A sequence physically shaped hardware acceleration circuit

Abstract

The application discloses a sequence function annotation physical shape matching hardware acceleration circuit, and belongs to the technical field of digital integrated circuit and sequence data processing. The circuit comprises a sequence receiving and coding unit, a physical shape parameter lookup table unit, a sliding window signature unit, a mode prototype storage unit, a parallel matching array and a threshold value determination and label output unit. The physical shape parameter lookup table unit maps a sequence stream into shape parameters position by position; the sliding window signature unit outputs physical shape signatures through blocking and normalization; the parallel matching array calculates the similarity between the signatures and prototypes in parallel; and the threshold value determination and label output unit outputs a matching mode ID, a physical label and a matching degree. The mode prototype storage adopts double buffers, and supports dictionary hot loading zero waiting switching. The application realizes sequence physical shape matching with four period fixed delay, and can be integrated into a sequence function annotation chip for gene sequencing, proteomics, chemical informatics and the like.

Description

Technical Field

[0001] This invention belongs to the field of digital integrated circuits and sequence data processing technology. Specifically, it relates to a hardware acceleration circuit for physical shape extraction, matching, and interpretable tag output of linear sequence streams. It can be integrated as a standardized IP core into a high-throughput functional labeling chip for real-time functional determination of sequence segments without relying on known functional modes. Background Technology

[0002] In high-throughput functional annotation scenarios, determining the function of sequence fragments is a fundamental operation. This type of operation is widely used in fields such as genome sequence functional annotation, protein structure and function prediction, RNA secondary structure analysis, and chemical molecule functional classification. Each sequence data point processed by a microarray requires functional determination of the window fragment, which is a core bottleneck in the high-throughput analysis workflow.

[0003] There are currently three main ways to implement sequence function annotation:

[0004] (I) Matching based on known functional patterns. Known functional patterns (DNA / RNA motifs, protein domains, chemical substructures) are statistically represented as a weight matrix. Matching scores are calculated window-by-window for the input sequence and compared with a threshold. The limitation of this approach is that the functional pattern matching model loses the physical structural information corresponding to the sequence fragments—the shape characteristics of the DNA double helix (minor groove width, propeller twist, etc.), the dihedral conformation of the protein backbone, and the three-dimensional shape characteristics of the molecule—resulting in extremely weak predictive power for fuzzy functional patterns. For example, the PWM prediction correlation coefficient for DNA transcription factors TP53 is only about r≈0.15, indicating that its binding specificity comes from the DNA shape rather than the letter sequence.

[0005] (ii) Sequence pattern exhaustive matching. Perform fragment pattern exhaustive or fuzzy matching on the sequence. The limitations of this approach are: the functional sequence space is extremely sparse, the exhaustive space explodes exponentially with the fragment length, and the matching results have no physical semantics—it cannot explain "why this sequence fragment has this function".

[0006] (III) Software-layer physical shape prediction. Published physical shape feature systems have been proven to significantly improve functional prediction, but all of them are executed at the software layer. Software-layer execution requires extracting shape signatures for each sample and matching them with dictionary prototypes. The single-window processing latency is in the microsecond to millisecond range, and in high-throughput scenarios (millions to billions of windows in a single run), the accumulated latency is on the order of hours, which cannot meet the requirements of real-time annotation.

[0007] The above three solutions have the following drawbacks:

[0008] Defect 1: Existing matchers rely on known functional patterns (sequence motifs / structural domains), losing physical shape information. Physical shape determines function, but existing hardware lacks circuitry that matches based on physical shape rather than letter sequence.

[0009] Defect 2: While physical shape features have been proven to significantly improve functional prediction, this improvement remains at the software level. The computational cost of shape signature extraction and multi-prototype matching becomes the throughput bottleneck for high-throughput real-time annotation, rendering the software solution ineffective in high-throughput scenarios.

[0010] Defect 3: Existing matchers output binary "match / not match" or abstract scores without physically interpretable labels (such as "small groove width narrows" or "high alpha spiral tendency"). Downstream components cannot directly obtain the mechanism semantics and require an additional interpretation layer.

[0011] Defect 4: After the functional mode dictionary is updated, the existing solution needs to pause the pipeline for recompilation or reconfigure the hardware. It lacks a dictionary hot-loading mechanism and is incompatible with the timing of real-time data stream processing.

[0012] In summary, there is currently no dedicated hardware circuit in the chips that can extract physical shape signatures from the input sequence stream, match them with the physical shape dictionary, and output physically interpretable tags—all without relying on known functional modes, and supporting online hot updates of the dictionary. Summary of the Invention

[0013] The technical problem to be solved by the present invention

[0014] This invention aims to address multiple technical bottlenecks in existing sequence function annotation hardware architectures. Specific technical problems are as follows:

[0015] (i) Sequence function annotation relies on known function patterns (motifs / domains). Sequences with fuzzy function patterns and no known patterns cannot be reliably annotated, and there is a lack of hardware circuits that match the physical shape.

[0016] (ii) Physical shape signature extraction and dictionary matching remain at the software level, and the computational load becomes the throughput bottleneck for high-throughput real-time annotation;

[0017] (iii) The existing matcher has no physically interpretable label output, and the downstream cannot obtain the mechanism semantics;

[0018] (iv) Updating the functional mode dictionary requires pausing pipeline reconfiguration and lacks a hot reloading mechanism.

[0019] Inventive concept

[0020] This invention includes the following technical solution: using a physical shape pattern dictionary as configuration input, and through a fixed-delay pipeline consisting of a parameter lookup table, a sliding window signature unit, a matching array, a threshold determination unit, and a tag output unit, the entire chain of "sequence → physical shape signature → dictionary matching → physical tag" is hardware-based.

[0021] This includes the following three aspects:

[0022] (i) Hardware extraction of physical shape signature. The sequence stream is mapped position by position to physical shape parameters using a parameter lookup table. After sliding window segmentation and intra-block normalization (z-score), the physical shape signature vector is output. The signature encodes a shape pattern (relative fluctuation) rather than an absolute value, which is insensitive to signal amplitude but sensitive to shape, and is suitable for various sequence types.

[0023] (ii) Multi-prototype parallel matching hardwired. The signature is matched in parallel with N pattern prototypes using a matching array (parallel multiplication tree + norm unit + argmax selector), and the maximum matching degree S is output. The matching result is given by comparing the maximum matching degree S with the configurable threshold θ_c.

[0024] (iii) Dictionary hot loading. The pattern prototype storage uses double-buffered SRAM. After the new dictionary is written to the background buffer, the loading selection signal is flipped and switched with zero wait. After switching, the system automatically enters the observation mode (with tolerance relaxed within a predetermined number of steps) to prevent misjudgment due to instantaneous fluctuations in the early stage of loading.

[0025] Overall technical solution

[0026] This invention implements end-to-end hardware-based sequence physical shape matching, replacing the traditional software-based item-by-item calculation scheme with a deterministic delay of four clock cycles, achieving single-step matching and physical tag output of sequence segments after dictionary configuration. Through a four-stage fixed-delay pure hardware pipeline, using parameter lookup tables, sliding window signature units, matching arrays, threshold comparators, and tag encoders as core components, the following matching chains are solved in hardware parallelism:

[0027] (1) Physical shape parameter extraction: Look up the parameter table position by position in the input sequence stream and output the physical shape parameters;

[0028] (2) Shape signature normalization: Sliding window block aggregation, intra-block normalization, output signature vector;

[0029] (3) Parallel matching: The signature is parallelly multiplied by the dot product / norm of N pattern prototypes, and the matching degree and the maximum matching degree S are output.

[0030] (4) Threshold determination and label output: S and The comparison provides a judgment. The argmax mode ID is used to look up the physical label table, and the physical label plus the matching degree is output.

[0031] Specific technical solutions

[0032] (a) Sequence receiving and encoding unit. Used to receive the input sequence stream, encode the sequence symbols into parallel lookup indices, and output them to the parameter lookup table according to the pipeline cycle; includes a sequence shift register to support position-by-position advancement of the sliding window.

[0033] (ii) Physical Shape Parameter Lookup Table Unit. Connected to the sequence receiving and encoding unit, this unit stores the physical shape parameters of the sequence fragments and outputs the shape parameters position by position for the input sequence stream. The parameter table is a read-only ROM or a writable SRAM, loaded according to the dictionary version. For DNA sequences, the parameters include four parameters: MGW / ProT / Roll / HelT (expanded to 1024 items through reverse complementation, each 8 bits fixed-point); for protein sequences, the parameters include φ / ψ dihedral angles; for molecular sequences, the parameters include volume / charge / hydrophobic characteristics.

[0034] (III) Sliding Window Signature Unit. Connected to the parameter lookup table unit, the physical shape parameters are aggregated in blocks according to the sliding window, and normalization is performed within each block (mean and standard deviation hardware statistical unit), outputting the physical shape signature vector. The normalization makes the signature encoding shape pattern (relative fluctuation) rather than an absolute value.

[0035] (iv) Pattern Prototype Storage Unit. Connected to the sliding window signature unit, the SRAM stores N physical shape pattern prototype vectors, which can be updated online; a dual-buffer structure is adopted, the main waterline is not affected during the writing of the new dictionary to the background buffer, and the load selection signal is switched with zero wait after the writing is completed; a dictionary version register is configured, and the threshold determination automatically enters the observation mode after switching to prevent misjudgment due to instantaneous fluctuations in the early stage of loading.

[0036] (v) Parallel matching array. Connect the sliding window signature unit to the pattern prototype storage unit, and calculate the similarity between the signature vector and N prototype vectors in parallel. The dot product is implemented using a parallel multiplication tree. The norm is approximated by summing squares and taking the square root. The output is N matching degrees and the maximum matching degree S (argmax selector), with single-cycle output.

[0037] (vi) Threshold determination and tag output unit. Connect the parallel matching array and set the maximum matching degree S with the configurable threshold. Compare: → Determine "matching pattern i", and output the matching pattern ID, matching degree S, and matching result; → Output a "No Match" bypass signal. Look up the physical tag table using the matching mode ID (mode ID → physical description) and output the physical tag. The threshold value... Adjust by configuring registers.

[0038] Basic pipeline timing rules: Physical shape parameter extraction and sliding window aggregation are completed in cycle 1 by LUT lookup and shift register; normalization is completed in cycle 2 by the mean / standard deviation statistical unit; parallel matching is completed in cycle 3 by the multiplication tree and norm unit; threshold determination and tag output are completed in cycle 4 by the comparator and encoder. The entire operation is completed in four fixed clock cycles, with no branches, no arbitration waits, and no software intervention.

[0039] Beneficial effects

[0040] (i) This invention uses physical shape signatures instead of letter sequence matching, without relying on known functional patterns—sequences without known functional patterns can be captured by shape patterns, and fuzzy functional patterns can be reliably distinguished by shape patterns. Experimental data shows that with a small number of labeled samples (10 samples / class), the shape matching accuracy is 0.643, and the letter PWM matching accuracy is 0.373, with shape matching outperforming letter matching by approximately 0.27.

[0041] (ii) Compared to software layer shape prediction (single-window signature and matching approximately 5176ns), this invention outputs matching results with a fixed four-clock-cycle output (4ns at 1GHz), achieving a single-window speedup of approximately 1294 times. In high-throughput scenarios, the software end-to-end latency increases linearly with the number of windows (10... 9 (Window approximately 1.4 hours), once the hardware pipeline is full, throughput remains constant, outputting one matching result per cycle (10). 9 The window pipeline computation takes about 1 second, with an end-to-end speedup of about 5000 times (the hardware processing time is the same as the main pipeline computation time; this circuit is integrated into the main chip as an IP core, and data passes directly to the pipeline through the on-chip bus without the need for additional data transfer).

[0042] (iii) This circuit has a scale of approximately 19k gates and is implemented using ASIC with a power consumption of mW. It can be deployed in edge devices (nanopore sequencers, point-of-care diagnostic devices, portable analyzers) and has an energy efficiency improvement of 3-4 orders of magnitude compared to CPUs (tens of W) and GPUs (300W+).

[0043] (iv) Output physically interpretable labels and quantitative matching degree - downstream can directly obtain the mechanism semantics and can be connected to the subsequent analysis process, rather than binary "match / mismatch" or abstract scoring.

[0044] (v) Matching array dimensions are configurable—the same hardware can switch between DNA physical shape signature (62-dimensional), protein dihedral signature (2-dimensional), RNA base pair geometric signature, and molecular shape signature (5-dimensional), enabling cross-domain reuse of unified physical shape matching hardware.

[0045] (vi) Dual buffers support dictionary hot loading. During the writing of the new mode to the background buffer, the mainstream waterline is completely unaffected, the switching is zero-wait, and the observation mode after switching prevents misjudgment.

[0046] Terminology Definition

[0047] Sequence stream: refers to linear sequence data such as DNA, RNA, proteins, and chemical molecules, which are composed of discrete symbols.

[0048] Physical shape signature: refers to the vector obtained by extracting physical shape parameters and normalizing the sequence fragment through sliding window, which encodes the shape pattern (relative fluctuations) rather than the absolute value.

[0049] Physical shape parameters: These refer to the physical structural characteristics of a sequence fragment. For DNA, these include minor groove width (MGW), propeller twist (ProT), roll angle, and helical twist (HelT); for proteins, these include the dihedral angle of the main chain. For RNA, this includes base pair geometry parameters; for molecules, it includes volume, charge distribution, and hydrophobicity.

[0050] Physical shape pattern dictionary: refers to the set of physical shape feature prototypes obtained by clustering from known functional sequences. Each prototype contains a signature vector and a physically interpretable label.

[0051] Matching degree S: refers to the similarity between the signature and the pattern prototype, with a value range of [-1, 1].

[0052] Physical labels: refer to interpretable mappings from pattern IDs to physical descriptions (such as "CTCF dominant, narrowing groove width" or "high α-spiral tendency"), which include shape semantics.

[0053] Observation mode: refers to the working mode that relaxes the tolerance of threshold judgment within a predetermined number of steps after dictionary hot loading, to prevent misjudgment due to instantaneous fluctuations in the early stage of loading. Attached Figure Description

[0054] Figure 1 This is a schematic diagram of the overall architecture of the hardware acceleration circuit for sequence physical shape matching of the present invention. The left side of the diagram shows the sequence stream input, the top of the diagram shows dictionary loading, and the right side of the diagram shows the physical tag and matching degree output; the solid arrow indicates the direction of the forward data flow; the dotted-dash arrow indicates the configuration control signal; the circled labels ① to ⑥ indicate the six modules in sequence (① sequence receiving and encoding unit, ② physical shape parameter lookup table unit, ③ sliding window signature unit, ④ pattern prototype storage unit, ⑤ parallel matching array, ⑥ threshold determination and tag output unit); the four-stage pipeline boundaries are separated by vertical dashed lines and labeled T1 / T2 / T3 / T4, where ① and ② are executed in the T1 stage pipeline, ③ in the T2 stage, ⑤ in the T3 stage, ⑥ in the T4 stage, and ④ is the configuration storage unit located outside the pipeline. Detailed Implementation

[0055] Example 1: Determination of DNA transcription factor binding sites and hardware resources

[0056] like Figure 1 The overall architecture shown in this embodiment uses DNA CTCF binding site determination as an example to verify the functional correctness and timing performance of the matching circuit.

[0057] Configuration parameters: Physical shape pattern dictionary (39 prototypes, 62-dimensional fixed points), threshold The input is a 19bp sequence window (CTCF binding site).

[0058] Pipeline timing:

[0059] Period T1: The sequence receiving and encoding unit encodes the 19bp window as a lookup index; the parameter lookup table unit outputs the four parameters MGW / ProT / Roll / HelT position by position (15 sliding window positions × 4 parameters, each 8 bits).

[0060] Period T2: The sliding window signature unit completes the four-block partitioning (MGW 15+ProT 15+Roll 16+HelT 16=62 dimensions) and intra-block z-score normalization, outputting a 62-dimensional signature vector.

[0061] Period T3: The parallel matching array will perform a parallel dot product / norm with the signature and 39 prototypes, outputting a maximum matching degree S=0.941.

[0062] Period T4: Threshold Decision Unit Comparison , determine "match"; argmax outputs the mode ID; the label output unit outputs the physical label and the matching degree S=0.941.

[0063] Control: After a single-base mutation (geometric critical site mutation) within the same window, S dropped to The result is "mismatch".

[0064] Actual results: Software-level single-window processing takes about 0.1~1ms, while this invention outputs a fixed four-clock-cycle (4ns) output, resulting in a hardware speedup of about 5 orders of magnitude; the total latency is still 4 cycles when processing 1024 windows in parallel.

[0065] Hardware resource and timing verification: Based on the 28nm standard process and 1GHz main frequency, the equivalent gate count of each module in this circuit is estimated as follows: sequence reception and encoding unit approximately 0.3k gates, physical shape parameter lookup table approximately 4k gates, sliding window signature unit approximately 2k gates, pattern prototype storage (double buffer) approximately 2k gates, parallel matching array approximately 10k gates, threshold determination and tag output approximately 0.5k gates, totaling approximately 19k gates. The pipeline critical path consists of four levels of combinational logic: LUT lookup, normalization statistics, multiplication tree, and comparator. Static timing analysis shows sufficient establishment and hold margins, with no timing violations.

[0066] Example 2: Motionless Site Annotation and Matching Grading

[0067] This embodiment verifies the physical shape annotation and matching degree classification output of weak motif transcription factor (without reliable PWM) sites.

[0068] Configuration parameters: dictionary The input is the TP53 response element candidate window (its PWM predictive power r≈0.15, and the sequence matching cannot be determined).

[0069] Matching process: The candidate window passes through the T1~T3 pipeline to obtain the maximum matching degree S=0.62; the T4 threshold judgment S=0.62≥ 0.5, and the matching mode and physical label are output.

[0070] Matching degree graded output (threshold register grading): When S≥0.7, output high confidence match (pattern ID, tag, and high confidence flag); when 0.5≤S<0.7, output medium confidence match (pattern ID, tag, and medium confidence flag); when 0.3≤S<0.5, output low confidence or edge match (pattern ID and low confidence flag, for downstream verification); when S<0.3, output no match bypass signal "unmarked".

[0071] This embodiment illustrates that even if the PWM cannot determine the location, the physical shape matching circuit can still output an interpretable label.

[0072] Example 3: Rapid Pre-annotation of Protein Secondary Structure

[0073] This embodiment verifies the application of the same circuit structure in the pre-annotation of protein secondary structures.

[0074] Configuration parameters: Input amino acid sequence window (19 residues), physical shape parameters are as follows. Dihedral angle (2D), the prototype is based on a known secondary structure ( spiral, The prototypes were obtained by clustering (folded and randomly curled), with 8 prototypes in each class.

[0075] Pipeline timing:

[0076] Period T1: The sequence receiving and encoding unit encodes the 19-residue window as a lookup index; the parameter lookup table unit outputs position by position. Dihedral angle.

[0077] Period T2: The sliding window signature unit completes dihedral sliding window aggregation and intra-block normalization, and outputs a 2D signature vector.

[0078] Period T3: The parallel matching array performs a parallel dot product / norm of the signature and the prototype, and outputs the maximum matching degree S.

[0079] Period T4: Threshold determination unit compares S with Output the secondary structure labels and matching degree.

[0080] The comparison and verification results with the letter PWM scheme are as follows: When the number of labeled samples per class is 10, the physical shape matching accuracy is 0.643, while the letter PWM accuracy is 0.373; when the number of labeled samples per class is 20, the physical shape matching accuracy is 0.565, while the letter PWM accuracy is 0.341; when the number of labeled samples per class is 40, the physical shape matching accuracy is 0.626, while the letter PWM accuracy is 0.401; when the number of labeled samples per class is 80, the physical shape matching accuracy is 0.573, while the letter PWM accuracy is 0.393. Physical shape matching is superior to letter statistics in all aspects, and the advantage is more significant with a small number of samples. With 10 samples, physical shape matching is about 0.27 times better than letter PWM. In the cross-protein generalization test, the accuracy for leave-one-protein tests ranges from 0.42 to 0.84, with a mean of about 0.57.

[0081] Application scenarios: This circuit can provide enhancement of local physical shape features of sequences at the input of structural prediction models such as AlphaFold, or provide fast structural prediction when homologous sequences (MSA) cannot be obtained, serving as a front-end supplement to the MSA route.

[0082] Example 4: RNA and Molecular Shape Labeling

[0083] This embodiment verifies the feasibility of the same circuit structure for RNA and chemical molecule sequence types.

[0084] RNA base pairing annotation:

[0085] The input is a three-dimensional RNA atomic flow, with the physical shape parameter being the Watson-Crick standard atomic pair spatial distance. AU pairings contain two hydrogen bonds: N1↔N3 and N6↔O4; GC pairings contain three hydrogen bonds: N1↔N3, N2↔O2, and O6↔N4; and GU wobbling pairings contain two hydrogen bonds: N1↔O2 and O6↔N3. A four-cycle pipeline sequentially performs atomic coordinate extraction, standard atomic pair distance calculation (atom pairs are locked using a lookup table, and distance calculations are performed using a subtractor and square root approximation unit), distance window determination and non-adjacent constraint (an empirical window of 2.3 to 3.8 Å, with the absolute value of the residue difference greater than 1 excluding false positives from adjacent base stacking), hydrogen bond counting, and output of pairing type labels (three hydrogen bonds correspond to GC pairings, and two hydrogen bonds correspond to AU or GU pairings). Using yeast tRNA^Phe (PDB 1TRA) as the benchmark, the base pairing type discrimination accuracy is 0.80 to 0.87, compared to 0.54 for the control pure distance threshold method. Standard atom pair locking is implemented by a lookup table ROM, distance calculation is completed by a subtractor and a square root approximation unit, distance window determination is implemented by a comparator, non-adjacent constraints are implemented by counting logic, and hydrogen bond counting synthesis is completed by an adder tree.

[0086] Molecular shape annotation:

[0087] The input is a molecular structure data stream, with the physical shape parameter being the eigenvalue ratio of the inertial tensor. The inertial tensor is a 3x3 matrix; three similar eigenvalues ​​correspond to spherical molecules, two similar eigenvalues ​​with one much smaller correspond to disk-shaped molecules, and one eigenvalue much larger than the other two correspond to rod-shaped molecules. A four-cycle pipeline sequentially completes the extraction of 3D atomic coordinates of the molecules, inertial tensor calculation (nine multiplications and additions), shape factor lookup table determination (eigenvalue ratios are mapped to three shape categories: spherical, disk-shaped, or rod-shaped), and physical shape label output. Using fifteen real drug molecules as a test set, the prediction accuracy is 0.80, and the majority class baseline is 0.67.

[0088] Unified benefits: Standard atom pairs and inertial tensors are physical priors and can work without labeled samples; pairing type and shape labels are physically interpretable labels; the same four-cycle pipeline can switch from DNA to RNA and molecular types simply by rewriting the lookup table content and computational dimensions.

[0089] Example 5: Dictionary Hot Loading

[0090] This embodiment verifies the zero-wait switching capability of the dual buffers during runtime dictionary update.

[0091] Configuration parameters: The initial dictionary is loaded into Buffer A. At step N, a new mode write is triggered, and the new prototype is written to Buffer B, completing in two cycles. The load selection signal toggles, and the matching array automatically switches to Buffer B in the next clock cycle, with zero waiting time. After switching, the threshold determination automatically enters observation mode, reducing the sensitivity of the matching determination within a predetermined number of steps to prevent misjudgments due to transient fluctuations in the initial stage of new mode loading.

[0092] Key point: The main waterline is completely unaffected during dictionary updates and continues to advance matching using the old prototype in Buffer A.

Claims

1. A hardware acceleration circuit for physical shape matching of sequence function annotation, characterized in that, include: The sequence receiving and encoding unit is used to receive the input sequence stream, encode the sequence symbols into a parallel lookup index, and output in a single cycle. The physical shape parameter lookup table unit, connected to the sequence receiving and encoding unit, is used to store the physical shape parameters of the sequence segments and outputs the physical shape parameters position by position for the input sequence stream, with a single-cycle output. The sliding window signature unit is connected to the physical shape parameter lookup table unit. It is used to aggregate the physical shape parameters by sliding window blocks, normalize the parameters within each block, and output the physical shape signature vector in a single cycle. The pattern prototype storage unit, connected to the sliding window signature unit, is used to store N physical shape pattern prototype vectors. It adopts a dual buffer structure, supports online hot loading of the dictionary, and outputs all prototype vectors in a single cycle. A parallel matching array, connecting the sliding window signature unit and the pattern prototype storage unit, is used to calculate the similarity between the signature vector and N prototype vectors in parallel, and output N matching scores and the maximum matching score, with single-cycle output. The threshold determination and tag output unit is connected to the parallel matching array and is used to compare the maximum matching degree with the configurable threshold, output the matching mode ID, physical tag and matching degree determination, and output in a single cycle. The entire circuit, from sequence input to tag output, is fixed at four pipeline clock cycles. All operational paths are pure combinational logic architectures, with no software scheduling and no microcode operations.

2. The physical shape matching hardware acceleration circuit for sequence function labeling according to claim 1, characterized in that, The physical shape parameter is of any of the following types: DNA sequence minor groove width (MGW), propeller twist (ProT), roll angle (Roll), and helical twist (HelT). Main chain dihedrals of protein sequences ; Base pair geometry parameters of RNA sequences; The size, charge distribution, and hydrophobicity of chemical molecules.

3. The physical shape matching hardware acceleration circuit for sequence function labeling according to claim 1, characterized in that, The normalization of the sliding window signature unit is implemented using hardware statistical units of the mean and standard deviation within the block, making the signature encoding shape pattern (relative fluctuations) rather than an absolute value.

4. The physical shape matching hardware acceleration circuit for sequence function labeling according to claim 1, characterized in that, The pattern prototype storage unit adopts a dual-buffer structure, including a first buffer and a second buffer. The current match uses the prototype vector in one of the buffers, and the new dictionary can be written to the other buffer without affecting the ongoing match. After the write is completed, the load selection signal is toggled, and the match array automatically switches to the new dictionary in the next clock cycle. The switching process does not require pausing the pipeline.

5. The physical shape matching hardware acceleration circuit for sequence function labeling according to claim 4, characterized in that, The threshold determination unit automatically enters the observation mode after the loading selection signal flips, and reduces the sensitivity of the matching determination within a predetermined number of steps to prevent misjudgment due to instantaneous fluctuations in the early stage of new dictionary loading.

6. The physical shape matching hardware acceleration circuit for sequence function labeling according to claim 1, characterized in that, The parallel matching array includes: Parallel dot product multiplication tree, used to calculate the inner product of the signature vector and each prototype vector; Norm unit, used to calculate the norm of the signature vector and the norm of each prototype vector; The argmax selector is used to select the maximum value from N matching degrees.

7. The physical shape matching hardware acceleration circuit for sequence function labeling according to claim 1, characterized in that, The configurable threshold of the threshold determination unit Adjust by configuring registers.

8. The physical shape matching hardware acceleration circuit for sequence function labeling according to claim 1, characterized in that, The threshold determination unit is based on the maximum matching degree S and... The output classification is determined after comparison: S≥0.7 outputs a high confidence flag, 0.5≤S<0.7 outputs a medium confidence flag, 0.3≤S<0.5 outputs a low confidence flag, and S<0.3 outputs an unmatched bypass signal.

9. The hardware acceleration circuit for physical shape matching of sequence function annotation according to claim 1, characterized in that, The physical tag is an interpretable mapping table from pattern ID to physical description, implemented in read-only memory or writable SRAM, and loaded with dictionary version.

10. The physical shape matching hardware acceleration circuit for sequence function annotation according to claim 1, characterized in that, The functions of each stage of the four-stage production line are as follows: T1 level: The sequence receiving and encoding unit outputs the lookup index, and the physical shape parameter lookup table unit outputs the shape parameter position by position; Level T2: The sliding window signature unit completes block division and normalization, and outputs the signature vector; T3 level: Parallel matching array completes parallel dot product, norm and argmax selection, and outputs the maximum matching degree; Level T4: The threshold determination unit completes threshold comparison and classification determination, and the label output unit outputs the physical label and matching degree.

Images