Siyuan Jiang

AE-PocketMiner adds self-attention to a geometric GNN to predict cryptic pockets and their allosteric coupling from one structure, validated by simulations and thiol labeling on CB1, δ-secretase and VP35.

NA-MPNN extends the ProteinMPNN architecture to a single protein-DNA-RNA graph, beating specialized tools at both RNA sequence design and protein-DNA binding specificity prediction.

De novo designed two-helix transmembrane proteins bind the outer surface of the dopamine D1 receptor and act as anchors, biased, negative or positive allosteric modulators, rescuing loss-of-function mutants.

ProRNA3D-single pairs protein and RNA language models through geometric attention to predict protein-RNA complexes without MSAs, outperforming AlphaFold 3 when evolutionary information is scarce.

T-REX is an LLM-driven campaign controller that switches between rescue, exploration and exploitation during de novo binder design, raising structurally unique hit throughput about 2.4-fold across seven targets.

AlloBench builds an updatable pipeline producing 2,141 curated allosteric sites and benchmarks 10 prediction tools on a leakage-free test set, finding all fall below 60% accuracy.

AlloFusion fuses ProtT5 embeddings, PSSM and biochemical features in a residue-level 1D CNN, predicting allosteric site residues without pocket detection and outperforming pocket-based PASSer methods.

A five-agent pipeline that searches 26 biomedical databases, counters LLM bias toward famous genes, and ranks candidates, recovering about 19% more validated CRISPR screen hits than general-purpose models.

ProteinMPNN replaces physics-based energy functions with a message-passing network, raising native sequence recovery from 32.9% to 52.4% and rescuing designs that previously failed experimentally.

BoltzDesign1 freezes Boltz-1 and optimizes the Pairformer distogram instead of backpropagating through diffusion, designing binders for proteins, small molecules, metals, nucleic acids and post-translational modifications.

A review of sequence- and structure-based allosteric site prediction methods and four design cases showing how predicted sites become lead compounds.

RhoDesign inverse-folds RNA 3D backbones into novel aptamer sequences using RhoFold-predicted training structures, producing fluorogenic Mango-III-like aptamers with hit rates rising from 22% to 75%.

KeyAlloSite uses weak, non-contact coevolutionary couplings to rank allosteric pockets and single out the few pocket residues that transmit allosteric signals, supported by asciminib and enzyme engineering cases.

A training-free optimizer that rescues borderline de novo designs by alternating RF3 gradient-guided mutation with Boltz prediction and inverse-folding cycles, sharply raising filter-passing rates.

By letting low-pLDDT AlphaFold structures train only at high diffusion noise levels, a 16.7M-parameter model beats the 200M-parameter Proteina on designability and diversity for long proteins.

DeepSSInter combines ESM2, SaProt, and geometric transformers to predict inter-protein residue contacts without MSA, outperforming existing methods in accuracy while being two orders of magnitude faster.

Allosteric communication is dominated by fast local modes when the two sites are close and by slow global modes when they are far apart; combining both channels lets CorrSite2.0 identify 97.2% of known allosteric sites.

An optimized Boltz-2 pipeline with medicinal chemistry filters and ADMET models identified active molecules or binders for 6 of 10 prospective targets while testing only 28-96 molecules each.

Odin-Multi optimizes a single binder sequence against multiple targets and off-targets inside frozen AlphaFold2, turning cross-reactivity and specificity into design objectives rather than screening outcomes.

BoltzGen is an all-atom diffusion model that encodes amino acid identity with virtual atoms so sequence and structure are generated jointly, reaching 66% target success on nine novel protein targets.

AlphaFold 3 swaps Evoformer for a Pairformer and adds an atom-level diffusion module, predicting complexes of proteins, nucleic acids, ligands and modified residues more accurately than specialized tools.

La-Proteina models Cα coordinates explicitly and compresses sequence and side chains into 8-dimensional per-residue latents, reaching 68.4% all-atom co-designability and generating proteins up to 800 residues.

Benchmarking AlphaFold3, Chai-1, Boltz-2 and RF2NA on 11 protein–aptamer complexes shows that mismatched and shuffled aptamers also yield stable, favorable complexes, so the models miss binding specificity.

AF2BIND feeds AlphaFold2 a target protein plus 20 amino-acid baits and trains a logistic regression on the pair representation to predict small-molecule binding residues de novo.

Splitting GNM slow modes into subsets and computing transfer entropy identifies global information sources that match known active and allosteric sites in 23 proteins, revealing layered, bidirectional allosteric pathways.

RhoFold+ fuses an RNA language model pretrained on ~23.7 million sequences with MSA information to predict single-chain RNA 3D structures, outperforming existing methods including AlphaFold3.

SwitchCraft frames multistate protein design as sequence optimization through Boltz-1 under composable constraints, producing allosteric switches, ligand discriminators and de novo fluorescent biosensors in silico.

RFdiffusion3 diffuses every polymer atom explicitly, unifying binder, protein-DNA, small-molecule and enzyme active-site design in a 168M-parameter model that runs 10x faster than RFD2.

RFdiffusion2 diffuses directly on atomic coordinates, scaffolds all 41 active sites of the AME benchmark, and yields five experimentally active designed enzymes from minimal chemical constraints.

RFdiffusion turns a structure prediction network into a generative one by denoising fine-tuning, solving monomer, symmetric assembly, motif scaffolding and binder design with roughly 100-fold higher experimental success than Rosetta.

resTrain injects distance restraints into AlphaFold by optimizing a learned bias on the pair representation, boosting antibody-antigen docking success from 44% to 72% with a single coarse contact.

Multiplying the Pairformer input pair representation by a scalar (1+β) steers AlphaFold 3 and Boltz-2 toward alternative conformations without retraining, raising per-state success from 0.60 to 0.73.

Boltz-2 adds an affinity module to an open co-folding model, approaching FEP-level accuracy on binding affinity at over 1000x the speed while also improving structure prediction.

Boltz-1 is an MIT-licensed, fully open model that matches AlphaFold3 and Chai-1 on biomolecular complex prediction, and its Boltz-steering variant raises physical validity to 97%.

AlignIF inverts the idea of multiple sequence alignment into multiple structure alignment for RNA inverse folding, reaching 0.595 sequence recovery and 30 experimentally functional designs.

BoltzProt-1 adds an interaction-scoring model, BoltzPPI, after generation, raising the confirmed-binder rate from 3.3% to 8.0% while 58% of confirmed binders pass every developability criterion.

AFToolkit drops MSA search, masks templates after the first recycle, and feeds AlphaFold2's internal representations to an SVM to predict ΔΔG for substitutions, insertions and deletions.

Causal interventions on ESMFold, OpenFold and Boltz-1 reveal a shared two-stage folding trunk: early blocks write sequence chemistry into pairwise space, late blocks turn it into geometry.

Probes, SAEs and steering show Boltz-1's diffusion module keeps geometry but drops sequence chemistry, and a highly decodable strand direction cannot steer structure.

Sparse autoencoders decompose ESM-2 embeddings into thousands of interpretable biological features that reveal missing database annotations and causally steer the model's sequence predictions.

Sparse autoencoders on the AlphaFold3 and Boltz-2 diffusion modules reveal interpretable features, divergent denoising strategies, and internal signals that predict de novo design success better than pLDDT.

Probing, activation patching and MSA ablations show that AlphaFold 3 compresses MSA coevolution into a structural latent space, and a handful of diverse homologs suffices.

SiteAF3 fixes the receptor, diffuses only the ligand around a specified pocket and injects pocket and hotspot information via the MSA, improving AlphaFold3 across ligand types.

MiniFold strips ESMFold's AlphaFold2-derived folding trunk and structure module down to their essentials, matching its accuracy while running 10-20x faster with far less memory.

FlashBind decouples virtual screening into fast docking plus lightweight E(3)-equivariant geometric scoring, achieving 50x speedup over Boltz-2 with comparable enrichment and prospective wet-lab validation.

ODesign unifies proteins, RNA, DNA, small molecules and ions in one all-atom generative model, matching or beating specialist tools across eleven benchmarks with orders-of-magnitude higher design throughput.

A four-round blind competition across 57 pseudoknots and ~50,000 sequences shows that deep learning, powered by RibonanzaNet, matches experienced human designers in de novo RNA pseudoknot design.

NACraft is a training-free framework that designs RNA and DNA aptamers by writing design goals as loss functions and optimizing sequences through structure-model feedback.

RFD2-MI Based deep learning model for de novo design of metalloproteases