM overhead.

CD45: cumulative strict SUArchived TM0.6 curves are loaded after initialization.
Select a target and method to replay the archived campaign trajectories.

Evidence-linked decision traces

Intermediate audit cards show how T-ReX converted structured evidence into the next exploit, rescue or explore action as SU accumulated. These are public supervisor summaries from the saved logs, not private chain-of-thought and not an exhaustive dump of every tick.

CD45 · first SU feedback

example intermediate stage

exploit
Evidence
Live SU increased after early generation, triggering a higher exploit fraction.
LLM audit rationale
The supervisor links route yield, duplicate pressure and strict-gate margins to the next action.
Selected next job
Continue the productive route while keeping rescue/explore capacity.

SC2RBD · hard-target pivot

example intermediate stage

explore
Evidence
Cheap direct-scoring routes were weak, so the supervisor tested a slower BindCraft route.
LLM audit rationale
Hard-target evidence justified waiting for delayed interface-optimization signal instead of stopping immediately.
Selected next job
Probe a different generator/configuration and monitor for strict SU.

Generated complexes

The default view is a six-target T-ReX nomination panel. Each candidate is a strict-success Foldseek-TM0.6 representative ranked first within the SU top-10 by the balanced joint score (strict margins + AF2 ipTM + AF2 ipSAE). Physics/affinity values are in-silico post-hoc predictions, not experimental measurements.

Loading the molecular viewer…
If opened directly from a file URL, serve this directory over HTTP so the PDB asset can be fetched.

How to read this page

The endpoint hierarchy mirrors the paper: generation throughput first, independent quality second, candidate nomination last.

Primary endpoint

Strict successpLDDT ≥ 90, iPAE ≤ 7/31 (0.2258), and binder scRMSD < 1.5 Å.
Structure-unique (SU)Strict successes clustered by Foldseek on the binder chain at TM0.6.
Unique sequencesAll strict successes clustered with MMseqs2 at the sequence-identity threshold; this does not replace SU.
Independent scoresAF2 ipTM/ipSAE/actifpTM and other metrics are post-hoc diagnostics, not SU credit.

Fair-comparison rules

Method-only baselinesComplexa, BindCraft and BoltzGen use their official default configurations.
Controller baselinesPUCT and ε-greedy use the same registered action space and execution stack as T-ReX; only proposal/allocation logic differs.
Compute accountingRates use worker GPU-hours. T-ReX LLM/controller GPU time is not added to the worker denominator.
Top-K interpretationTop-10/50/100 panels are selected after strict filtering and structure clustering; threshold counts use the full scored SU pool.

Data provenance

The dashboard keeps endpoint counts, campaign replay and candidate nomination as separate auditable layers. Each layer is rendered only from a reconciled artifact; the page does not claim to be a live scheduler monitor.

LayerScopeSource used here
Endpoint comparisonSix targets · TM0.6 · 144 worker-GPU-h nominal endpointPublication endpoint table
Campaign replaySix targets · T-ReX, PUCT and official-default armsReconciled cumulative scaling-curve CSV
Candidate panelSix T-ReX candidates plus two contextual referencesStrict-success SU top-10 balanced-rank records and copied PDBs
Interactive demo snapshot assembled from the archived TM0.6 endpoint, MMseqs2 sequence-throughput table, controller EvidenceSummary checkpoints and three copied PDB assets. Original experiment directories, logs and generated samples remain unchanged.