Searcharxiv⌕ Search

arXiv subjects

Yingchao Liu

Publications and source records attributed to Yingchao Liu.

8 recordsLinked to original sources
arXiv2026

EasyNano: rapid epitope-targeted nanobody CDR design via differentiable distogram optimization with ESMFold2

Computational design of nanobodies that bind user-specified protein epitopes could transform therapeutic development, but current methods either rely on stochastic sampling requiring days of GPU computation or inverse folding approaches unable to target epitopes directly. Here we present EasyNano, a practical pipeline for rapid, epitope-targeted nanobody complementarity-determining region (CDR) design that operates in approximately 10-20 minutes on a high-end personal workstation. EasyNano optimizes CDR residue logits via gradient descent through the ESMFold2 pairwise distance distogram, using the lightweight ESMFold2-Fast model (721M) as a differentiable oracle guided by a composite loss including a dedicated epitope proximity term. A full ESMFold2 (1.3B) CA-coordinate structure prior prevents framework pose drift. The wild-type logit initialization bias emerges as a critical practical parameter controlling CDR mutability. Across six target-framework pairs spanning self-recovery and de novo design scenarios, EasyNano improves ipTM by up to +0.559 -- from 0.143 to 0.702 (Ty1/RBD) -- and achieves a 4.6-fold improvement (ipTM 0.117 to 0.538) on a manually docked AQP4-targeting framework, while preserving ipTM on already-strong binders. Random CDR baselines (n=30 per target) confirm statistical significance (5.7 sigma above random mean for Ty1). Multi-seed analysis reveals diverse local minima, underscoring the importance of replicate runs. Kabsch cross-validation against crystal structures confirms that designed CDRs preserve the framework pose basin. EasyNano demonstrates that ESMFold2-based differentiable optimization provides a fast, practical, and epitope-specific approach to nanobody CDR design.

q-bio.QM↗
arXiv2026

Interpretable enzyme function prediction via sparse autoencoder features of ESMC across the microbial protein universe

Microbial genomes and metagenomes contain millions of proteins whose enzymatic functions remain unknown, the enzyme dark matter. While deep learning has improved protein function prediction, most methods are black boxes relying on sequence or structural similarity, limiting discovery of novel catalytic activities. The ESMC-6B protein language model and its sparse autoencoder with a 16,384-dimensional codebook of interpretable biological concepts, each annotated by GPT-5, creates a new opportunity: using these features directly as semantic signatures for enzyme function. Here, we show that ESMC-SAE features enable accurate and interpretable enzyme commission (EC) number prediction without task-specific training or GPU-intensive computation. On a balanced benchmark of 4,868 microbial SwissProt enzymes across 161 EC3 subclasses, ESMC-SAE binary features achieve 78.9% top-1 and 88.5% top-5 accuracy, 37.6% higher than 3-mer baselines (57.3%). In leave-one-EC3-class-out evaluation simulating discovery of novel enzyme classes, SAE features recover the EC1 superclass in 47.7% of cases (3.3x random, 14.3%), versus 26.6% for sequence methods. Discriminative features correspond to mechanistically interpretable concepts: catalytic triad geometry for hydrolases, NAD(P)H-binding Rossmann folds for oxidoreductases, phosphate-binding P-loops for transferases. We also survey the ESM Atlas of 7.7 million clusters and identify 169,859 dark enzyme-like candidates across all major microbial phyla. Our results establish a paradigm for enzyme function discovery in microbial dark matter: interpretable by design, scalable without GPU clusters, and applicable to the billions of proteins in the ESM Atlas.

q-bio.QM↗
arXiv2026

TurboESM: Ultra-Efficient 3-Bit KV Cache Quantization for Protein Language Models with Orthogonal Rotation and QJL Correction

The rapid scaling of Protein Language Models (PLMs) has unlocked unprecedented accuracy in protein structure prediction and design, but the quadratic memory growth of the Key-Value (KV) cache during inference remains a prohibitive barrier for single-GPU deployment and high-throughput generation. While 8-bit quantization is now standard, 3-bit quantization remains elusive due to severe numerical outliers in activations. This paper presents TurboESM, an adaptation of Google's TurboQuant to the PLM domain. We solve the fundamental incompatibility between Rotary Position Embeddings (RoPE) and orthogonal transformations by deriving a RoPE-first rotation pipeline. We introduce a head-wise SVD calibration method tailored to the amino acid activation manifold, a dual look-up table (LUT) strategy for asymmetric K/V distributions, and a 1-bit Quantized Johnson-Lindenstrauss (QJL) residual correction. All experiments are conducted on ESM-2 650M, where our implementation achieves a 7.1x memory reduction (330 MB to 47 MB) while maintaining cosine similarity > 0.96 in autoregressive decoding across diverse protein families, including short peptides, transmembrane helices, enzyme active site fragments, and intrinsically disordered regions. We further implement a Triton-based fused decode attention kernel that eliminates intermediate dequantization memory allocations, achieving a 1.96x speedup over the PyTorch two-step path for the KV fetch operation alone; however, TurboESM incurs a prefill overhead of 21-27 ms relative to the original model due to KV quantization and packing, making it most suitable for memory-bound scenarios rather than latency-critical short-sequence workloads. Analysis reveals that PLMs exhibit sharper outlier profiles than large language models (LLMs) due to amino acid vocabulary sparsity, and our method effectively addresses these distributions.

q-bio.QM↗
arXiv2025

UniOTalign: A Global Matching Framework for Protein Alignment via Optimal Transport

Protein sequence alignment is a cornerstone of bioinformatics, traditionally approached using dynamic programming (DP) algorithms that find an optimal sequential path. This paper introduces UniOTalign, a novel framework that recasts alignment from a fundamentally different perspective: global matching via Optimal Transport (OT). Instead of finding a path, UniOTalign computes an optimal flow or transport plan between two proteins, which are represented as distributions of residues in a high-dimensional feature space. We leverage pre-trained Protein Language Models (PLMs) to generate rich, context-aware embeddings for each residue. The core of our method is the Fused Unbalanced Gromov-Wasserstein (FUGW) distance, which finds a correspondence that simultaneously minimizes feature dissimilarity and preserves the internal geometric structure of the sequences. This approach naturally handles sequences of different lengths and is particularly powerful for aligning proteins with nonsequential similarities, such as domain shuffling or circular permutations, which are challenging for traditional DP methods. UniOTalign therefore offers a new, mathematically principled, global matching paradigm for protein alignment, moving beyond the limitations of path-finding algorithms.

q-bio.QM↗
arXiv2025

Lie-RMSD: A Gradient-Based Framework for Protein Structural Alignment using Lie Algebra

The comparison of protein structures is a fundamental task in computational biology, crucial for understanding protein function, evolution, and for drug design. While analytical methods like the Kabsch algorithm provide an exact, closed-form solution for minimizing the Root Mean Square Deviation (RMSD) between two sets of corresponding atoms, their application is limited to this specific metric. The rise of deep learning and automatic differentiation frameworks offers a new, more flexible paradigm for such optimization problems. We present Lie-RMSD, a novel, fully differentiable framework for protein structural alignment. Our method represents the rigid-body transformation (rotation and translation) as a 6-dimensional vector in the Lie algebra se(3) of the special Euclidean group SE(3). This representation allows the RMSD to be formulated as a loss function that can be directly minimized by modern gradient-based optimizers. We benchmarked our framework by aligning two allosteric conformations of Adenylate Kinase (PDB IDs: 4AKE and 1AKE). We demonstrate that a suite of standard optimizers (SGD, Adam, AdamW, and Sophia) can robustly converge to the global minimum, achieving precision effectively identical to the analytical Kabsch algorithm. This work validates the accuracy of the Lie algebra-based gradient descent approach and establishes a robust foundation for its extension to more sophisticated and biologically relevant scoring functions where no analytical solutions exist.

q-bio.QM↗
arXiv2025

Dimer-Enhanced Optimization: A First-Order Approach to Escaping Saddle Points in Neural Network Training

First-order optimization methods, such as SGD and Adam, are widely used for training large-scale deep neural networks due to their computational efficiency and robust performance. However, relying solely on gradient information, these methods often struggle to navigate complex loss landscapes with flat regions, plateaus, and saddle points. Second-order methods, which use curvature information from the Hessian matrix, can address these challenges but are computationally infeasible for large models. The Dimer method, a first-order technique that constructs two closely spaced points to probe the local geometry of a potential energy surface, efficiently estimates curvature using only gradient information. Inspired by its use in molecular dynamics simulations for locating saddle points, we propose Dimer-Enhanced Optimization (DEO), a novel framework to escape saddle points in neural network training. DEO adapts the Dimer method to explore a broader region of the loss landscape, approximating the Hessian's smallest eigenvector without computing the full matrix. By periodically projecting the gradient onto the subspace orthogonal to the minimum curvature direction, DEO guides the optimizer away from saddle points and flat regions, enhancing training efficiency with non-stepwise updates. Preliminary experiments on a Transformer toy model show DEO achieves competitive performance compared to standard first-order methods, improving navigation of complex loss landscapes. Our work repurposes physics-inspired, first-order curvature estimation to enhance neural network training in high-dimensional spaces.

cs.LG↗
arXiv2023

Edge-pinning effect of graphene nanoflakes sliding atop graphene

Edge effect is one of the detrimental factors preventing superlubricity in laminar solid lubricants. Separating the friction contribution from the edge atom and inner atom is of paramount importance for rational design of ultralow friction across scales in van der Waals heterostructures. To decouple these contributions and provide the underlying microscopic origin at the atomistic level, we considered two contrast models, namely, graphene nanoflakes with dimerized and pristine edges sliding on graphene monolayer based on extensive ab initio calculations. We found the edge contribution to friction is lattice orientation dependence. In particular, edge pinning effect by dimerization is obvious for misaligned contact but suppressed in aligned lattice orientation. The former case providing local commensuration along edges is reminiscent of Aubry's pinned phase and the contribution of per edge carbon atom to the sliding potential energy corrugation is even 1.5 times more than that of an atom in bilayer graphene under commensurate contact. Furthermore, we demonstrated that the dimerized edges as high frictional pinning sites are robust to strain engineering and even enhanced by fluorination. Both structural and chemical modification in the tribological system constructed here offers the atomic details to dissect the undesirable edge pinning effect in layered materials which may give rise to the marked discrepancies in measured friction parameters from the same superlubric sample or different samples with the same size and identical preparation.

cond-mat.mtrl-sci↗