Darin Tsui

Darin Tsui, Amirali Aghazadeh

Protein language models leverage evolutionary information to perform state-of-the-art 3D structure and zero-shot variant prediction. Yet, extracting and explaining all the mutational interactions that govern model predictions remains difficult as it requires querying the entire amino acid space for $n$ sites using $20^n$ sequences, which is computationally expensive even for moderate values of $n$ (e.g., $n\sim10$). Although approaches to lower the sample complexity exist, they often limit the interpretability of the model to just single and pairwise interactions. Recently, computationally scalable algorithms relying on the assumption of sparsity in the Fourier domain have emerged to learn interactions from experimental data. However, extracting interactions from language models poses unique challenges: it's unclear if sparsity is always present or if it is the only metric needed to assess the utility of Fourier algorithms. Herein, we develop a framework to do a systematic Fourier analysis of the protein language model ESM2 applied on three proteins-green fluorescent protein (GFP), tumor protein P53 (TP53), and G domain B1 (GB1)-across various sites for 228 experiments. We demonstrate that ESM2 is dominated by three regions in the sparsity-ruggedness plane, two of which are better suited for sparse Fourier transforms. Validations on two sample proteins demonstrate recovery of all interactions with $R^2=0.72$ in the more sparse region and $R^2=0.66$ in the more dense region, using only 7 million out of $20^{10}\sim10^{13}$ ESM2 samples, reducing the computational time by a staggering factor of 15,000. All codes and data are available on our GitHub repository https://github.com/amirgroup-codes/InteractionRecovery.

Darin Tsui, Kunal Talreja, Daniel Saeedi et al.

Protein language models (pLMs) have emerged as powerful predictors of protein structure and function. However, the computational circuits underlying their predictions remain poorly understood. Recent mechanistic interpretability methods decompose pLM representations into interpretable features, but they treat each layer independently and thus fail to capture cross-layer computation, limiting their ability to approximate the full model. We introduce ProtoMech, a framework for discovering computational circuits in pLMs using cross-layer transcoders that learn sparse latent representations jointly across layers to capture the model's full computational circuitry. Applied to the pLM ESM2, ProtoMech recovers 82-89% of the original performance on protein family classification and function prediction tasks. ProtoMech then identifies compressed circuits that use <1% of the latent space while retaining up to 79% of model accuracy, revealing correspondence with structural and functional motifs, including binding, signaling, and stability. Steering along these circuits enables high-fitness protein design, surpassing baseline methods in more than 70% of cases. These results establish ProtoMech as a principled framework for protein circuit tracing.

Darin Tsui, Aryan Musharaf, Yigit Efe Erginbas et al.

The growing adoption of machine learning models for biological sequences has intensified the need for interpretable predictions, with Shapley values emerging as a theoretically grounded standard for model explanation. While effective for local explanations of individual input sequences, scaling Shapley-based interpretability to extract global biological insights requires evaluating thousands of sequences--incurring exponential computational cost per query. We introduce SHAP zero, a novel algorithm that amortizes the cost of Shapley value computation across large-scale biological datasets. After a one-time model sketching step, SHAP zero enables near-zero marginal cost for future queries by uncovering an underexplored connection between Shapley values, high-order feature interactions, and the sparse Fourier transform of the model. Applied to models of guide RNA efficacy, DNA repair outcomes, and protein fitness, SHAP zero explains predictions orders of magnitude faster than existing methods, recovering rich combinatorial interactions previously inaccessible at scale. This work opens the door to principled, efficient, and scalable interpretability for black-box sequence models in biology.

Darin Tsui, Kunal Talreja, Amirali Aghazadeh

Computing the Fourier transform of a $q$-ary function $f:\mathbb{Z}_{q}^n\rightarrow \mathbb{R}$, which maps $q$-ary sequences to real numbers, is an important problem in mathematics with wide-ranging applications in biology, signal processing, and machine learning. Previous studies have shown that, under the sparsity assumption, the Fourier transform can be computed efficiently using fast and sample-efficient algorithms. However, in most practical settings, the function is defined over a more general space -- the space of generalized $q$-ary sequences $\mathbb{Z}_{q_1} \times \mathbb{Z}_{q_2} \times \cdots \times \mathbb{Z}_{q_n}$ -- where each $\mathbb{Z}_{q_i}$ corresponds to integers modulo $q_i$. Herein, we develop GFast, a coding theoretic algorithm that computes the $S$-sparse Fourier transform of $f$ with a sample complexity of $O(Sn)$, computational complexity of $O(Sn \log N)$, and a failure probability that approaches zero as $N=\prod_{i=1}^n q_i \rightarrow \infty$ with $S = N^δ$ for some $0 \leq δ< 1$. We show that a noise-robust version of GFast computes the transform with a sample complexity of $O(Sn^2)$ and computational complexity of $O(Sn^2 \log N)$ under the same high probability guarantees. Additionally, we demonstrate that GFast computes the sparse Fourier transform of generalized $q$-ary functions $8\times$ faster using $16\times$ fewer samples on synthetic experiments, and enables explaining real-world heart disease diagnosis and protein fitness models using up to $13\times$ fewer samples compared to existing Fourier algorithms applied to the most efficient parameterization of the models as $q$-ary functions.

Thomas Walton, Darin Tsui, Aryan Musharaf et al.

Autoregressive models have transformed protein engineering by enabling the generation of novel protein sequences beyond those found in nature. However, their sequential inference introduces significant latency, limiting their utility in high-throughput protein screening. Speculative decoding accelerates generation by employing a lightweight draft model to sample tokens, which a larger target model then verifies and refines. Yet, in protein sequence generation, draft models are typically agnostic to the structural and functional constraints of the target protein, leading to biologically implausible outputs and a shift in the likelihood distribution of generated sequences. We introduce SpecMER (Speculative Decoding via k-mer Guidance), a novel framework that incorporates biological, structural, and functional priors using k-mer motifs extracted from multiple sequence alignments. By scoring candidate sequences in parallel and selecting those most consistent with known biological patterns, SpecMER significantly improves sequence plausibility while retaining the efficiency of speculative decoding. SpecMER achieves 24-32% speedup over standard autoregressive decoding, along with higher acceptance rates and improved sequence likelihoods.

Darin Tsui, Kunal Talreja, Amirali Aghazadeh

Predicting protein function from amino acid sequence remains a central challenge in data-scarce (low-$N$) regimes, limiting machine learning-guided protein design when only small amounts of assay-labeled sequence-function data are available. Protein language models (pLMs) have advanced the field by providing evolutionary-informed embeddings and sparse autoencoders (SAEs) have enabled decomposition of these embeddings into interpretable latent variables that capture structural and functional features. However, the effectiveness of SAEs for low-$N$ function prediction and protein design has not been systematically studied. Herein, we evaluate SAEs trained on fine-tuned ESM2 embeddings across diverse fitness extrapolation and protein engineering tasks. We show that SAEs, with as few as 24 sequences, consistently outperform or compete with their ESM2 baselines in fitness prediction, indicating that their sparse latent space encodes compact and biologically meaningful representations that generalize more effectively from limited data. Moreover, steering predictive latents exploits biological motifs in pLM representations, yielding top-fitness variants in 83% of cases compared to designing with ESM2 alone.