2026-09-25
T-REX uses two local 27B LLMs to orchestrate binder-design tools. On seven targets at 144 GPU-hours, unique-hit throughput is 2.48x the best generator and 2.43x PUCT.
The protein-binder toolbox is no longer the scarce part. Proteína-Complexa, BindCraft and BoltzGen can propose backbones, ProteinMPNN can rewrite sequences, and AlphaFold2 can score them. What is scarce is a way to spend a finite GPU budget across those tools as a campaign unfolds. Which generator works on this target, which hyperparameters to keep, and which near-miss is worth a rescue attempt are not known in advance, and they change. A scalar hit count also hides the failure mode: no output, still waiting on evaluation, a few thresholds short, or a structural duplicate. Those call for different next jobs.
Ellen D. Zhong's group at Princeton CS treats the campaign as an online allocation problem. Each target gets 144 H100 worker GPU-hours (four GPUs for 48 hours). The objective is structurally unique hits. Jinyeop Song at MIT Physics is a co-author.
T-REX stands for Target-adaptive Rescue, Explore, eXploit. The shared action space has eight families: four Complexa search variants (beam, best-of-n, Feynman-Kac steering, MCTS), BindCraft, BoltzGen, ProteinMPNN redesign, and AF2 evaluation. A design counts as a hit only if binder pLDDT is at least 90, normalized iPAE is at most 7/31, and binder scRMSD is below 1.5 Å, the Complexa protocol. Qualified backbones are clustered with Foldseek at TM-score 0.6; one cluster is one SU.
Raw logs never go straight into the model. Deterministic code first builds an EvidenceSummary covering structure quality, novelty, per-route yield and cost, the evaluation queue, parent lineages and GPU occupancy, then labels one of seven campaign states (low evidence, duplicate collapse, productive, productive with duplication, near-miss-enriched, stalled, deep stall). A Planner LLM writes evidence-linked hypotheses and next tests. A Check + Build layer validates them. A Supervisor LLM tags each job as Rescue, Explore or eXploit, ranks them, and proposes a budget mix. A deterministic controller actually fills GPUs, asynchronously, as soon as a worker frees up. Both LLMs are local Qwen3.6-27B-FP8; LLM inference is excluded from the 144 hours.
Rescue repairs a promising near-miss. Explore tries an underused method or setting. eXploit puts more compute on routes that still produce new structures. PUCT and epsilon-greedy share the same action space but allocate from a scalar reward, with no contextual LLM reasoning.
Seven targets come from the Complexa benchmark: CD45, BetV1, CbAgo, HER2-AAV, SC2RBD, PDL1, IL7RA. T-REX produced 1,461 TM0.6 SUs and led every target against five controls.
| Target | T-REX | PUCT | Best generator-only | vs PUCT | vs best GO |
| CD45 | 2.090 | 0.479 | 0.347 (CX) | 4.36× | 6.02× |
| BetV1 | 0.444 | 0.236 | 0.118 (CX) | 1.88× | 3.76× |
| CbAgo | 0.806 | 0.375 | 0.361 (BC) | 2.15× | 2.23× |
| HER2-AAV | 1.417 | 0.361 | 0.722 (BC) | 3.92× | 1.96× |
| SC2RBD | 0.382 | 0.278 | 0.153 (BC) | 1.38× | 2.50× |
| PDL1 | 3.347 | 1.618 | 1.951 (CX) | 2.07× | 1.72× |
| IL7RA | 1.660 | 0.660 | 1.236 (BC) | 2.52× | 1.34× |
Units are SU per GPU-hour. Geometric-mean gain is 2.43× over PUCT and 2.48× over the strongest generator-only baseline on each target. No single generator won every target. The lead holds under TM 0.5 and 0.8, 70% sequence clustering, and tighter post-hoc filters (actifpTM, ESMFold2 ipTM, PRODIGY). Independent repeats on CD45, SC2RBD and CbAgo also beat PUCT and epsilon-greedy.
Of 1,454 confirmed generation or redesign starts, Explore was 74% of jobs in low-evidence states, Rescue 51% in stall or deep stall, and eXploit 82% in productive-with-duplication. Dropping the Rescue priority cut stall recovery from 26/27 to 14/20, stretched median stall recovery from 1.77 to 5.04 GPU-hours, and raised time spent stalled from 3.8% to 27.7%.
The bottleneck in computational binder design is shifting from "is there a stronger generator" to "how do you schedule heterogeneous tools under a budget." The portable piece of T-REX is a structured campaign state: a 27B model only proposes and ranks, while validation and scheduling stay in deterministic code. The stack is open source and the planner can run locally. Groups already running BindCraft or BoltzGen can sit a controller on top.
This is an orchestration gain. Generator weights stay frozen for the whole campaign.
All scoring is in silico. The three generators have experimentally confirmed binders in their original papers; the 1,461 T-REX hits do not. Real programs still need affinity, specificity and developability. Generator-only baselines exclude post-cutoff AF2 evaluation from the budget because it cannot change what gets generated; adaptive controllers count evaluation inside the 144 hours, so the comparison is not fully symmetric. LLM serving is also outside worker hours. The action space and hyperparameter bounds are fixed up front; no new model is introduced mid-campaign.