The Closed Loop: NISE Teaches Two Neural Networks to Negotiate Until a Drug Has a Pocket

Designing a protein that grips a small-molecule drug is not the same problem as predicting how a natural protein folds. You are not matching a sequence to a structure you already know exists in nature. You are inventing a pocket, a sequence, and a ligand pose at the same time — and every choice constrains the others.

For years, the field's wins in this space leaned on brute-force screening or on approximating drug functional groups as if they were amino acids. The hit rates were respectable in narrow cases, but "zero-shot" design — building a binder for an arbitrary clinically used drug from scratch — remained out of reach for deep learning.

That gap is what makes the July 2026 publication of neural iterative selection–expansion (NISE) in Nature worth pausing on. Researchers led by Nicholas F. Polizzi at the University of Washington describe a closed loop between two reciprocal neural networks that designed drug-binding proteins for two chemically distinct therapeutics — the anticancer payload exatecan and the anticoagulant apixaban — with experimental success rates of 100% and 83%, respectively, and affinities reaching the picomolar range.

The paper is open access. The code is on GitHub. And the central claim is not that AI has solved drug discovery, but that a specific optimization recipe — climbing a joint probability landscape rather than minimizing a physics energy — can produce binders that traditional Rosetta pipelines miss.

Why self-consistency was not enough

Structure prediction cracked a Nobel-adjacent problem: given a sequence, infer the fold. Models such as AlphaFold2, RoseTTAFold-All Atom (RFAA), Boltz-1/2, and AlphaFold3 extended that to protein–ligand co-structures — input a sequence and a SMILES string, output a predicted complex.

The protein-design community had already turned "self-consistency" into a design principle for protein and peptide binders: design a sequence, predict its structure, keep designs where prediction matches intent. Polizzi's team argues that backbone-only self-consistency is a weak filter for small molecules. In their exatecan trajectories, most designed sequences were backbone-consistent, but only a minority placed the ligand in the intended orientation; alternative binding modes that exposed polar drug atoms to solvent dominated early rounds.

NISE adds a third axis: sequence–structure–ligand self-consistency. A design must agree with itself on backbone coordinates and on where the drug sits.

The loop: expansion, prediction, selection

NISE starts from a protein backbone with a docked ligand — in the exatecan case, coordinates stripped from a COMBS/Rosetta design so the algorithm could not inherit a tuned sequence. Each round has two moves:

1. Expansion: LASErMPNN samples many sequences for the current backbone–ligand pair at high temperature, exploring broadly rather than greedily.
2. Selection: A co-structure predictor (RFAA in the exatecan work; Boltz-2 in the apixaban campaign) folds each sequence with the ligand. Designs with high self-consistency and high ligand confidence (pLDDT) become the next round's inputs — backbone and ligand atoms updated together.

Critically, the loop does not call a physics energy function. Optimization is driven by neural confidence and geometric agreement. When the authors swapped the structure predictor for Rosetta minimization and selected on ligand energy instead, ligand pLDDT did not improve and sequence quality (negative log-likelihood) did not tighten — evidence, they argue, that both networks in the reciprocal loop matter.

The algorithm resembles iterative coordinate ascent on a joint distribution: sample from P(sequence | structure, ligand) and take a confidence-weighted argmax over P(structure, ligand | sequence), climbing toward a high-probability mode learned from Protein Data Bank co-crystal statistics.

LASErMPNN: sequence design that sees the ligand

The sequence-design network, LASErMPNN (ligand-aware sequence engineering message-passing neural network), is not a cosmetic rename of LigandMPNN. It introduces a pretrained ligand encoder — trained on synthetic ligands to predict quantum-chemical atom properties such as partial charge — whose embeddings are locked during protein-sequence training. It decodes side-chain dihedral angles alongside amino-acid identity and includes ligand nodes in every encoding round.

On held-out benchmarks, LASErMPNN recovered binding-site residues on a rigorously held-out streptavidin–biotin test and ranked the native sequence of a prior de novo drug binder (PiB) fourth among 1,001 designed sequences — without ever having seen PiB in training.

Head-to-head, LASErMPNN designs tended to be less overpacked near the ligand than LigandMPNN retrainings, with lower van der Waals repulsion — a subtle but chemically meaningful difference when the goal is a sculpted pocket rather than a crushed ligand.

Exatecan: from 120 nanomolar to lactone protection

Exatecan, a camptothecin-class topoisomerase I inhibitor used as an antibody–drug conjugate payload, is a punishing target. Its lactone ring hydrolyzes in plasma with a half-life on the order of hours; the ring-open form binds albumin and loses potency. There are almost no exatecan structures in public structural databases, so models must generalize from sparse related chemistry.

The team generated four-helix bundle scaffolds, docked exatecan with COMBS, and ran 16 traditional COMBS–Rosetta designs in parallel as a baseline. Three of those expressed binders bound exatecan, with the tightest at Kd = 8 µM.

NISE, starting from one of those poses with its sequence discarded, produced four selected designs. All four bound. The highest-affinity construct, dubbed EPIC (exatecan–protein interaction construct), reached Kd = 120 nM — roughly 360-fold tighter than human serum albumin under the same assay conditions and 70-fold better than the best traditional design in the head-to-head set.

Over 35 NISE iterations, the algorithm remodeled a 3₁₀ helix into a canonical coil, narrowed a wide helix–helix gap, and translated the ligand deeper into the bundle — burying the labile lactone while preserving camptothecin-class specificity. EPIC did not bind unrelated drug classes such as apixaban or dexamethasone.

Neural proofreading

The authors then asked whether LASErMPNN could maturate its own design without a new experimental campaign. "Neural proofreading" — revisiting binding-site residues at lower sampling temperature on an energy-relaxed RFAA model — suggested two substitutions. Single mutants improved affinity more than tenfold; the double mutant EPIC(Q51N/M97L) reached Kd = 1.2 nM, a 100-fold gain over EPIC with no new structural data. Crystal structures at 2.0 Å and 2.2 Å resolution confirmed the designed binding mode, including bidentate hydrogen bonding from Asn51 in the higher-affinity variant.

Functionally, EPIC variants repartitioned exatecan toward the intact lactone form: in PBS at pH 7.4, more than 99% of drug remained ring-closed for at least 50 hours in the presence of the tightest binder — a practical demonstration for delivery and payload stabilization, not just a binding curve on a slide.

Apixaban: 80 picomolar on a different fold

To test generality, the team switched scaffolds to the mixed α/β NTF2 fold and targeted apixaban, a Factor Xa inhibitor chemically unrelated to camptothecins. CARPdock-generated poses seeded NISE with Boltz-2 as the co-structure predictor.

Six of six tested NISE designs bound apixaban-FITC; the tightest, APEX (apixaban-binding protein exemplar), achieved Kd = 80 pM for apixaban — nearly 10,000-fold tighter than the next-leading method cited in the paper for that target. Competition experiments confirmed binding to unlabeled apixaban at comparable affinity.

Different fold, different drug class, same algorithmic skeleton: reciprocal neural networks, confidence-based selection, no Rosetta energy at the center of the loop.

What the community is weighing

The Nature Chemical Biology research highlight published 20 July 2026 frames NISE as overcoming simultaneous optimization of sequence, structure, and ligand conformation — the triad that defeated prior deep-learning loops.

Open-source release matters here: the polizzilab/NISE repository on GitHub documents workflows pairing LASErMPNN with Boltz-2x, with thirteen tagged releases as of early 2026. Reproducibility is part of the claim.

Industry conversation around AI protein design — including recent Hacker News threads on Isomorphic Labs' closed Drug Design Engine — often returns to the same caution: generating binders is not the same as developing drugs. Hit rates and picomolar constants do not shortcut toxicology, formulation, or clinical trials. NISE's authors are explicit about applications in drug delivery, sensing, and catalysis rather than a direct pipeline to approved medicines.

That honesty is the right frame. The advance is methodological: a general recipe for zero-shot small-molecule binder design where physics-based iterative selection–expansion stalled, validated on two real drugs with X-ray confirmation and functional readouts (hydrolysis protection) that go beyond affinity alone.

The wider biotech stack

NISE lands in a month when experimental data consortiums are also moving. A-Alpha Bio's Atlas Consortium, announced 22 July 2026 with founding members including GSK, Boltz, Cradle, and Dyno Therapeutics, aims to generate standardized antibody–antigen affinity datasets at a scale individual labs cannot build alone — an adjacent bottleneck in biologics rather than small-molecule pockets.

The through-line is the same: neural networks are only as useful as the loops that connect them to measurement. NISE closes the design–prediction loop for small-molecule binders. Consortia like Atlas try to close the data loop for antibodies. Neither replaces the wet lab. Both compress the distance between hypothesis and test.

For camptothecin payloads, anticoagulant sequestration, or enzyme-adjacent sensing proteins, the closed loop is now a published, executable protocol — two networks negotiating until the drug has a pocket worth ordering from a DNA synthesis vendor.

Sources

• Nature — Zero-shot design of drug-binding proteins via neural iterative selection−expansion (24 June 2026)
• Nature Chemical Biology — Designed to bind (20 July 2026)
• bioRxiv — Zero-shot design of drug-binding proteins via neural selection-expansion (22 April 2025)
• GitHub — polizzilab/NISE (repository; latest release v0.9.3, 12 February 2026)
• GlobeNewswire — A-Alpha Bio Announces Industry-First Data Consortium (22 July 2026)
• Hacker News — The Isomorphic Labs Drug Design Engine unlocks a new frontier beyond AlphaFold (February 2026)