1. Most readers would be confused about the novelty or advances of NIFFLR. It’s based on the alignment of constructed exon-exon junction sequences, a strategy used by many tools long ago. Does this strategy work for mini-exons that are shorter than 10 nt or 20 nt? This strategy also heavily depends on annotated exons. Can it identify any novel splice donors or acceptors? For tissues or organisms with incomplete annotation, how will this strategy be affected? All of above need be carefully evaluated and described in details.

Response: The novelty of NIFFLR lies in its approach of constructing transcripts from existing exons guided by long reads. NIFFLR shows significant improvements in sensitivity and quantification accuracy compared to existing tools. Since NIFFLR relies on existing exons, it cannot identify novel donor or acceptor sites; however, it can detect novel introns, i.e. novel donor-acceptor pairs. As we state in the manuscript, NIFFLR is designed for genomes with high-quality annotations, such as the human genome. NIFFLR strategy works well for exons that are longer than the minimum alignment K-mer size (default value is 12).  Any exon shorter than this value will not be detected in the alignment and will likely be skipped. Fortunately, the percentage of transcripts containing such mini-exons is small.  Human RefSeq annotation (mail chromosomes only) contains 379021 exons, and only 598 are shorter than 12 bp.

2. The aligner (psa_aligner) used by the authors needs comprehensively assessment, e.g. whether psa_aligner is suitable for noisy long-read data. An evaluation and comparison with the commonly used long-read aligner Minimap2 should be included.

Response: psa_aligner is not a novel tool. It is used in the MaSuRCA assembler [REF] to map super-reads built from Illumina reads to noisy long reads. Evaluating or comparing alignment methods was not the focus of our study.  In our tests we found that Minimap2 lacked sufficient sensitivity, particularly by missing alignments of short exons. We decided to use instead psa_aligner which only produces pseudo-alignments.

3. The NIFFLR programming script is poorly written and difficult to install and use for analysis. For example, when I tried to run NIFFLR, I received an error message from python (not NIFFLR) and could not finished the analysis. The error, “IndexError: list index out of range”, occurred during the “performing filtering and quantification” step.

Response: Thank you for raising this concern and for sharing the error message. We reviewed the relevant part of the NIFFLR pipeline but we were unable to replicate the issue. Based on the error message reported “ IndexError: list index out of range” during the filtering and quantification step, the problem is likely due to unexpected formatting in the reference GFF/GTF file. The quantification.py script assumes that each line in the reference file follows the standard GFF/GTF structure, containing at least 9 tab-separated fields, with the final (ninth) field being a semicolon-delimited string of attributes (e.g., with the transcript ID as the first entry). If any line deviates from this structure, it can trigger an IndexError.

To improve stability and usability, we have significantly reworked the pipeline and released a new version, v2.0.0, which is described in the revised manuscript.  This new version is easier to install and use and provides improved compatibility with input files that may slightly deviate from standard GTF/GFF specifications.

4. The authors used simulated data to evaluate NIFFLR and other tools. However, this evaluation may result in biased results. Many reports have noted that existing simulators are not suitable for evaluation. For example, according to a recent paper published by Dr. Hagen Tilgner and colleagues (Mikheenko A, et al., 2022 [Ref 1]), NanoSim randomly selects a starting position in a transcript to simulate truncation based on a uniform distribution. However, in real long-read data, a uniform distribution cannot be observed.

Response: We agree that simulated data have limitations and may not perfectly capture all features of real long-read sequencing. However, simulated data is the only way to evaluate software on datasets where the ground truth is fully known.  Nanosim is a peer-reviewed, widely used tool for generating simulated long reads, and its use is standard practice for benchmarking transcriptome assembly methods. Therefore, we believe it is appropriate to use NanoSim to produce simulated reads for the comparisons presented in this study.

5. A direct comparison of the tools for transcript identification and quantification using real long-read data with ground truth need be included. Many previous studies used SIRV E2 for evaluation, which contains 69 synthesized transcript isoforms with different abundances. I actually tried to run NIFFLR to analyze SIRV data, but it failed. NIFFLR produced an "exon extraction failed" error when processing the GTF annotation of SIRV, while I did not encounter any errors when using other tools. This is another example of poor programming of NIFFLR.

Response: We thank the reviewer for the suggestion. We have added an evaluation of the ability of different programs to identify isoforms in the SIRV E2 dataset to the Results section. The new NIFFLR version 2.0.0 now works correctly with the SIRV GTF file, resolving the previous “exon extraction failed” error.

6. I am stunned that the authors used transcript abundance from short-read data to evaluate the performance of long-read tools. The bias of short-read data has been so widely reported. For example, a Nature Methods paper published by Chen et al. 2025 [Ref 2] provided important evidence on this. In addition to the bias, the novel transcript isoforms in the data that are similar to annotated isoforms would also confuse short-read quantification.

Response: We understand why the reviewer is raising this objection, as indeed biases of short-read data have been widely reported. However, transcript quantification using short-read data has long been a standard in the field, and thousands of studies have successfully used short reads to quantify transcript expression (e.g. Romeo-Cardeillac et al., BMC Genomics, 2024; Kuehl et al., Nature, 2025; D'Sa et al., Sci. Adv., 2025). Widely used tools such as StringTie2 (Kovaka et al., Genome Biology, 2019), Salmon ( Patro et al., Nat Methods, 2017) and Kallisto (Bray et al., Nat Biotechnol, 2016) all rely on short reads for quantification. While it is true that short-read data can exhibit biases and that novel isoforms similar to annotated transcripts may complicate quantification, these limitations notwithstanding, short-read quantification still provides a reliable reference that can be used to assess the accuracy of long-read quantifications, as has been done in previous studies (Pardo-Palacios et al., 2023, Tang et al., 2020).  We therefore consider it appropriate to use short-read transcript quantification as a benchmark for evaluating long-read tools. This approach provides a practical and widely accepted reference for assessing quantification performance, while acknowledging that no method is entirely free of bias.

7. The authors need to provide sufficient evidence before arbitrarily judging existing tools. For example, they described ESPRESSO’s strategy for identifying novel splice junctions as “a stringent criterion that limits its ability to discover novel junctions”. However, they did not provide any evidence regarding the sensitivity of detecting novel splice junctions. Novel splice junctions can be further divided into junctions with novel splice sites and junctions as novel combination of annotated splice sites. Both types need to be compared between ESPRESSO and NIFFLR before the authors can draw such a conclusion.

Response:  We have removed the statement that ESPRESSO uses “a stringent criterion that limits its ability to discover novel junctions” from the introduction. For context, our statement that ESPRESSO uses “a stringent criterion that limits its ability to discover novel junctions” was not based on empirical benchmarking but rather on the algorithmic description provided in the ESPRESSO manuscript. As described in the manuscript, ESPRESSO classifies novel junctions into two types: junctions with novel splice sites (Novel Not in Catalog, NNC) and novel combinations of annotated splice sites (Novel In Catalog, NIC). According to the ESPRESSO manuscript, a novel transcript isoform is reported only if (i) each splice junction is supported by at least two perfectly aligned reads and (ii) the combination of junctions is not a substring of any other novel isoform. Perfectly aligned reads are defined as having no mismatches or indels within 10 nt of splice sites.   

In contrast, NIFFLR does not rely on spliced alignments and does not impose strict local alignment quality requirements. Instead, it uses a custom approximate aligner to map reference exons directly to long reads and assembles transcript models by optimizing exon tilings. This makes NIFFLR more tolerant of sequencing errors and more flexible in capturing novel combinations of annotated exons. Subsequent filtering and quantification steps in NIFFLR help exclude spurious transcripts. We note, however, that NIFFLR does not detect novel splice sites, as its transcript models are built from a predefined set of reference exons.

8. Why was BamBu (Chen Y, et al., 2023 [Ref 3]) not included in the evaluation and comparison? BamBu was published online more than two years ago and has been widely used for long-read data analysis. To demonstrate the advantages of their tool, the authors will need to compare it with BamBu.

Response: We added Bambu to all evaluations.

Major concerns:

1. It’s unclear why certain parameter choices were made for the implementation, such as k = 12 or coverage >= 35%, at the level of the exon to read alignment or Amin / Gmin at the level of choosing the most optimal path. The authors mention that they performed optimization but the results of these experiments are not included. Adding these results showing that this performance is optimal would increase the confidence in the tool and these decisions taken. Response: The default value of K=12 was empirically chosen as it provided the best balance between sensitivity and precision of the  psa_aligner on multiple data sets. Values of 11 or lower introduced numerous false positive alignments and slowed down NIFFLR without improving sensitivity, whereas values above 12 resulted in missed alignments of short exons.  In general, the optimal value of K depends on the read error rate and the minimum exon length in the annotation. For lower-error rate transcriptome sequencing data, K can be increased to 15–17 to achieve faster alignment. This information has been now added to our manuscript in section….

2. One of my main concerns is that, based on my understanding of the method, the only novel transcripts that can be discovered by NIFFLR are those that use already-annotated exons. This could substantially limit its ability to identify biologically relevant novel isoforms, which frequently involve novel splice sites or exons and are described in previous long-read RNA-seq studies. The authors should consider alternative methods to include novel splicing as part of their novel isoform discovery.

Response: We agree with the reviewer’s observation that using a known set of exons is a key limitation of NIFFLR.  NIFFLR is designed to be used with well-annotated genomes, such as the human genome, where nearly all exons are known and most novel transcripts arise from alternative splicing. We have clearly stated this limitation in the manuscript.  Despite this constraint, we believe NIFFLR provides tremendous value through its ability to quantify known isoforms and to detect and quantify novel alternatively spliced isoforms with high sensitivity and accuracy, rivalling the best published tools.

3. Related to point 2, it is hard to assess the performance of the tool relative to other reported findings in the field because it only contains novel-in-catalog novel transcripts according to the very commonly-used SQANTI classification (Pardo-Palacios F, et al., 2024 [Ref 1]). It is common in the field to examine the proportions of transcripts from each of these categories as a quality control metric.

Response: We designed our comparison of tools using simulated data so that not all reference transcripts present in the data were provided to the programs, including NIFFLR.  This way we can evaluate each program’s ability to recover novel transcripts that are present in the sample, but absent from the supplied reference annotation. To measure precision, we report the number of false positive transcripts as well as the precision values in Table 1. 

4. My main concern in this paper however is the quality of the benchmarking. There have been many papers that have performed long-read RNA-seq benchmarking in the past and have employed various metrics to evaluate the results: (Dong X, et al., 2023 [Ref 2], Chen Y, et al., 2022 [Ref 3], Pardo-Palacios F,  et al., 2024 [Ref 4]), 

The manuscript would benefit from adopting such standard benchmarking strategies that have become widely-accepted. As it stands, it is very tough to contextualize its results in the field as a whole. Some more point-by-point recommendations would be at least the following:

4a. For the simulations, as referenced in point 2, it is unrealistic to expect that novel transcripts will only arise from novel combinations of known exons (NICs). The authors should consider using existing simulated novel transcript ground truth datasets that exist, such as the one from LRGASP.

Response: We appreciate the reviewer’s suggestion. For this study, we chose to use transcripts produced by Nanosim, a publicly available and peer-reviewed tool specifically designed to generate realistic long-read transcriptome data. Our goal was to ensure a controlled and reproducible comparison under well-defined simulation parameters. We agree that other simulated datasets, such as those from LRGASP, provide valuable complementary benchmarks and will consider incorporating them in future evaluations.

4b. For the quantification benchmarking, it is more common to report a correlation metric between the ground truth and the estimated quantification from the tool. This would make these results more interpretable and comparable to other benchmarking efforts in the field. Please see our answer to point 3 below.

4c. Also related to the quantification benchmarking, there are no significance values reported for any of the pairwise comparisons; just written speculation on the visual appearance of the plots. Statistical analyses would increase the confidence of these results.

Response: The reviewer makes a good point. To address it, we have added values of Pearson correlation coefficient (PCC) values to the box-and-whisker plots in the manuscript. These values provide a quantitative measure of concordance between methods and allow readers to assess the significance of the pairwise comparisons more rigorously.

4d. Finding more isoforms is not necessarily a metric of a “better” tool, but is referenced as if it is in the text related to Figure 5. In fact, the percentage of reported “novel” transcripts for various GTEx tissues is surprisingly high and therefore the high number of transcripts could be indicative of over-calling novel transcripts and therefore poor specificity. Instead, the authors should additionally overlap the discovered isoforms with those discovered by Glinos et al. in their original paper (or other external datasets) to see how well their method recapitulates what others have already said about the dataset.

Response: We completely agree that the sheer number of novel isoforms is not, by itself, a measure of a “better” tool.  However, in light of our other benchmarks on both simulated and real data, we believe that the novel isoforms identified by NIFFLR are likely to be highly reliable, and the high percentage of novel transcripts observed in GTEx tissues could indeed reflect incomplete annotation rather than overcalling. Importantly, many independent studies (such as Glinos et al.) have reported that existing annotations capture only a subset of true transcripts, and numerous bona fide isoforms remain unannotated. In this context, NIFFLR’s detection of additional isoforms is consistent with these observations.  We have updated the Results section to make this more clear:

“…The number of isoforms identified by NIFFLR far exceeds the number reported by FLAIR ( Glinos et al., 2022), which identified 93,718 transcripts across 21,067 genes, of which 77% were novel. 34,876 transcripts in 11,840 gene loci were in common between the set of transcripts identified by ( Glinos et al., 2022) and by this study…”  

5. Would the authors be able to speculate on any specific use case (a specific cohort, a specific technology, etc) where the exons-to-reads alignment approach might be especially beneficial compared to traditional minimap2-based approaches? This might add some insight to the discussion.

Response: NIFFLR is designed to be used with well-annotated genomes, such as the human genome, where almost all exons are known and most of (or all) novel transcripts are due to alternative splicing. We stated the limitation in the manuscript.  Using exons from the reference annotation allows us to avoid performing spliced alignment to high-error transcriptomic reads, thereby achieving higher sensitivity and more accurate quantification of reference transcripts.

Minor concerns:

1. As the development of the method is of central importance to the paper, the implementation could be expanded on or explained a bit more. In particular, the concepts of the “implied” starts / ends of the alignments were confusing in Figure 1 and related text. Similarly, Figure 2 was a bit overly-complicated, and perhaps the authors could consider presenting the possible transcript paths separately as alternative transcripts in genome browser format (ie IGV or UCSC).

Response:  We have completely revised the Methods section to improve clarity and added a workflow diagram of the algorithm to the Supplementary materials as Figure S1.  Here is the updated algorithm description from the Methods:

After building the alignments, we assign each long read to a gene locus using a majority vote approach. Specifically, for each read, we compute the total number of K-mers in all LCSs for matching exons across different gene loci and assign the read to the locus L whose exons collectively have the highest total number of matching K-mers. Alignments of exons that belong to different gene loci are then discarded. Next, we build the transcript matching the read by finding the best tiling of the read using exons that belong to locus L—that is, the sequence of exons that maximizes read coverage while minimizing gaps or overlaps in the implied alignment coordinates. The long read defines a 5’ to 3’ forward direction, which provides a natural topological order. We therefore sort the aligned exons by their alignment start coordinates if they align in the forward direction, or by their alignment end coordinates if they align in the reverse direction. Since we only kept alignments of exons that all belong to a single gene locus L, the exons must all align either in the forward or reverse direction. For simplicity, we describe the algorithm below assuming all exons are aligned in the forward direction; the reverse case is handled the same by reversing the read.

We represent the exon tiling problem as a directed graph, where nodes correspond to exons, and node weights are exon lengths.  An edge connects the 3’ end of an exon A to the 5’ end of exon B if the absolute value of the distance between their aligned positions on the read is less than 20bp. The weight of the edge between exons A and B is defined as this distance.  Next, we choose the “start” nodes.  A valid start node is an exon that is not connected on the 5’ end and whose 5’ end lies upstream of the read start (i.e., has a negative coordinate relative to the read origin), indicating an overhang. If no such exon exists, we select the exon(s) with the smallest positive coordinate among all exons aligned to the read. If multiple exons share the same alignment start due to alternative splicing, we use all such exons as alternative start nodes. We select “end” nodes in a similar manner: an end node is an exon not connected on the 3’ end and whose 3’ ends extends beyond the end of the read (i.e., has an overhang). If none do, we choose the exon(s) whose 3’ end coordinate is closest to the 3’ end of the read.

We solve the exon tiling problem by finding the path through the graph that starts from any start node and ends at an end node and minimizes a penalty function. The penalty is defined as the sum of edge weights along the path plus an overhang penalty, calculated as defined as 0.1× (|5’ overhang of the start exon| + |3’ overhang of the end exon|), where |a| denotes the absolute value of a. Ideally, there should be no gaps or overlaps between aligned exons in the transcript, resulting in a perfect path with zero weight.  However, because psa_aligner computes approximate alignments, exon start and end alignment coordinates are imprecise estimates. In case of a tie, we select the path with the larger total node weight (the sum of the lengths of the exons on the path). If a read can be spanned by a single exon — either because only one exon maps to the read, or because that exon simultaneously has the start closest to the 5′ end and the end closest to the 3′ end of the read— we report that single exon as the path. If multiple exons individually span the read, we select the exon that satisfies the condition of being a valid start exon or end exon and has the smallest total overhang length. Finally, if no valid path is found, we report the exon with the longest alignment to the read. Figure 2 illustrates an example of a valid exon path. Once the best path is identified, we examine the genomic coordinates of the exons, which are encoded in their sequence IDs. We discard the path if any exons in it overlap in genomic coordinates, as this likely indicates a long read is chimeric a substantial local genome rearrangement that NIFFLR cannot handle.

2. At the end of the first paragraph of main text on page 7, the authors state “This result demonstrates that when novel isoform discovery and quantification are the primary goals, NIFFLR is the best tool” when referring to the simulation experiment where they measured isoform detection / assembly only. This means this result has no relevance for quantification. Response: We thank the reviewer for this observation. We agree that the original placement of the statement implied a conclusion about quantification based solely on isoform detection/assembly, which could be misleading. To address this, we have moved the statement to the end of the last paragraph in the “Comparison on simulated long reads” section, immediately following the discussion of the quantification benchmarking results. This placement more accurately reflects the context of the data and avoids conflating isoform discovery with quantification performance.

3. If the authors mean “false positive transcripts” by “spurious transcripts”, they should simply refer to them as the latter as there is no definition for “spurious” transcripts in the text.

Response:  We have changed the wording from “spurious transcripts” to “false positive transcripts”.

4. In table 2, IsoQuant has the same number of known and total isoforms; implying it found no novel isoforms. I am highly doubtful this is correct and is probably a typo.

Response: Thank you for pointing this out, indeed this was a typo and we corrected it.

5. Figure 3 is missing y axis labels. Furthermore, the meaning of the box and whiskers are elaborated on in the results, which should just be in the legend. Additionally, Figure 3b is just a zoomed in duplicate of two plots from 3a and is unnecessary.

Response:  We have added the y-axis label to Figure 3. We also wish to clarify that Figure 3b is not a zoomed-in duplicate of plots from 3a. While Figure 3a shows quantification error across all transcripts detected by each tool, Figure 3b focuses specifically on the shared subset of transcripts quantified by both NIFFLR and IsoQuant. This allows for a fair, direct comparison between the two programs on the transcripts shared by both, highlighting differences in quantification performance that could be obscured when considering all transcripts.

6. For Figure 4 the authors describe a strange method of depth normalization to compare the long-read RNA-seq to short-read RNA-seq transcript quantification estimates. They should use normal TPM / CPM normalization. This figure also has unnecessary details about the plot in the results section which should be in the legend (same as in Figure 3).

Response:  TPM and transcript coverage are linearly related, and many long-read transcriptome quantification programs report read counts rather than TPMs.  Transcript coverage computed from the Illumina data is directly comparable to the number of long reads covering each transcript. By normalizing the Illumina coverage by the total sequencing depth, we made Illumina read coverage into a metric equivalent to long read counts, enabling a fair comparison between long- and short-read quantifications.   

7. The authors make no references to their supplementary material PDF in the main body of the text. They should include references so that readers know where to find the calls made to perform the benchmarking etc.

Response: Thank you for noting that, we added the missing references to Supplementary Materials.

Major comments:

1. A key limitation of the method is its reliance on a known set of exons: all predicted transcripts must be combinations of known exons.  It appears that even slight variations in the boundaries of exons must be previously annotated for the method to predict transcripts with those variant exons. This general issue is brought up in the last two sentences of the discussion.  The authors suggest that when the full exon set is not known, short read data could be used to delineate the exons.  That may be a feasible strategy, but it is not implemented or evaluated in this work, nor do the evaluations consider novel exons.  This work would be strengthened by evaluations that include the more realistic scenario of novel exons (including 5' and 3' end variants of known exons).  By definition, NIFFLR will not be able to detect or quantify transcripts including these exons, but the impact of these exons on the precision and quantification accuracy of other transcripts should be measured.  Alternatively, the authors could implement the exon delineation strategy that they mentioned and examine NIFFLR's performance in combination with this strategy.  Admittedly, the evaluation on the real data set potentially includes reads from transcripts with novel exons, but it is difficult to discern from the results presented how any novel exons impacted the method's performance.

Response: We agree with the reviewer’s observation that using a known set of exons is a key limitation of NIFFLR.  NIFFLR is designed to be used with well-annotated genomes, such as the human genome, where almost all exons are known and most novel transcripts are due to alternative splicing. We have added an analysis of the synthetic SIRV E2 data, which includes alternative exons whose boundaries differ by only a few bases, and we show that NIFFLR and ESPRESSO were able to find all but one of the isoforms that are present in the sample.   

2. A number of details of the method are omitted or unclear.  A more formal and detailed presentation of the algorithm is needed.  For example, (A) what algorithm is used to solve the "exon tiling problem"? (B) what, precisely, is the objective function? (C) is the algorithm guaranteed to find the optimal solution? (D) how is an edge allowed between overlapping exons? (E) what is the intuition behind distributing reads "proportionally to the size of the container transcripts"?  Personally, I would like to see a more mathematical presentation and a figure depicting the steps of the entire procedure, particularly those detailed in the paragraph beginning "By design, all intron junctions...".

Response: We added a flowchart of the method as Supplementary Figure S1 and substantially revised the Methods section to improve clarity. The revised text is below:

We represent the exon tiling problem as a directed graph, where nodes correspond to exons, and node weights are exon lengths.  An edge connects the 3’ end of an exon A to the 5’ end of exon B if the absolute value of the distance between their aligned positions on the read is less than 20bp. The weight of the edge between exons A and B is defined as this distance.  Next, we choose the “start” nodes.  A valid start node is an exon that is not connected on the 5’ end and whose 5’ end lies upstream of the read start (i.e., has a negative coordinate relative to the read origin), indicating an overhang. If no such exon exists, we select the exon(s) with the smallest positive coordinate among all exons aligned to the read. If multiple exons share the same alignment start due to alternative splicing, we use all such exons as alternative start nodes. We select “end” nodes in a similar manner: an end node is an exon not connected on the 3’ end and whose 3’ ends extends beyond the end of the read (i.e., has an overhang). If none do, we choose the exon(s) whose 3’ end coordinate is closest to the 3’ end of the read.

We solve the exon tiling problem by finding the path through the graph that starts from any start node and ends at an end node and minimizes a penalty function. The penalty is defined as the sum of edge weights along the path plus an overhang penalty, calculated as defined as 0.1× (|5’ overhang of the start exon| + |3’ overhang of the end exon|), where |a| denotes the absolute value of a. Ideally, there should be no gaps or overlaps between aligned exons in the transcript, resulting in a perfect path with zero weight.  However, because psa_aligner computes approximate alignments, exon start and end alignment coordinates are imprecise estimates. In case of a tie, we select the path with the larger total node weight (the sum of the lengths of the exons on the path). If a read can be spanned by a single exon — either because only one exon maps to the read, or because that exon simultaneously has the start closest to the 5′ end and the end closest to the 3′ end of the read— we report that single exon as the path. If multiple exons individually span the read, we select the exon that satisfies the condition of being a valid start exon or end exon and has the smallest total overhang length. Finally, if no valid path is found, we report the exon with the longest alignment to the read. Figure 2 illustrates an example of a valid exon path. Once the best path is identified, we examine the genomic coordinates of the exons, which are encoded in their sequence IDs. We discard the path if any exons in it overlap in genomic coordinates, as this likely indicates a long read is chimeric a substantial local genome rearrangement that NIFFLR cannot handle.

3. With respect to the software package, I was ultimately able to compile the software after installing the Boost and zlib development libraries.  It would be helpful to have these compilation dependencies noted in the README.  For ease of use by the community, it would also be helpful to have the package available via conda and/or Docker.  Finally, please provide a small test dataset with the software such that users can make sure that their installation is working and can see what the inputs and outputs look like.

Response: We added information about software dependencies to the README file.  We also included a script that runs NIFFLR on the SIRV E2 dataset, that downloads the reads from SRA using fastq-dump. In addition, we have made NIFFLR available on Bioconda, where it can now be installed as ‘nifflr’.

Minor comments:

4. From the filtering steps of the algorithm, could there be an output provided that might indicate the presence of novel exons at a given locus?  For example, could the user be provided the number of reads that mapped to a locus but that were not ultimately assigned to a transcript?

Response: At this time, we do not provide this directly, but this information is available in an intermediate GTF file that lists all reads contributing to each transcript.  The file is named .gtf. If a transcript listed in that file is not present among the final output transcripts, it indicates that the reads supporting that transcript were ultimately discarded.

5. For quantification evaluations, the box and whisker plots of log2 fold changes are helpful, but I would have also liked to have seen scatterplots of true vs. predicted counts (on log-scaled axes) along with correlation values, as has been common for short-read quantification methods.  Such scatterplots help to visualize potential subsets of transcripts that have biased predictions and trends relative to the magnitude of expression.

Response:  We added the scatter plots to the Supplementary Materials as Figures S2-S6.  We also included the Pearson correlation coefficient values in the plots shown in the main manuscript.