Xiwei Cheng (@xiwei393) header
Xiwei Cheng (@xiwei393) profile picture

PhD Student @Northeastern. Ex @UCSD, RUC, ByteDance.

Joined October 2023
Excited to share our work on MolRSI! 🚀 A step toward AI scientists that continuously learn from experience and improve their own discovery strategies. Grateful to be part of this amazing team!
What if AI could discover better ways to discover molecules? Meet MolRSI: an AI scientist that recursively improves its own search policy, with results validated in the wet lab. 🧪 Antibiotic discovery: 15.9% hit rate vs 8.7% for an expert-designed non-agentic workflow. 🧬 De novo pMHC binder design: 8 selective binder hits across 2 targets vs 0 for Claude Science. Paper: arxiv.org/abs/2610.39714 Project page: yikunpku.github.io/MolRSI/ How it works 👇
1
7
255
Elucidating enzyme–substrate specificity through co-folding foundation model 1 Boltz2ESI is presented as an end-to-end enzyme–substrate specificity predictor that replaces “pocket definition + rigid docking” with native enzyme–ligand co-folding, aiming to capture ligand-induced active-site plasticity directly from sequence + substrate SMILES. 2 The core idea is a two-stage structural protocol built on the Boltz-2 co-folding foundation model: (i) MSA-guided co-folding to localize a putative active site, then (ii) an MSA-free re-folding of the cropped pocket with the ligand to force pose resolution from biophysical complementarity rather than evolutionary alignment bias. 3 From the refined co-folded complex, the method extracts Boltz-2 internal structural representations (single/token features, pairwise interaction trunk, and a complex distogram), then fuses them with global evolutionary context (ESM3 residue embeddings) and ligand chemistry priors (Uni-Mol2 3D embedding + Morgan fingerprints). 4 The interaction module is geometry-conditioned and uses a 4-block Pairformer stack with attention/message passing masked to focus on protein–ligand (and ligand–ligand) interactions; it is initialized from the pretrained Boltz-2 affinity head so the downstream model inherits thermodynamic priors while being trained for catalytic specificity labels. 5 Training/evaluation uses ESIBank (323,783 curated pairs; 8,124 enzymes; 34,417 substrates) with hard negatives generated via EC-hierarchy sampling, and two validation regimes: random split and a stringent “unknown substrate and enzyme” split where neither enzyme nor substrate identities overlap between train and test. 6 On ESIBank, Boltz2ESI reports higher performance than sequence-only ESP and structure+docking EZSpecificity, including under the “unknown substrate and enzyme” split (AUROC 0.7664 vs 0.7198 for EZSpecificity and 0.6778 for ESP). Zero-shot Boltz-2 affinity is near-random (AUROC ~0.53), supporting the paper’s point that binding affinity alone does not determine catalytic specificity. 7 Stratifying decoys by shared EC digits, Boltz2ESI maintains stronger discriminative resolution down to deep EC levels (including enzymes sharing the same four-digit EC), suggesting the co-folded geometry helps separate tight sub-family specificities that sequence similarity and rigid scaffolds often miss. 8 Ablations attribute complementary roles to modalities: removing ESM3 embeddings causes the largest single drop (AUROC to 0.6605, AUPR to 0.2100 on the hard split), while replacing co-folded complexes with AlphaFold2/AlphaFill + AutoDock-GPU docking poses collapses precision (AUPR 0.3331 to 0.2169), highlighting that complex quality and induced-fit modeling are key for high-precision substrate discrimination. 9 A mechanistic case study on citrate synthase (large ligand-dependent domain closure) shows co-folding better recovers the catalytically competent closed-like geometry than rigid docking; the resulting Boltz2ESI score is much higher for the co-folded complex (0.9770) than for the rigid-docked setup (0.6860) when the enzyme is held out from training. 10 In an external pathway task (withanolide biosynthesis), Boltz2ESI is used to rank 107 candidate plant P450s from transcriptome data without pathway-specific training: it ranks CYP87G1 8th and CYP88C7 4th (better than EnzymeCAGE’s 13th and 10th), but ranks CYP749B2 17th (worse than EnzymeCAGE’s 6th), which the authors attribute to a more complex multi-step transformation and/or training-data coverage effects. 📜Paper: biorxiv.org/content/10.64898… #ComputationalBiology #Bioinformatics #ProteinStructure #EnzymeSpecificity #DeepLearning #FoundationModels #Cofolding #Biocatalysis #MetabolicPathways #NaturalProducts
9
1
51
3,745
Xiwei Cheng retweeted
An incredibly packed room for our morning speakers Nathan Frey @nc_frey from Anthropic and Vivek Natarajan @vivnat from Google DeepMind at the AI Scientist Summer Workshop. We still have lots of fantastic speakers throughout the day! Checkout the schedule at ai-scientist-workshop.github…
We're organizing an AI Scientist Summer Workshop in Boston on August 4! We have already received an incredible amount of interest, so please register soon to secure a spot. We have an incredible line up of speakers from Harvard, MIT, Anthropic, Google DeepMind, AI2, and Microsoft Research. Marinka Zitnik (HMS) @marinkazitnik Vivek Natarajan (Google) @vivnat Nathan Frey (Anthropic) @nc_frey Paul Liang (MIT) @pliang279 Michael Brenner (Harvard, Google) Markus Buehler (MIT) @ProfBuehlerMIT Yuanqi Du (MSR) @YuanqiD Bodhisattwa Majumder (AI2) @mbodhisattwa Rafael Gómez-Bombarelli (MIT, Lila Sciences) @RGBLabMIT We're also welcoming contributed talks and posters. Please fill in the registration form to submit a talk or poster 👇 Thank you to my fellow organizers @WengongJin @YuanqiD
9
9
5
107
21,713
Xiwei Cheng retweeted
Introducing DrugSAGE: a self-evolving LLM agent that builds state-of-the-art drug discovery models by accumulating and reusing experience across tasks. It ranks #1 among 9 SOTA agents, and transfers solutions to new tasks at zero test-time search. Most LLM agents pay the full, expensive search cost on every new task and discard everything they learned. DrugSAGE remembers. 📄 arxiv.org/abs/2605.15461
4
15
2
71
10,319
Xiwei Cheng retweeted
Thrilled to announce the AI Scientist Summer Workshop @AdaFang_ @YuanqiD 📅 Aug 4, 2026 📍Microsoft Research New England, Cambridge MA Join speakers from Anthropic, Google, Microsoft, AI2, Harvard, and MIT to discuss the future of AI Scientists! Register: ai-scientist-workshop.github…
5
21
4
153
26,721
Protein-Nucleic Acid Complex Modeling with Frame Averaging Transformer • This study introduces FAFormer, an equivariant transformer architecture specifically designed for modeling protein-nucleic acid interactions, using frame averaging (FA) to enhance geometric context in predictions. • FAFormer improves contact map prediction accuracy by over 10% relative to baseline models, making it highly effective for identifying binding sites between proteins and nucleic acids. • By integrating FA at each transformer layer, FAFormer maintains E(3) equivariance, capturing 3D structural relationships more effectively than traditional transformers or GNNs. • Benchmarking against RoseTTAFoldNA shows FAFormer’s superior performance in contact prediction and aptamer screening, with a 20-30x speed advantage in certain tasks. • FAFormer’s unique architecture includes a Biased MLP Attention Module, Local Frame Edge Module, and Global Frame FFN, each leveraging FA to enhance structural representation in protein-aptamer complexes. • The model’s fast, unsupervised approach to aptamer screening offers promising potential for nucleic acid-based drug discovery, avoiding the need for labeled SELEX data. @WengongJin @ZhenqiaoSong @huangtinglin 💻Code: github.com/Graph-and-Geometr… 📜Paper: arxiv.org/abs/2406.09586 #Bioinformatics #ProteinInteractions #EquivariantModels #DrugDiscovery #MachineLearning #AIforBiology #Aptamer
8
29
2,862
Decomposed direct preference optimization for structure-based drug design • DECOMPDPO introduces a novel structure-based optimization framework, addressing challenges in drug design by aligning diffusion models with pharmaceutical requirements using multi-granularity preferences. • The method decomposes molecules into substructures, enabling flexible optimization at both global and local levels. This decomposition allows DECOMPDPO to optimize specific molecular features without compromising the whole molecule. • DECOMPDPO integrates a physics-informed energy constraint to ensure generated molecules maintain realistic conformations, enhancing the validity of its results. • One of the standout features is its ability to fine-tune pretrained diffusion models for multiple protein families, achieving a 95.2% success rate for generating molecules with medium-high affinity. • In molecular optimization tasks, DECOMPDPO achieved 100% medium-high affinity and a 52.1% success rate, outperforming baseline models such as RGA and DecompOpt. • The approach proves adaptable for two key scenarios: molecule generation across various protein targets and targeted optimization for specific subpockets, addressing practical pharmaceutical needs. • Using the CrossDocked2020 benchmark, DECOMPDPO demonstrated superior performance in both affinity metrics and molecular properties. • This framework shows that introducing decomposition improves multi-objective optimization efficiency, especially in scenarios where objectives may conflict. • DECOMPDPO highlights its potential for practical applications in pharmaceutical research, bridging gaps in data quality and enhancing the molecule generation process. • This study positions DECOMPDPO as a pioneering approach by leveraging preference alignment in structure-based drug design, a novel contribution that sets it apart from existing diffusion-based models. @QuanquanGu @Xiangxin_Zhou @xiwei393 📜Paper: arxiv.org/abs/2407.13981 #drugdesign #computationalbiology #deeplearning #diffusionmodels #AIinPharma #moleculegeneration #DECOMPDPO
8
35
5,904
Reprogramming Pretrained Target-Specific Diffusion Models for Dual-Target Drug Design 1. This study presents COMPDIFF and DUALDIFF, two new diffusion models that are reprogrammed from pretrained single-target models to design drugs that bind to two biological targets simultaneously, a task known as dual-target drug design. 2. Dual-target drugs have significant therapeutic potential, including overcoming drug resistance and improving efficacy, particularly in complex diseases like cancer. 3. The authors propose a dataset for dual-target drug design, derived from synergistic drug combinations, where two drugs produce an effect greater than the sum of their individual effects. 4. COMPDIFF and DUALDIFF use protein-ligand binding priors to align two pockets in 3D space, enabling efficient SE(3)-equivariant composed message passing for the generative sampling process. 5. DUALDIFF, in particular, introduces a more granular compositional process by incorporating SE(3)-equivariant messages at every layer of the neural network, leading to high-quality dual-target drug designs. 6. The methods leverage the knowledge gained from single-target pretraining and can be applied to dual-target drug design in a zero-shot manner, bypassing the need for extensive dual-target training data. 7. COMPDIFF and DUALDIFF significantly outperform traditional linker design methods and other baselines, such as Pocket2Mol and TargetDiff, in terms of binding affinity and molecular diversity for dual-target drug design. 8. The work highlights the growing potential of generative diffusion models in rational drug design, especially for complex and multi-target scenarios. @musicsuperman @Xiangxin_Zhou 📜Paper: arxiv.org/abs/2410.20688 #DrugDiscovery #DualTargetDrug #DiffusionModels #DeepLearning #GenerativeModels #AI #Bioinformatics #Chemoinformatics #ComputationalBiology
1
4
19
4,411
Xiwei Cheng retweeted
1/ 🔍 Wonder about the answer? cryo-EM ✖️ Foundation Model 🟰 ❓ cryo-EM ✖️ Flow Matching 🟰 ❓ cryo-EM ✖️ Diffusion Transformer 🟰 ❓ Excited to introduce our new work--cryoFM, the first cryo-EM foundation model for protein densities with flow matching, which generalizes to four tasks without finetuning. 📃 Paper: arxiv.org/abs/2410.08631 🏠️ Project: bytedance.github.io/cryostar… Joint work with Yilai Li @li_yilai , Jing Yuan @eugenejyuan , Quanquan Gu @QuanquanGu
2
33
6
100
29,949
Xiwei Cheng retweeted
Xiwei @cheng_xiwe11231 and Yuwei @yuwei_yang_ will present our work, "DecompOpt: Controllable and Decomposed Diffusion Models for Structure-based Molecular Optimization", tomorrow at #ICLR2024. 📍PS #5, Halle B #5 ⏰Thu 9 May 10:45 a.m. -12:45 p.m. CEST 🔗iclr.cc/virtual/2024/poster/…
Excited to share our work on 💊structure-based drug design💊, DecompOPT (openreview.net/pdf?id=Y3Bbxv…) accepted by #ICLR 2024. Many thanks to my coauthors @cheng_xiwe11231, Yuwei Yang, @baoy_nlp, Liang Wang and @QuanquanGu. (🧵1/6)
4
7
4,262
Xiwei Cheng retweeted
Excited to share our work on 💊structure-based drug design💊, DecompOPT (openreview.net/pdf?id=Y3Bbxv…) accepted by #ICLR 2024. Many thanks to my coauthors @cheng_xiwe11231, Yuwei Yang, @baoy_nlp, Liang Wang and @QuanquanGu. (🧵1/6)
1
11
2
39
15,997