Amplification bias in sequencing-based spatial transcriptomics: sources, mechanisms, impacts, and mitigation strategies Authors: Yuting Shan, Yanyan Piao, Qinyu Ge Published: 2026 ABSTRACT AbstractSpatial transcriptomics (ST) has emerged as a powerful approach for profiling gene expression in spatial tissue context; yet, its quantitative accuracy remains substantially compromised by amplification bias introduced during the complex library preparation process. These biases arise at multiple stages and accumulate throughout the experimental workflow, distorting transcript abundance, reducing detection sensitivity, and ultimately confounding downstream spatial analyses. This review systematically analyzes amplification bias in ST. We examine how input templates, oligonucleotide components characteristics, enzymatic properties, and experimental conditions collectively contribute to amplification bias, and discuss how these factors propagate through the workflow to generate systematic distortions in data. We further review and critically compare existing strategies for mitigation, encompassing both experimental optimizations and computational approaches and propose a practical decision framework for selecting amplification-bias mitigation strategies according to platform type, sample quality, and RNA input levels. Finally, we outline key challenges and future directions, emphasizing the need for integrative solutions that jointly consider experimental design and computational modeling. This work provides practical guidance for improving data fidelity and interpretation in ST. Introduction Spatial transcriptomics (ST) technologies, which combine gene expression profiling with spatial localization in intact tissues, have undergone rapid evolution since their pioneering introduction by Ståhl et al. in 2016 [1]. Among current ST platforms, sequencing-based ST has become one of the predominant approaches owing to its high-throughput and transcriptome-wide profiling capability. By associating the measured transcripts with their native tissue coordinates via physical or chemical barcoding, ST enables deeper insights into tissue microenvironments, developmental stages, and patterns of disease invasion [2, 3]. ST has rapidly evolved from proof-of-concept to increasingly refined platforms with improved spatial resolution and molecular sensitivity. However, these advances come with substantial experimental complexity and are inherently constrained by trade-offs among resolution, sensitivity, and tissue coverage. A typical sequencing-based ST workflow encompasses tissue processing, library preparation, sequencing, and downstream analysis (Fig. 1), and each step may contribute to technical variability and bias. Notably, formalin-fixed paraffin-embedded (FFPE) tissues, as a key resource for retrospective analyses and large-scale cohort studies in clinical and translational research, have been increasingly incorporated into ST workflows [4, 5]. However, FFPE processing introduces RNA crosslinking and degradation, reducing molecular accessibility and integrity. Given the sensitivity of ST to starting molecule quality, low-quality RNA inputs are more prone to exacerbating systematic biases during amplification. Overview of the workflow of amplification bias in sequencing-based ST and structural basis of amplification related bias. (A) Schematic overview of a typical ST workflow, including tissue processing, sequencing library preparation, sequencing, and downstream computational analysis. Steps (1)–(6) denote key amplification-related stages: (1) reverse transcription, (2) second-strand cDNA synthesis, (3) fragmentation and adapter ligation, (4) index PCR, (5) bridge PCR, and (6) sequencing by synthesis. The sequencing datasets undergo data preprocessing (including quality control, normalization, and data integration) and downstream analyses such as spatial clustering, cell type annotation, deconvolution, identification of spatially variable genes (SVGs), spatial region detection, spatial neighborhood analysis, cell–cell communication inference, and spatial trajectory inference. (B) Schematic of the capture oligonucleotide used in the 10× Genomics Visium platform. (C) Schematic of the sequence architecture of an ST sequencing library. ST Chip: spatial transcriptomics capture array chip. ST studies are further constrained by the scarcity of clinical specimens, limited tissue availability, small biopsy regions, and experimental design limitations, such as the use of small spatially barcoded capture regions in high-resolution platforms [6]. These factors collectively lead to inherently low RNA input quantities, thereby increasing reliance on amplification during library preparation. Messenger RNAs (mRNAs) are reverse transcribed into complementary DNA (cDNA), providing a stable substrate for subsequent amplification. While this step is largely standardized, amplification strategies vary across multiple ST platforms. For example, sequencing-based methods rely on multi-step polymerase chain reaction (PCR) workflows to generate sufficient material for sequencing [7], whereas imaging-based methods use in situ rolling circle amplification (RCA) [8]. Specifically, amplification, whether spatially in situ amplification or bulk nucleic acid amplification, is essential for achieving sufficient detection sensitivity and library yield. However, this extensive reliance on amplification introduces nonlinear kinetics, sequence-dependent biases, and stochastic sampling effects. These amplification-induced distortions can have profound quantitative and spatial consequences in ST datasets. They may compress the apparent dynamic range of gene expression and lead to the systematic underestimation of low-abundance transcripts [9]. In severe cases, amplification bias can disrupt the linear relationship between true transcript abundance and observed read counts [10]. Moreover, amplification-induced variability may be misinterpreted as genuine spatial heterogeneity, generating spurious expression hotspots or gradients that substantially compromise the accuracy of spatially differential expression analysis. Consequently, amplification bias may have a substantial impact in ST, as it influences not only expression quantification but also the accurate resolution of spatial expression patterns, and therefore warrants particular attention. ST has been extensively examined from multiple technical perspectives. Prior reviews have largely emphasized methodological advances, including platform development [11–13], innovations in spatial indexing strategies [14, 15], integrative multi-omics applications [16–18], and clinical translation [19]. In contrast, the origins, quantitative consequences, and mitigation strategies of amplification bias remain insufficiently explored. A systematic examination of this issue is critical for improving the accuracy and interpretability of ST data. Here, we synthesize current understanding of amplification bias in ST, delineating its molecular determinants, mechanisms of emergence, and downstream effects on quantitative measurement and spatial pattern inference. We further summarize recent advances in bias correction, normalization, and quality control strategies, and propose a framework for selecting strategies to mitigate amplification bias, with the aim of supporting more robust and reproducible ST analysis. Sources and mechanisms of amplification bias As illustrated in Fig. 1A, the ST typical workflow can be divided into several major stages. It begins with tissue preparation, which includes sectioning, fixation, and permeabilization of biological samples to maintain tissue integrity and allow RNA accessibility. Then, mRNA molecules are captured by oligonucleotide arrays or microbeads carrying spatial barcodes, enabling the recording of spatial information for each transcript. The captured mRNAs are then reverse transcribed into cDNA, converting the spatially barcoded RNA molecules into stable templates. Subsequently, amplification and a series of library construction steps are carried out to generate sequencing-qualified libraries. Finally, sequencing and computational analysis reconstruct spatially resolved gene expression profiles, which facilitate the identification of cell-type organization, molecular gradients, and tissue-specific regulatory mechanisms. Amplification bias arises from a variety of molecular and technical factors that collectively distort the quantitative representation of nucleic acid fragments (Fig. 2). Here, based on the components and conditions of the amplification process, we classify these factors into four categories: template-related bias, primer- and adapter-related bias, enzymatic bias, and thermal cycling instruments- and parameters-dependent bias. In following sections, we will discuss each category separately and explore how these sources of bias affect the quantitative accuracy of nucleic acid measurements. Determinants and mechanisms of amplification bias formation. (A) Factors contributing to bias formation, including sample-processing variables, capture-related probe designs, and intrinsic template properties, such as nucleic acid abundance, secondary structure, fragment length, and sequence composition. (B) Mechanisms of bias generation, including competitive template interactions, enzymatic kinetic selectivity, stochastic amplification effects, instrument performance, and program parameters, which together can promote biased template expansion, particularly under low-input conditions. ST Chip: spatial transcriptomics capture array chip. Template-related bias Although all ST technologies are designed to resolve spatial gene expression patterns, they differ substantially in their underlying design principles and experimental workflows. Here, we focus on key features of the input RNA for bias formation during library construction in sequencing-based ST technologies. Tissue sections serve as the primary spatial substrate in ST, preserving the native architecture of biological samples while simultaneously imposing physical constraints on molecular accessibility and capture efficiency. Spatial coding is achieved by permeabilizing tissue sections to release intracellular nucleic acids or enable the entry of exogenous barcoded probes. RNA capture in ST depends on effective permeabilization and reagent penetration [20]. Tissue structural density varies markedly across tissue types. Parenchymal tissues (e.g. liver and brain) are relatively less dense and more permissive to reagent penetration, which facilitates RNA release and capture, whereas fibrotic or highly cellular tumor tissues are more compact and less permeable. Insufficient penetration reduces mRNA capture efficiency or leads to capture failure [21], reducing the effective complexity of library input and altering RNA abundance. Suboptimal permeabilization conditions may compromise RNA integrity, resulting in structural or chemical damage such as strand breaks or base modifications, which in turn reduce primer binding and polymerase extension [22]. Following release, RNA molecules may diffuse prior to capture, resulting in cross-spot contamination. Limited spot size, as well as probe density and accessibility, lead to co-capture of transcripts from multiple cells and contamination from ambient RNA, introducing technical spatial heterogeneity. In FFPE samples, formaldehyde fixation induces chemical modifications of nucleic acids. Formaldehyde reacts with the exocyclic amino groups of nucleobases, particularly adenine, guanine, and cytosine, leading to the formation of imine and hemiaminal adducts as well as aminal cross-links. These modifications not only generate intra- and intermolecular nucleic acid cross-links but also induce extensive RNA–protein cross-linking [23]. Compared with fresh-frozen tissues, the reduced RNA accessibility in FFPE samples restricts transcript release and capture efficiency, increasing the difficulty of template acquisition and in situ reverse transcription in ST workflows. Sequencing-based ST relies on barcoded capture and labeling, making it inherently more sensitive to variability in template availability and capture efficiency. The RNA molecules that can be incorporated into downstream analytical workflows, in terms of both abundance and composition, are largely determined by spatial barcoding strategies and capture probe design. Poly(dT)-based capture probe depends on transcript 3′-end integrity and preferentially enriches mRNAs with intact poly(A) tails, limiting the detection of non-polyadenylated RNAs. In FFPE tissues, a key challenge is RNA fragmentation. Delayed tissue penetration during fixation allows continued activity of endogenous nucleases, while the high-temperature treatment during paraffin embedding further introduces physical damage, together promoting irreversible RNA degradation. This degradation process leads to substantial loss of 5′-end information, which ultimately manifests as a pronounced 3′ coverage bias at the data level. Notably, poly(A)-based capture can introduce gene-specific biases. For example, in mouse lens tissue, highly expressed genes such as Crybb3 and Cryaa are systematically underdetected [13]. In comparisons between Visium and bulk RNA-seq, underdetected genes are enriched for poly(T) motifs near their 3′ ends [24]. These motifs can form secondary structures with the poly(A) tail, hindering hybridization with poly(dT) probes and reducing capture efficiency. These factors introduce bias that is independent of the underlying biological state and may be further amplified during downstream amplification and sequencing steps. Template-related biases can also arise from intrinsic sequence characteristics of nucleic acid templates, including guanine‑cytosine (GC) content, sequence composition, local sequence context, and secondary structure, which affect strand accessibility and amplification efficiency. Extreme GC content or repetitive motifs in templates significantly increase the likelihood of stable secondary structure formation [25]. Such structures interfere with primer binding or polymerase extension through intramolecular base pairing, thereby suppressing or prematurely terminating amplification [26, 27]. Hairpins with longer stems and smaller loops exert stronger suppression of template amplification due to their greater thermodynamic stability. Fan et al. reported that stable stem–loop structures with stems exceeding 20 nucleotides in regions adjacent to the amplicon can severely impede PCR, rendering the formation of the correct product nearly impossible [28]. Apart from that, the physical integrity and abundance of templates are also critical, damaged or low-copy templates are particularly prone to uneven amplification [29]. In the ST workflow, each step is tightly coordinated through sequential template–product relationships, forming a highly coupled reaction cascade. The product generated at each stage in turn functions as the template for the subsequent step, thereby enabling the progressive transfer and amplification of transcript information throughout the workflow (Fig. 1A (1)–(6)). Collectively, these template-intrinsic features give rise to the earliest layer of bias, forming the basis upon which subsequent primer-, enzyme-, and processing-related influences accumulate. Capture probe-, adapter-, and primer-related bias For ST, reverse transcription is typically performed in situ within tissue sections to preserve the spatial context of gene expression. On array-based platforms such as 10× Genomics Visium and Slide-seq, capture probes are designed to include a PCR handle, a spatial barcode, a unique molecular identifier (UMI), and a poly(dT) sequence (Fig. 1B). These capture probes serve a dual function: they hybridize to polyadenylated mRNA while simultaneously providing spatial and molecular information, and they also act as primers to initiate reverse transcription. The integration of multiple functions increases the length of capture probes, which can reach 60 nt or more on the Visium [30]. Due to variability in tissue section thickness, some studies employ sections thicker than 15 μm to maximize the recovery of intact cellular RNA. Under these conditions, primers and templates must diffuse over longer distances, which can introduce gradient effects and reduce transcript capture efficiency. In addition, the density of capture probes across the array is not perfectly uniform [31], potentially generating technical spatial heterogeneity in the resulting transcriptomic data. The adapter sequence serves as a bridge connecting the cDNA to the sequencing platform and typically contains the P5/P7 sequencing adapters, primer binding sites for sequencing, and index sequences for multiplexing (Fig. 1C). Its contribution to bias is similar to that of template preferences: the GC content and secondary structure of adapter sequences affect the amount of template available for PCR by directly influencing the accessibility of primer binding sites. Additionally, excess adapter concentrations will promote adapter-dimer formation [32], which diverts sequencing capacity and reduces effective depth for genuine cDNA. Consequently, the detection of low-abundance transcripts, which strongly depends on adequate depth, is disproportionately impaired [33, 34]. Primers serve as essential elements in polymerase-driven amplification reactions, providing the starting point for DNA synthesis. Their concentration, GC content, secondary structure, and melting temperature (Tm) determine binding specificity and amplification efficiency [35]. For example, compared to primers targeting regions with low GC content, primers targeting GC-enriched motifs typically exhibit higher amplification efficiency [36]. In addition, primers with a GC content of 40%–60%, minimized secondary structure formation (including hairpins, self-dimers, and cross-dimers), and closely matched Tm (ΔTm ≤ 1–2°C) generally exhibit superior performance [37]. In some cDNA second-strand synthesis steps, random hexamer primers are used to replace or supplement oligo(dT) primers to improve the uniformity of full-length transcript coverage. Although random hexamer primer pools are theoretically capable of representing all possible six-nucleotide combinations, their actual synthesis may still introduce compositional biases. Moreover, the occurrence probability of these hexamers within the local sequence context of the transcriptome is non-uniform [38]. Previous studies have shown that the use of random hexamer primers introduces nucleotide composition bias at read start sites, and also affects the positional uniformity of read distribution along expressed transcripts [39]. Enzymatic bias During reverse transcription, PCR, RCA, and in vitro transcription (IVT), various enzymes act as biological catalysts driving nucleic acid synthesis and amplification. Their activity, specificity, fidelity, and stability directly influence reaction efficiency, amplification uniformity, and overall library quality. Importantly, enzymatic reactions are unable to correct existing biases and may instead exacerbate them. The influence of various enzymes on library construction has been extensively studied [27, 40, 41]. Structural differences that determine the functions of reverse transcriptases (RTs) and DNA polymerases (DNAPs) result in variations in their catalytic performance. Although primers with high GC content have been reported to perform efficiently, the polymerase itself typically shows reduced efficiency when amplifying high GC or long templates. In addition, the fidelity of polymerases directly determines the likelihood of errors (base substitutions, insertions, and deletions) introduced during amplification [42]. RTs introduce bias through two intrinsic limitations: low fidelity due to the absence of proofreading activity [43] and incomplete reverse transcription caused by RNase H-mediated degradation of structured RNA template. Fu et al. reported in Pixel-seq that additional amplification is required to compensate for regionally reduced replication efficiency during gel stamp transfer [44]. This observation suggests that enzymatic replication or amplification on a 2D solid interface is subject to spatially heterogeneous reaction efficiency, and these patterns may also partially reflect edge effects arising from reagent diffusion. In necrotic or inflammatory regions, endogenous protease activity is often elevated. Infiltrating immune cells secrete a range of proteases that contribute to extracellular matrix degradation and tissue remodeling [45, 46]. In such protease-rich microenvironments, exogenous enzymes used in molecular assays may be subject to degradation, leading to reduced enzymatic stability and decreased reaction efficiency. In FFPE tissues, residual paraformaldehyde (PFA) in localized areas can also inhibit enzyme activity by impairing enzyme structure [47, 48]. Such factors may reduce reverse transcription efficiency in specific spatial regions, potentially compromising the accuracy of the resulting transcriptomic data. During RCA, strand-displacing polymerases such as Phi29 DNAP are sensitive to template length and structure, resulting in length-dependent amplification bias correlated with the helical periodicity of B-form DNA [49]. In vitro, Phi29 efficiently amplifies adenine-rich circular templates, whereas guanine-rich templates with stable secondary structures are amplified less efficiently [50]. The template encodes the sequence information, and its properties determine the binding efficiency of both primers and enzymes. Primer–template interactions influence the initiation efficiency of enzymatic reactions, whereas the enzyme’s activity, fidelity, and processivity directly govern the efficiency and accuracy of nucleic acid synthesis. These molecular-level factors are interdependent (Fig. 2), and even subtle differences in binding or extension can be exponentially amplified during enzymatic reactions. Thermal cycling instruments- and parameters-dependent bias As most platforms perform reverse transcription in situ, spatial variations in surface heating or edge effects can lead to differences in amplification efficiency and should be considered as potential sources of technical spatial heterogeneity. For sequencing-based ST platforms, RNA is transferred to tube systems for amplification after spatial barcoding and reverse transcription (Fig. 1A (1)–(2)). Amplification variability at this stage is no longer spatially constrained but instead driven by reaction composition and thermocycling condition. During the early stages of amplification, hybridization between nucleic acid molecules and the kinetic and thermodynamic properties of polymerases typically introduce only minor biases. However, as amplification proceeds, platform characteristics such as heating and cooling rates and the uniformity of thermal modules, together with dynamic adjustments of amplification parameters including annealing time and the number of cycles, can substantially amplify these subtle differences. An excessively steep thermal profile, corresponding to overly rapid heating and cooling rates, limits the time the sample spends above the critical denaturation temperature, resulting in incomplete denaturation or poor amplification of GC-rich templates [51]. In addition, reaction conditions also contribute to bias. Lower annealing temperatures reduce the energetic barrier for mismatches, increasing the amplification of nonspecific products, whereas higher annealing temperatures suppress efficient hybridization, resulting in selective underrepresentation [52, 53]. Cycle number is the principal contributor to bias amplification: additional cycles allow early, subtle disparities to expand exponentially into substantial abundance shifts, particularly under low-template conditions where stochastic drift becomes prominent [54, 55]. In addition, significant bias can be introduced in the downstream steps of Illumina sequencing [51], arising from cluster amplification and sequencing by synthesis, both involving polymerase-mediated primer extension (Fig. 1A (5)–(6)). In summary, amplification bias emerges from the interplay of multiple factors spanning molecular, technical, and experimental dimensions. While molecular features define the inherent direction of bias, instrument and reaction conditions govern its amplification strength and observability. The mechanistic relationships are summarized in Table 1. Table 1Amplification biases in sequencing-based spatial transcriptomics capture array chip: mechanisms, consequences, and mitigation strategies.Amplification bias mechanismStage in ST workflowMechanistic origin and processDownstream effectMitigation strategiesTemplate-intrinsic constraintsRT;PCR-Low template input induces stochastic sampling variability;-RNA integrity differences and poly(A)-based capture preferences bias transcript availability;-GC content, RNA secondary structure, and transcript length cause sequence-dependent reverse transcription inefficiency and polymerase stalling.-Dropout of low-abundance transcripts;-Systematic underrepresentation of GC-rich or structured genes;-Distorted gene expression distributions.-Improved RNA preservation and fixation protocols;-Compartmentalized PCR;-Reaction additives (e.g. low-melting agarose);-Spike-in controls for transcript quantification bias.Priming/adapter affinity biasRT;Adapter ligation;PCRSequence-dependent hybridization efficiency and ligation bias of primers/adapters lead to uneven capture and amplification of transcripts.-Reduced transcriptome complexity;-Uneven transcript representation;-Barcode imbalance;-Biased quantification of gene expression.-Ligation-free or template-switching methods;-Template-switching oligonucleotide (TSO) strategy;-Balanced hexamer pool design;-Spike-in controls for transcript quantification bias.Enzyme kinetic heterogeneityRT;PCRVariability in reverse transcriptase and DNA polymerase processivity, fidelity, and GC sensitivity induces transcript-dependent amplification efficiency differences.-Artificial sequence inflation;-Increased technical noise;-Reduced accuracy in expression quantification and variant calling.-Use of high-processivity and thermostable reverse transcriptases with improved tolerance to RNA secondary structures;-Use of RNase H-deficient reverse transcriptases to prevent premature degradation of RNA templates;-Enzyme cocktail optimization;-Compartmentalized PCR;-Spike-in controls for transcript quantification bias.Amplification dynamics distortionPCREarly stochastic variation in low-input templates is exponentially amplified during PCR, causing nonlinear amplification kinetics and dynamic range compression.-Over-amplification of dominant transcripts, loss of low-abundance signals;-Artificial spatial gradients;-False spatial patterns.-Reduction of PCR cycle number;-Implementation of linear amplification methods;-Use of UMIs for deduplication;-Computational normalization of amplification depth;-Spike-in controls for transcript quantification bias;-Compartmentalized PCR. Amplification biases in sequencing-based spatial transcriptomics capture array chip: mechanisms, consequences, and mitigation strategies. Impact of amplification bias on ST data analysis and biological interpretation ST datasets exhibit higher sparsity than single-cell RNA sequencing (scRNA-seq), with dropout rates often exceeding 90% versus ~80%–90% in scRNA-seq datasets [56]. Because only a limited fraction of transcripts is successfully captured and amplified, quantitative measurements in ST may be more susceptible to amplification-derived distortions. In addition, stochastic effects during reverse transcription and early PCR cycles may lead to insufficient amplification of low-copy transcripts [57], whereas templates with intrinsically higher amplification efficiency may become progressively overrepresented. This may result in systematic distortions of transcript abundance estimates. Amplification bias effects alter the proportional relationship between observed read counts and true transcript abundance. Studies in microbiome sequencing have shown that template-dependent amplification efficiency differences can substantially distort relative abundance estimates [58, 59]. Although these observations were derived from microbiome datasets, they suggest a potential mechanism by which amplification efficiency differences can substantially distort sequencing-based quantitative measurements. In ST, many widely used downstream analyses, such as spatial clustering, differential expression, and pathway enrichment, typically rely on relative expression patterns or ordering rather than absolute RNA molecule counts. The extent to which these effects quantitatively propagate into specific ST analytical workflows remains insufficiently characterized and warrants further systematic evaluation. Beyond transcript quantification, excessive PCR amplification can reduce library complexity. Based on analyses of bulk RNA-seq and scRNA-seq datasets, Parekh et al. reported a negative association between PCR cycle number and statistical power for detecting genes with modest expression changes (log₂ fold change = 0.5), which corresponded to an observed decrease of ~2.4% per additional PCR cycle [60]. These observations suggest that increased amplification may exacerbate sampling noise and reduce effective library diversity in bulk RNA-seq and single-cell RNA-seq datasets. However, it should be noted that such quantitative estimates have not yet been systematically validated in ST datasets, where amplification dynamics may further interact with tissue-specific capture efficiency and spatial heterogeneity. In addition to expression-level distortions, FFPE tissues introduce a distinct class of amplification-associated artifacts. Formaldehyde-induced nucleobase modifications can promote polymerase misincorporation during PCR amplification, resulting in characteristic C > T and G > A substitutions [61, 62]. These artifacts can increase false-positive variant calls and may be misinterpreted as region-specific molecular features, thereby confounding spatial genomic analyses based on archival clinical samples. Tumors are characterized by extensive genomic alterations, including mutations and copy number variations, along with a highly heterogeneous microenvironment. In studies of tumor spatial organization and heterogeneity, cell–cell communication and signaling pathway activity are typically inferred from coordinated expression patterns [63, 64]. Accordingly, tumor spatial architecture and microenvironment are interpreted through spatial molecular states and the organizational patterns they form. In ST, particularly in tumor tissues, direct quantitative evidence for amplification bias remains limited. This is partly because amplification is an inherent component of ST library preparation workflows and partly because ST studies have generally prioritized biological interpretation over systematic methodological benchmarking. Taken together, systematic uncertainty in sequencing-derived measurements and its impact on biological interpretation remain unresolved challenges in ST. Unlike sequencing-based ST platforms, most imaging-based technologies rely on direct molecular imaging or targeted hybridization rather than PCR-based amplification. Although some imaging-based methods employ RCA to enhance fluorescence signals, these reactions are initiated by target binding and primarily serve as in situ signal amplification rather than nucleic acid amplification. Consequently, amplification-related biases are expected to be less pronounced in imaging-based approaches compared with sequencing-based platforms. Correction and optimization strategies for amplification bias PCR-free library preparation is generally considered the most direct and theoretically optimal approach to minimize or eliminate amplification‑introduced distortions in gene expression profiles [65, 66]. However, in ST, RNA input amounts are typically far below the requirements for PCR-free library preparation (tens to hundreds of nanograms), making amplification unavoidable due to limited sample availability and experimental constraints. Although existing ST platforms incorporate mitigation strategies in their original design (Table 2), a systematic synthesis and evaluation of amplification bias reduction approaches remains lacking. This section summarizes mitigation strategies at both experimental and computational levels, aiming to support robust experimental design and downstream data analysis in ST. Table 2Overview of spatial transcriptomics platforms with detection principles, amplification strategies, and bias mitigation approaches.MethodsDetection principleAmplification strategyMitigation strategiesRef.ISSImaging-basedRT and RCA-LNA-modified primer;-Control probe density;-Reaction condition optimization[8]Xenium (10× Genomics, USA)Imaging-basedRCA-like signal amplification-Dual-arm probe design for target validation;-Target-dependent padlock probe ligation followed by in situ RCA signal amplification;-Error-correcting barcode[67]MERFISHImaging-basedNon-enzymatic optical amplification-Array-synthesized oligo pools;-Avoid predicting probe sequences that form strong secondary structures[68]seqFISH+Imaging-basedNon-enzymatic optical amplification-Hybridization time optimization;-BLAST screening for probe specificity;-Standardized GC content;-Utilization of error-correcting barcodes;-Thorough elution, re-hybridization in each round[69]STSequencing-basedRT and PCR-UMI calibration corrects counting bias caused by PCR replication[1]Visium/Visium HD (10× Genomics, USA)Sequencing-basedRT and PCR-UMI calibration corrects counting bias caused by PCR replication;-High-density continuous capture grids and multi-scale binning reduce spatial boundary effects and sampling noise[70, 71]Ex-STSequencing-basedRT and PCR-Longer, staged RT improves completeness;-UMI calibration corrects counting bias caused by PCR replication[72]DBiT-seqSequencing-basedRT and PCR-UMI calibration corrects counting bias caused by PCR replication;-Maxima H Minus reverse transcriptase;-Kapa HiFi Enzyme minimizes PCR-introduced errors and bias;-Control the number of PCR cycles[73]Decoder-seqSequencing-basedRT and PCR-UMI calibration corrects counting bias caused by PCR replication;-Maxima H- reverse transcriptase and KAPA HiFi hotstart high-fidelity enzyme[74]Slide-seqSequencing-basedRT and PCR-3′ end strategy reduces length-dependent bias;-Limit the number of PCR cycles;-UMI calibration corrects counting bias caused by PCR replication;-Compare to bulk and scRNA-seq[75]Seq-ScopeSequencing-basedBridge amplification;RT and PCR-Spatial smoothing and template input control;-High-fidelity polymerase;-Optimized PCR cycles;-UMI calibration corrects counting bias caused by PCR replication[76]Stereo-seqSequencing-basedRCA-based DNB array;RT and RCA-Uniform array fabrication and spatial normalization algorithm;-Limit amplification times to prevent local signal saturation;-GC-tolerant polymerases and optimized salt concentrations;-Increase RT efficiency and use optimized random primers[77]Abbreviations: ISS, in situ sequencing; RT, reverse transcription; RCA, rolling circle amplification; LNA, locked nucleic acid; ST, spatial transcriptomics; UMI, unique molecular identifier; PCR, polymerase chain reaction; Ex-ST, expansion spatial transcriptomics; scRNA-seq, single-cell RNA sequencing. Overview of spatial transcriptomics platforms with detection principles, amplification strategies, and bias mitigation approaches. Abbreviations: ISS, in situ sequencing; RT, reverse transcription; RCA, rolling circle amplification; LNA, locked nucleic acid; ST, spatial transcriptomics; UMI, unique molecular identifier; PCR, polymerase chain reaction; Ex-ST, expansion spatial transcriptomics; scRNA-seq, single-cell RNA sequencing. Experimental design strategies for reducing amplification bias External RNA spike-in controls serve as an external quantitative reference for assessing amplification bias by benchmarking observed read counts against known input concentrations. The External RNA Control Consortium (ERCC) spike-in RNA standards are synthetic transcripts characterized by defined sequences and minimal homology to endogenous RNAs [78]. They exhibit a broad dynamic range spanning approximately six orders of magnitude [79]. Owing to these properties, ERCC spike-ins are widely used as external controls to assess detection limits, quantification linearity, and amplification bias in sequencing-based assays [80, 81]. However, in ST, their applicability is limited. ERCC molecules are short, single-exon synthetic RNAs that differ substantially from endogenous transcripts. The absence of native ribonucleoprotein complexes, together with the lack of complex secondary structures [82], allows ERCC molecules to diffuse more freely within tissue environments. Furthermore, as exogenous spike-in controls, ERCC molecules do not experience tissue-dependent RNA release and therefore fail to reflect the efficiency of tissue permeabilization or the release of RNA from cells and the extracellular matrix. Their relatively short and variable poly(A) tails (20~26 nt) also introduce capture biases in oligo(dT)-based ST platforms [83]. These factors cause the recovery of ERCC spike-ins to diverge from that of endogenous transcripts, limiting their utility as process-matched controls in ST. Newer spike-in standards such as Spike-In RNA Variants (SIRVs) have been developed to better recapitulate transcript length diversity, poly(A)-tail characteristics, and alternative splicing complexity found in endogenous transcriptomes. In ST experiments, a relatively low proportion of spike-in controls is generally recommended, considering the extremely low RNA input per spot and the frequently degraded nature of FFPE-derived tissues. Another key parameter is the timing of spike-in addition. Spike-ins introduced concurrently with the sample can capture overall technical variability across the experimental workflow. In contrast, late addition restricts their utility to library preparation-related biases, while failing to capture spatial process-related biases introduced during early steps of ST, including tissue permeabilization and in situ capture. UMIs, random tags attached to each molecule at an early stage and used to collapse reads derived from the same molecule, represent one of the most widely adopted strategies for avoiding amplification bias and improving the accuracy of estimating original molecule abundance [84, 85]. Typically, UMIs are incorporated into cDNA during reverse transcription using oligo(dT) probes containing spatial barcodes, allowing each transcript to be uniquely labeled prior to PCR amplification. As catalysts driving amplification reactions, the key properties of enzymes widely used in ST, as well as those with potential applications in this field, are summarized in Table 3. During the reverse transcription stage, RTs with reduced RNase H activity, such as SuperScript II, SuperScript III, and SuperScript IV, demonstrate enhanced fidelity and an improved capacity to navigate complex template secondary structures [86]. The RNase H-deficient reverse transcriptase Maxima H Minus demonstrates superior sensitivity and uniformity in sequencing applications involving ultra-low RNA inputs (0.5 pg) [87]. Bst DNAP performs well on long, highly structured viral RNA, mainly due to its relatively high reaction temperature (65°C) [27]. Elevated temperatures facilitate the disruption of RNA secondary structures, thereby increasing template accessibility, while high processivity of Bst DNAP at 65°C reduces pauses and dissociation events during reverse transcription. For ST, these examples indicate that properties such as catalytic selectivity, template-binding strength, processivity, thermal stability, and the ability to resolve RNA secondary structures should be prioritized in directed evolution to achieve optimal performance. Table 3Polymerase performance.ReactionEnzymeStructureRNase H activityProofreading activityTemperatures (°C)Fidelity(error rate)Amplicon size (kb)NA detection levelsGC-richAT-richRTAMV RTsWild type+No37–52 (42)10−5−1 ng - 1 μg––MMLV RTsWild type+No4210−5<71 ng - 5 μg–– SuperScriptIIEngineeredReducedNo37–4210−5≤12.31 ng - 5 μgHigh–IIINo5010−5≤12.310 pg - 5 μgHigh–IVNo50–5510−5≤12.310 pg - 5 μgHigh– Maxima H MinusEngineeredH−No50–5510−5≤201 pg - 5 μgHigh– Bst 3.0 DNAPEngineered−No65–7210−5≥0.3>10 copies––PCRTaq DNAPWild typeN/ANo7210−4≤50–80 copiesLow–AccuPrime Pfx DNAPEngineeredN/AYes6826× Taq≤1210 pg - 200 ngLow–Phusion DNAPEngineeredN/AYes7250× Taq≤205–100 ngHigh–Q5 High-Fidelity DNAPEngineeredN/AYes62a280× Taq≤20HighHighKapa HiFi DNAPEngineeredN/AYes68–72100× Taq≤110.1–500 ngHighHighPlatinum SuperFi II DNAPEngineeredN/AYes60 b300× Taq≤200.4–250 ngHighHighPfuUltra II DNAPEngineeredN/AYes68–7220× Taq≤1950–500 ngHigh–RCAPhi29 DNAPEngineeredN/AYes3010−6 - 10−7>70–––IVTT7 RNAPWild typeN/ANo3710−5<23.7–––RNase H activity is indicated as follows: +, full activity; Reduced, partial activity; H-, Lack of activity; −, absent. N/A, not applicable. N× Taq, N-fold higher fidelity relative to Taq DNAP. Dash (–), data not available.Abbreviations: RTs, reverse transcriptases; DNAP, DNA polymerase; RNAP, RNA polymerase. SuperScript II, SuperScript III; and SuperScript IV are commercial engineered MMLV‑derived reverse transcriptase variants from Thermo Fisher Scientific. aAccording to the recommendations from New England Biolabs, the optimal annealing temperature should be determined based on the specific primer pair used. bUniversal primer annealing at 60°C. Polymerase performance. RNase H activity is indicated as follows: +, full activity; Reduced, partial activity; H-, Lack of activity; −, absent. N/A, not applicable. N× Taq, N-fold higher fidelity relative to Taq DNAP. Dash (–), data not available. Abbreviations: RTs, reverse transcriptases; DNAP, DNA polymerase; RNAP, RNA polymerase. SuperScript II, SuperScript III; and SuperScript IV are commercial engineered MMLV‑derived reverse transcriptase variants from Thermo Fisher Scientific. aAccording to the recommendations from New England Biolabs, the optimal annealing temperature should be determined based on the specific primer pair used. bUniversal primer annealing at 60°C. The template-switching oligonucleotide (TSO) anneals to non-templated cytosine residues added by the terminal transferase activity of reverse transcriptase at the 5′ end of cDNA, promoting template switching and the incorporation of adapter sequences at both ends. This facilitates full-length mRNA capture and efficient amplification in ST [88]. The library preparation method SHERRY developed by Lin Di et al. mitigates amplification bias by reducing second-strand synthesis after reverse transcription [89, 90]. This approach employs Tn5 transposase to directly target RNA/DNA hybrids, circumventing conventional second-strand synthesis while maintaining coverage uniformity and reproducibility across varying input amounts. During high-throughput sequencing library preparation, common mitigation strategies for PCR-induced bias include reducing the number of PCR cycles, employing high-fidelity polymerases, optimizing primer design, and refining reaction conditions. Lu Meng et al. demonstrated that limiting multiplex PCR to fewer than 10 cycles enhances amplicon uniformity and reduces nonspecific products and primer dimers [91]. Studies evaluating commonly used polymerases for sequencing library preparation indicate that KAPA HiFi, Platinum SuperFi II, and Pfu DNA polymerases consistently perform well [41, 92]. With two or more key metrics across strong fidelity, processivity, and proofreading activity, these DNAPs ensure high library uniformity and minimize chimera formation, supporting the production of high-quality sequencing libraries. Q5® High-Fidelity DNA polymerase demonstrates similarly high fidelity and enhanced processivity, making it a promising alternative for optimizing NGS library construction. To address the stochastic and template-competition biases inherent to PCR amplification, compartmentalized amplification techniques have been proposed. For instance, Emulsion PCR (ePCR) improves amplification uniformity by compartmentalizing template molecules into aqueous–oil microdroplets, allowing each molecule to be amplified in an isolated microreactor [93]. This isolation prevents substrate competition and reduces template mispairing and PCR suppression, thereby minimizing amplification bias and improving library complexity compared to conventional PCR [94]. Similarly, techniques such as compartmentalized capillary PCR [95], microfluidic droplet PCR [96] physically partition the reaction system into numerous microchambers or capillary units, enabling single or few template molecules to amplify in isolated environments. Agarose PCR improves amplification sensitivity and enhances reaction uniformity by approximately two-fold through the restrictive effects of its polymeric network and intrinsic molecular crowding, providing a robust strategy for high-quality transcriptome sequencing of low-input samples [97]. In addition to exponential amplification, linear amplification methods such as IVT and linear amplification via transposon insertion (LIANTI), as well as quasi-linear approaches including multiple displacement amplification (MDA) and multiple annealing and looping-based amplification cycles (MALBAC), have been employed as alternative or complementary strategies to mitigate systematic bias during amplification. IVT achieves RNA amplification by introducing a T7 promoter into cDNA fragments [98, 99], whereas LIANTI implements random template amplification via Tn5 transposon-mediated insertion of a T7 promoter [100]. Quasi-linear methods, including MDA, utilize φ29 DNAP for multi-primed rolling circle-like amplification [101], and MALBAC performs initial amplification with semi-random primers followed by product looping to prevent re-amplification [102]. By amplifying each template approximately proportionally, these strategies reduce the impact of sequence-dependent amplification efficiency differences on library composition, thereby preserving relative transcript abundance and enhancing library uniformity in low-input or complex samples. Computational methods for mitigating amplification Despite careful experimental design, amplification-related biases may persist in ST datasets. To address these systematic errors, computational strategies can be organized into a multilayered correction framework spanning molecular-level deduplication, statistical imputation, bias-aware quantification, and spatially informed quality control. At the molecular level, UMI-based deduplication reduces artificial transcript inflation arising from PCR amplification and sequencing errors by identifying reads derived from the same original molecule. Widely used implementations such as UMI-tools [103] and Cell Ranger [104, 105] perform consensus-based error correction and read collapsing to improve transcript quantification accuracy. Beyond molecular deduplication, statistical expression recovery methods can further mitigate the downstream consequences of amplification-related noise, particularly sparsity and dropout events. For example, SAVER [106] is a typical probabilistic model-based strategy that was originally developed for UMI-based single-cell RNA-seq data, which treats observed counts as noisy realizations of underlying molecular abundance and combines the measured value with a data-driven model learned from correlated genes. In this way, unreliable low counts or zeros can be shrunk toward more plausible expression estimates, while highly supported observations are preserved. For ST datasets, this Bayesian recovery framework is particularly useful for reducing random technical variation, but it should be applied cautiously in spatial datasets because excessive correction may obscure genuine spatial heterogeneity or convert true biological zeros into nonzero estimates. Although SAVER-like frameworks do not directly model region-specific capture efficiency or other spatially structured technical biases, their probabilistic expression-recovery principles may provide a useful conceptual and methodological basis for developing correction approaches tailored to ST data. Compared with SAVER-like models, graph-based imputation methods have considerable potential for mitigating the downstream effects of technical noise in ST sequencing data. MAGIC [107] is a representative graph-based method originally developed for single-cell RNA-seq data. Its central assumption is that biologically related cells occupy neighboring positions on an underlying transcriptional manifold, whereas dropout events caused by inefficient molecular capture, stochastic amplification, or insufficient sequencing are comparatively random. MAGIC first constructs a cell–cell affinity graph and then applies Markov diffusion to propagate expression information across transcriptionally related cells. Through this process, weakly detected or missing transcripts can be partially recovered, and gene–gene relationships obscured by sparse measurements may become more apparent. Nevertheless, because MAGIC does not explicitly distinguish technical zeros from genuine biological absence, excessive diffusion may over-smooth cellular heterogeneity or introduce spurious gene–gene correlations. The graph-diffusion principle underlying MAGIC could be further adapted to ST data by incorporating spatial coordinates, tissue morphology, local capture efficiency, and anatomical boundaries into graph construction and information propagation. A recent preprint has proposed SpatialMAGIC [108], a hybrid framework that combines MAGIC-based graph diffusion with spatial self-attention to recover sparse expression signals while preserving local tissue organization. This emerging work illustrates the potential of adapting graph-based imputation to ST data; however, because it has not yet undergone peer review, its robustness, generalizability across ST platforms, and ability to avoid artificial spatial patterns require further independent benchmarking. Beyond stochastic noise, systematic biases arising from library preparation and transcript features require model-based bias correction strategies. LiBiNorm explicitly characterizes and corrects global and transcript length-dependent biases in Smart-seq and related protocols, substantially improving quantification accuracy for long transcripts [109]. Salmon integrates a bias-aware probabilistic framework that accounts for sequence composition, GC content, and positional fragment biases during abundance estimation, enabling real-time correction during inference [110]. These approaches highlight the importance of explicitly modeling technical bias rather than treating it as random noise. At the spatial level, quality control and spatially aware normalization methods further enhance data reliability by leveraging spatial context. SpotSweeper identifies spot- and region-level outliers based on local spatial consistency, reducing the impact of localized amplification and library preparation artifacts while preserving biologically meaningful heterogeneity [111]. SpaNorm further addresses spatially structured technical variation by considering both library size effects and spatial expression patterns [112]. By separating unwanted library size variation from biologically meaningful spatial signals, it can improve normalization while preserving genuine tissue heterogeneity. Together, SpotSweeper and SpaNorm provide complementary strategies for spatial quality control and bias-aware normalization in ST data. Collectively, these computational strategies constitute a hierarchical correction framework, ranging from molecular deduplication to statistical imputation, bias-aware modeling, and spatial quality control, thereby improving the accuracy and interpretability of ST data. For a comprehensive overview of ST analysis pipelines and computational frameworks, see Dries et al. [113]. A practical framework for selecting amplification bias mitigation strategies Numerous experimental and computational strategies have been developed to mitigate amplification-associated distortions in ST. To facilitate practical implementation, we propose a structured framework that integrates platform characteristics, sample quality, and RNA input levels to guide the selection of appropriate bias-mitigation strategies (Fig. 3). A decision tree for selecting amplification bias mitigation strategies in spatial transcriptomics. The first decision point concerns platform type. Imaging-based ST technologies are generally less susceptible to amplification-related distortions because transcript detection primarily relies on probe hybridization and localized signal amplification rather than extensive nucleic acid amplification. Consequently, technical performance is more strongly influenced by factors such as probe design, hybridization efficiency, and imaging quality. In contrast, sequencing-based ST platforms depend on reverse transcription and PCR amplification, making them inherently more susceptible to amplification-induced biases and therefore requiring more extensive bias-control measures. For sequencing-based workflows, universal mitigation measures, including UMI-based deduplication, PCR optimization, and routine quality control, should be regarded as baseline components of bias-control workflows. Additional interventions can be selected according to sample-specific risk factors such as RNA degradation or low transcript abundance. Regardless of the selected mitigation strategy, independent validation remains essential for assessing residual amplification bias and confirming the effectiveness of bias-reduction measures. Technical replicates, reference benchmarking datasets, spike-in controls where applicable, and post-correction performance assessment can provide complementary approaches for evaluating mitigation efficacy and ensuring the reliability of downstream biological interpretations. Emerging approaches in ST Emerging sequencing paradigms, such as long-read sequencing (LRS) technologies including Pacific Biosciences (PacBio) [114, 115] and Oxford Nanopore Technologies (ONT) platforms [116–118], mitigate amplification-associated bias by restructuring library architecture and sequencing strategies, enabling direct sequencing of full-length cDNA or RNA molecules while reducing reliance on short-read fragmentation and transcript assembly. Although amplification steps remain necessary for low-input samples, LRS preserves intact transcript structures within individual reads, partially decoupling amplification bias from transcript structure inference and thereby enhancing the reliability of isoform-level analyses. Recent proof-of-concept studies have further extended long-read sequencing to ST, integrating spatial barcodes to enable spatially resolved profiling of full-length transcript isoforms while minimizing the influence of amplification-induced artifacts on spatial expression patterns [119–121]. Despite current limitations in cost, throughput, and error rates, ongoing methodological and computational advances [122] position long-read sequencing as a promising complementary strategy for mitigating structural and transcript-level biases in amplified transcriptomic libraries. Commercial platforms such as 10× Genomics Xenium use targeted gene panels for spatial detection of predefined transcripts. Dual-arm (padlock) probes enable specific recognition of target transcripts, followed by signal amplification via RCA, which helps reduce amplification bias associated with RT- and PCR-based processes. Summary and outlook Over the past decades, the rapid advancement of ST has established a powerful framework for interrogating gene expression in spatial context within intact tissues. Owing to this capability, ST was recognized as the “Method of the Year 2020” by Nature Methods [123]. By preserving tissue architecture, ST enables precise localization of cell types and characterization of spatial gene expression patterns. This further reveals local cellular ecosystems and spatiotemporal dynamics of cellular development [124]. Despite these advances, the fidelity and reproducibility of ST data remain influenced by multiple experimental and analytical factors, including spatial indexing or barcoding strategies [14, 15], RNA capture efficiency [21], cross-tissue and cross-individual integration [125]. The amplification step in ST is necessary to generate sufficient signal from low-input RNA. It also introduces systematic bias due to the inherently low and heterogeneous starting material. However, amplification bias introduced during ST library preparation and sequencing has received limited systematic attention in the literature, despite being an important source of technical variation in ST workflows. Reverse transcription and PCR amplification do not correct deficiencies in template abundance or restore transcript complexity; instead, they further amplify subtle differences present in the input material, ultimately leading to substantial distortions in transcript abundance and spatial expression patterns. Unlike bulk RNA-seq, sequencing-based ST captures RNA at the interface between tissue sections and the capture surface through sequence-specific probe hybridization. The resulting RNA input is therefore shaped by tissue architecture, permeabilization efficiency, and probe accessibility, leading to low and heterogeneous template populations. Reverse transcription and PCR do not restore missing transcripts or correct input imbalances, but instead amplify preexisting differences in the captured material, resulting in distortions in transcript abundance and spatial expression patterns. During the reverse transcription stage, transcript sequence-dependent features, such as GC content, length, and local secondary structure, collectively influence reverse transcription efficiency and the integrity of cDNA products. Transcripts with moderate GC content and relatively weak secondary structures are more likely to be captured in a full-length form. Engineered RTs have been developed to address these limitations through multiple strategies. These include improving the ability to resolve stable secondary structures, enhancing processivity, and reducing RNase H activity. The reduction in RNase H activity helps prevent premature RNA degradation and improves template stability. These improvements have, to some extent, reduced sequence- and structure-dependent biases in reverse transcription and enhanced the overall uniformity of transcript capture. In PCR, amplification bias is primarily driven by template competition, enzymatic kinetic preferences and stochastic effects during early amplification cycles, and is further modulated by cycling conditions and instrument-specific parameters. High-fidelity DNA polymerases, improved primer design, and reduced PCR cycle numbers help decrease differences in amplification efficiency among templates. Reverse transcription and PCR should not be regarded as “black boxes” in ST workflows, as they are not simply a linear conversion from RNA input to DNA output, nor a proportional amplification from DNA input to DNA product. They represent complex biochemical processes whose behavior directly shapes the quality and interpretability of ST data. Systematic investigation of their underlying mechanisms and influencing factors is therefore essential, even when focusing specifically on reverse transcription and PCR themselves rather than the complete ST workflow. Linear amplification has been considered as a potential alternative to PCR, since it can better preserve relative transcript abundance and reduce the accumulation of amplification errors. However, it generally exhibits lower sensitivity than PCR and may generate substantial by-products, which should be carefully considered in practical applications. In contrast, compartmentalized amplification strategies reduce template competition and improve library uniformity by physically isolating template molecules. Such strategies are expected to be integrated into library preparation workflows or microfluidic systems. For samples with a high risk of bias, increasing technical replicates can, to some extent, improve the robustness of estimates and reduce the impact of stochastic technical variation on downstream biological inference. At the experimental design level, the incorporation of spike-in controls is a commonly used strategy for bias assessment and normalization. These controls can be used to characterize systematic bias and provide an external reference scale for normalization and correction. ERCC and SIRV spike-ins are complementary in their design: ERCC covers a wide dynamic range of synthetic transcripts with varying concentrations and lengths, whereas SIRV is primarily designed to model the structural complexity of transcript isoforms arising from alternative splicing. Their combined use can enhance calibration across different expression levels and transcript feature spaces. It should be noted that the input proportion of spike-in RNA should be carefully controlled to avoid disproportionate consumption of sequencing reads, which may otherwise reduce the detection sensitivity for low-abundance endogenous transcripts. Computational correction also plays an important role, particularly when experimental optimization alone is insufficient. UMI-based molecular barcoding and algorithms enable the PCR-derived duplicates to be collapsed, thereby improving the transcript quantification accuracy. Advances in statistical modeling and machine learning have further enabled the correction of systematic biases in sequencing data, including fragment length effects, GC-content bias, and amplification artifacts. Current normalization and standardization methods used in ST are largely adapted from scRNA-seq. These approaches often fail to adequately capture the spatially structured and tissue-dependent sources of variation in ST data, where library size and amplification-related technical effects may be confounded with spatial dependence, heterogeneous tissue architecture, and local differences in cellular composition. Systematic evaluation of spatial distribution patterns, amplification bias, and normalization strategies will be necessary to establish best practices in ST analysis. Promisingly, several ST-specific data correction tools are gradually being developed, including SpotSweeper, SpaNorm, and SpatialMAGIC. By integrating spatial histology with gene expression in a multimodal framework, technical biases can be further mitigated in complex tissue environments, enabling context-aware analysis of molecular expression. Overall, this review focuses on the mechanisms underlying amplification bias formation during ST library preparation, its consequences for data quality, and potential strategies for its mitigation and correction. Improving the robustness and reliability of ST requires the integration of optimized experimental designs with bias-aware computational frameworks, ultimately enabling more accurate interpretation of spatial gene expression and tissue organization. Key Points Spatial transcriptomics (ST) enables high-resolution mapping of gene expression within tissue architecture but is affected by amplification bias introduced during library preparation. Spatial transcriptomics (ST) enables high-resolution mapping of gene expression within tissue architecture but is affected by amplification bias introduced during library preparation. Amplification bias arises from multiple stages of the experimental workflow, including reverse transcription, polymerase chain reaction amplification, and library construction, leading to cumulative distortions in transcript representation. Amplification bias arises from multiple stages of the experimental workflow, including reverse transcription, polymerase chain reaction amplification, and library construction, leading to cumulative distortions in transcript representation. This systematic review of amplification bias mechanisms and existing mitigation strategies provides a useful reference for optimizing ST experimental design and for guiding the development of improved bias correction methods. This systematic review of amplification bias mechanisms and existing mitigation strategies provides a useful reference for optimizing ST experimental design and for guiding the development of improved bias correction methods. Future advances require integrative frameworks that jointly optimize experimental design and computational modeling to systematically reduce amplification-induced artifacts. Future advances require integrative frameworks that jointly optimize experimental design and computational modeling to systematically reduce amplification-induced artifacts. Supplementary Material