BIOTRAM: Biochemistry-informed Transformer for Relative Abundance via Metabarcoding

arXiv:2611.05418 · q-bio.GN, cs.LG · Submitted 2026-11-01 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Genomics Radio. Generated commentary on the latest computational biology and genomics papers.

Ines: We're looking at the paper: "BIOTRAM: Biochemistry-informed Transformer for Relative Abundance via Metabarcoding".

Marcus: Here's the gist of the paper. Arthropod monitoring at scale is often limited by metabarcoding, which identifies taxa but fails to account for relative abundance due to variations in DNA amplification rates and specimen body sizes.

Ines: BIOTRAM is proposed as a neural network designed to estimate per-taxon specimen counts by learning the read yield specific to each specimen.

Marcus: This estimation is achieved by dividing observed read counts for a taxon by a learned yield per specimen, which incorporates both the DNA contributed by that individual and the efficiency of its amplification.

Ines: The model's structure is intrinsically linked to the experimental pipeline, using Barcode Index Numbers (BIN) as its taxonomic unit.

Paper summary: Ines: So we're going over "BIOTRAM: Biochemistry-informed Transformer for Relative Abundance via Metabarcoding" again. The central idea here is that traditional metabarcoding just gives us a species list, and BIOTRAM proposes using a neural network to estimate the actual number of specimens per taxon in those samples.

Marcus: It claims this estimation works by learning the read yield specific to each specimen, and it does this by dividing the observed read counts for a taxon by a learned yield that incorporates both how much DNA that individual contributes and how efficiently its amplification goes.

Yuki: So why is this important? Because insect diversity is huge, but quantifying it reliably has been difficult for a long time. This paper suggests that linking the species list to physical properties like body size and primer binding efficiency can give us a better picture of the abundance we are missing.

Ines: The model produces latent predictions for these physical properties, and those predictions directly shape how the specimen counts relate to the observed read counts per BIN across different bottles. It’s not just guessing; it’s using biological and chemical knowledge baked into the model's structure.

Marcus: They frame it as a staged model, moving beyond a simple approach by basing it on the biochemical properties of the metabarcoding process itself, which is what makes it distinct from naïve models.

Yuki: From my perspective, this means we can use this to get relative abundance estimates across huge scales where manually counting individuals is just impossible. It’s about scaling our understanding of insect populations beyond just presence and absence.

Ines: The paper sets out a controlled evaluation under distribution shift, meaning they tested it against data from different collection locations, different time points, and even sequencing methods from other labs to see if it held up outside the training set.

Marcus: They also used a paired dataset of one thousand one hundred eighty-eight bottles processed with both metabarcoding and individual barcoding to create a ground-truthing dataset that's much larger than what existed before.

Conclusion: Ines: So, looking at the BIOTRAM paper again, it’s really about using biochemistry to inform a transformer to estimate relative abundance from metabarcoding data. The authors are demonstrating that this structure allows for the prediction of physical traits like body size and primer efficiency as intermediate steps in linking specimen counts to read counts.

Marcus: The core takeaway for me is that they managed to get significantly lower composition divergence, dropping it from zero point three five eight six down to zero point one two zero zero on raw read proportions, which shows the model is much better at capturing the true species mix.

Yuki: And for the wider field, this suggests that when we combine high-throughput sequencing with biological knowledge about how DNA is amplified, we can start getting estimates of insect abundance that are more reliable than just relying on species lists alone.

Ines: It’s about building a tool that doesn't just report what's there, but attempts to quantify the actual numbers behind those reports, which is a fundamental shift in how we analyze biodiversity data.

Emma Boehly, Scott C. Lowe, Ken A. Thompson, Anthony Fuller, Will Armstrong, Iuliia Zarubiieva, Neelu Madan, Tiancheng Gao, Sameer Padhye, Jayme Sones, © Kate H. J. Perez, © Bess Hardwick, IJ Brendan Furneaux, ⅸ Evgeny V. Zakharov, Dirk Steinke, Jeremy R. deWaard, Stephanie L. deWaard, Jaclyn McKeown, Norman Monkhouse, Cole Pace, Ū Sujeevan Ratnasingham, œ Raina Schmuck, Ł Amy Thompson, … Jorge L. León-Corts, Luciano Cagnolo, Julieta Soledad Arena, Jeremías Domnguez Mascruale, Errol Douwes, № Elli Groner, Ŋ Masae Iwamoto Ishihara, Shunsuke Matsuoka, Mailyn A. Gonzalez, … Alexandre Antonelli, œ Renata S. Sousa-Lima, Armin Niessner, Patrcia Guedes, œ Elizabeth G. Biro, Ž Zuania Coln-Piro, Dezene P. W. Huber, Santosh Mazumdar, Magda Bou Dagher Kharrat, Eddie P. Mondejar, Pablo Luis Tubaro, É Eugene Marais, ō Marie-Eugénie Maggia, Darso Alejandro Lijtmaer, Sean W. J. Prosser, Andrea Murillo-Vazquez, Paola Pulido-Santacruz, ℳ Paula Leitman, ř Thiago Arnt Dorigo, Nao Ito, Paul Knoops, Ŋ Maria L. Kuzmina, Hikaru Morishita, Keishi Kimoto, Dimby Raharinjanahary, ⅸ Rhian J. Smith, œ Solny A. Adalsteinsson, Kevin G. Borja-Acosta, Paul Amouroux, Angela M. Mendoza-Henao, É Angélica Diaz-Pulido, Anamaria Dal Molin, Luke L. Powell, ⅸ Diego Inclan, Alice Calvente, David Dunson, ń Brian Fisher, Tomas Roslin, Otso Ovaskainen, œ Paul D. N. Hebert

University of Guelph · Vector Institute · Centre for Biodiversity Genomics · Aalborg University · University of Helsinki · ⅸ University of Jyvaskyla Centre for Biodiversity Genomics Institute for the Frontera Sur Universidad Nacional de Córdoba Universidad del Rosario Independent Researcher Universidade Federal do Rio Grande do Norte Hochschule für Forstwirtschaft Rottenburg Faculdade de Cincias da Universidade de Lisboa Washington University in St. Louis University of Florida University of Northern British Columbia Bangladesh Forest Research Institute Alliance for Environment and Community Development Université Saint-Joseph European Forest Institute Mindanao State University Iligan Institute of Technology Premier Research Institute of Science and Mathematics Museo Argentino de Ciencias Naturales Universidad de Buenos Aires Gobabeb Namib Research Institute Andes Amazon Fund Universidad del Rosario Independent Researcher Universidade Pedaggica y Tecnológica Universidad Mayor

q-bio.GN, cs.LG

Submitted: 2026-11-01

Updated: 2026-11-01

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 90/100

The gist: Arthropod monitoring at scale is often limited by metabarcoding, which identifies taxa but fails to account for relative abundance due to variations in DNA amplification rates and specimen body sizes.

Key concepts

BIOTRAM
A neural network architecture that estimates the number of specimens for each taxon by learning how much DNA each individual contributes and how efficiently it is amplified. It uses Barcode Index Numbers (BIN) as its unit to link observed reads to estimated counts.
Read Yield Estimation
The core function of BIOTRAM is to estimate the specific read yield for every specimen. This involves dividing observed read counts for a taxon by a learned yield per specimen, which accounts for both the DNA content of that individual and the efficiency with which its DNA was amplified during sequencing.
Composite Loss Function
BIOTRAM minimizes a complex loss function that balances several objectives: accurately predicting count likelihood, maintaining composition similarity, ensuring presence is detected, and linking predicted reads per specimen to actual observed values. This comprehensive approach guides the model's learning process.

Terminology

Summary

Arthropod monitoring at scale is often limited by metabarcoding, which identifies taxa but fails to account for relative abundance due to variations in DNA amplification rates and specimen body sizes. p. 0

Where the paper says it

ca Abstract Arthropod monitoring at scale begins with metabarcoding of bulk samples: trap collections of hundreds to thousands of specimens processed as one unit, extraction of DNA from the bulked specimens, then sequencing of the DNA from all specimens together. Metabarcoding reveals which taxa such a sample contains, but due to factors such as differing DNA amplification rates and body sizes, its read counts do not neatly correspond to how abundant the different taxa are relative to one another.

BIOTRAM is proposed as a neural network designed to estimate per-taxon specimen counts by learning the read yield specific to each specimen. p. 0
Where the paper says it

This lack of relative abundance limits the data’s utility for the multitude of ecological analyses that depend on it. We propose BIOTRAM: a neural network that estimates per-taxon specimen counts by dividing the observed read counts by a learned read yield per specimen, which it factors into the DNA each specimen contributes and how efficiently that DNA is amplified.

This estimation is achieved by dividing observed read counts for a taxon by a learned yield per specimen, which incorporates both the DNA contributed by that individual and the efficiency of its amplification. p. 0
Where the paper says it

We propose BIOTRAM: a neural network that estimates per-taxon specimen counts by dividing the observed read counts by a learned read yield per specimen, which it factors into the DNA each specimen contributes and how efficiently that DNA is amplified. On a paired dataset of 1,188 bottles (800 k specimens) processed both by metabarcoding and by comprehensive individual barcoding for ground-truthing, BIOTRAM reduces the Jensen–Shannon divergence between the estimated and the true composition from 0.

The model's structure is intrinsically linked to the experimental pipeline, using Barcode Index Numbers (BIN) as its taxonomic unit. BIOTRAM predicts physical quantities such as body size and primer binding efficiency, which are then used to establish a relationship between specimen counts and observed read counts. p. 2

Where the paper says it

(3) Latent predictions of primer binding efficiency and body size (Section 4. Our model produces physically interpretable latent estimates of both properties, which directly shape the relationship between specimen counts and read counts per BIN and bottle.

The training objective minimizes a composite loss function: L = LNB + lambda comp L comp + lambda pres L pres + lambda rps L rps, comprising the negative-binomial count likelihood (LNB), composition divergence (L comp), presence loss (L pres), and a term linking predicted reads per specimen to observed values (L rps). p. 5
Where the paper says it

(2)’s prediction, and 𝜎(𝑧 (𝑏) 𝑖 ) for the predicted presence probability. Training minimizes a negative-binomial count likelihood (NB), which accommodates the over-dispersion of ecological counts, whose variance exceeds their mean, plus a per-bottle divergence on composition, a presence loss, and a term tying the predicted reads per specimen to its observed value: L = LNB + 𝜆comp Lcomp + 𝜆pres Lpres + 𝜆rps Lrps, (3) LNB = − 1 |R| ∑︁ (𝑖,𝑏) ∈R log NB𝑛 (𝑏) 𝑖 ; 𝑛ˆ (𝑏) 𝑖 , 𝛼 (𝑏) 𝑖 , Var 𝑛 (𝑏) 𝑖 = 𝑛ˆ (𝑏) 𝑖 + 𝛼 (𝑏) 𝑖 𝑛ˆ (𝑏) 𝑖 2 , (4) Lcomp = 1 |B| ∑︁ 𝑏∈ B JSD 𝑝 (𝑏) 𝑝ˆ (𝑏) , 𝑝 (𝑏) 𝑖 = 𝑛 (𝑏) 𝑖 Í 𝑗 ∈𝑏 𝑛 (𝑏) 𝑗 , 𝑝ˆ (𝑏) 𝑖 = 𝑛ˆ (𝑏) 𝑖 Í 𝑗 ∈𝑏 𝑛ˆ (𝑏) 𝑗 , (5) Lpres = − 1 |R| ∑︁ (𝑖,𝑏) ∈R h 𝑦 (𝑏) 𝑖 log 𝜎(𝑧 (𝑏) 𝑖 ) + 1 − 𝑦 (𝑏) 𝑖 log 1 − 𝜎(𝑧 (𝑏) 𝑖 ) i , (6) Lrps = 1 Í (𝑖,𝑏) 𝑤𝑖 ∑︁ (𝑖,𝑏) ∈R 𝑤𝑖 𝐻𝛿 𝜀 (𝑏) 𝑖 − 𝜀¯ 𝑠𝑖 !

A key distinguishing feature of BIOTRAM is its use of intermediate quantities that are hypothesized to correspond to physical measurements, offering interpretable predictions. p. 8

Where the paper says it

Do the predicted quantities mean what they are named? The model’s distinguishing property is its intermediate quantities, which should correspond to physical quantities and explain its predictions.

Although physical quantities like primer efficiency and body mass are provided as external input features, direct correlation is avoided due to potential circularity. p. 8
Where the paper says it

The model’s distinguishing property is its intermediate quantities, which should correspond to physical quantities and explain its predictions. Since the primer efficiency and body mass receive the corresponding external estimates as input features, correlating them with those estimates would be circular.

Instead, the model's predictive power is validated by checking if the empirical per-BIN (reads/specimen), derived from the metabarcoding pipeline, correlates with the predicted a bar i + m i (Equation 1), averaged across test bottles.

Performance evaluations on a paired dataset of 1,188 bottles (800k specimens) processed by both metabarcoding and comprehensive individual barcoding demonstrated significant improvements. p. 0

Where the paper says it

We propose BIOTRAM: a neural network that estimates per-taxon specimen counts by dividing the observed read counts by a learned read yield per specimen, which it factors into the DNA each specimen contributes and how efficiently that DNA is amplified. On a paired dataset of 1,188 bottles (800 k specimens) processed both by metabarcoding and by comprehensive individual barcoding for ground-truthing, BIOTRAM reduces the Jensen–Shannon divergence between the estimated and the true composition from 0.

BIOTRAM reduced the Jensen–Shannon divergence between estimated and true composition from 0. p. 0
Where the paper says it

We propose BIOTRAM: a neural network that estimates per-taxon specimen counts by dividing the observed read counts by a learned read yield per specimen, which it factors into the DNA each specimen contributes and how efficiently that DNA is amplified. On a paired dataset of 1,188 bottles (800 k specimens) processed both by metabarcoding and by comprehensive individual barcoding for ground-truthing, BIOTRAM reduces the Jensen–Shannon divergence between the estimated and the true composition from 0.

3586 to 0.1200 for raw read proportions. p. 0
Where the paper says it

On a paired dataset of 1,188 bottles (800 k specimens) processed both by metabarcoding and by comprehensive individual barcoding for ground-truthing, BIOTRAM reduces the Jensen–Shannon divergence between the estimated and the true composition from 0. 3586 for raw read proportions (the usual proxy for abundance) to 0.

Trained against four generic deep architectures with tuned loss weights, BIOTRAM achieved a count error 3. p. 0
Where the paper says it

3586 for raw read proportions (the usual proxy for abundance) to 0. Trained on the same objective and inputs as four generic deep architectures, all with tuned loss weights, it reaches a count error 3.

2% lower than the strongest baseline and exhibited the lowest composition divergence. p. 0
Where the paper says it

Trained on the same objective and inputs as four generic deep architectures, all with tuned loss weights, it reaches a count error 3. 2% below that of the strongest of them and the lowest composition divergence.

Furthermore, it demonstrated superior generalization capabilities, outperforming baselines on data from unseen regions, later collection dates, and different sequencing batches, as well as transferring effectively to data from an unseen laboratory. p. 0
Where the paper says it

2% below that of the strongest of them and the lowest composition divergence. It also generalizes to out-of-domain data better, beating all baselines on bottles from unseen regions, later collection dates and unseen sequencing batches, and transferring to data from an unseen laboratory.

The model's capacity is not solely defined by parameter count; BIOTRAM's architecture scales up to 3.2M parameters at its largest rung (10. p. 7

Where the paper says it

2% with tuned weights (Appendix I). Across a capacity ladder from 200k to 128M parameters, retuned at every rung, no model moves by more than 0.

5M), showing a margin over the best baseline of the same rung between 0. p. 6
Where the paper says it

1200, against 0. 1234 for the best baseline.

0018 and 0.0052 JSD, even at its worst size (JSD of 0. p. 7
Where the paper says it

Across a capacity ladder from 200k to 128M parameters, retuned at every rung, no model moves by more than 0. 0062 JSD, and BIOTRAM at its worst size (0.

121). The model's data efficiency improves substantially once sufficient data is available, as performance curves show BIOTRAM surpassing deep baselines after approximately 648 bottles, reaching a JSD of 0.1198 against TabICLv2’s 0. p. 6
Where the paper says it

036 BIOTRAM 0. 024 TabICLv2 (in-context, 28.

1227 (p = 0.001) when trained on all 864 bottles. p. 6
Where the paper says it

Data is partitioned at the bottle level. The 585 excess-free bottles are split 409 / 87 / 89 into train/val/test, and 455 of the 603 bottles with excess specimens (Appendix E) join the training set only, for 864 training bottles in all.

Read elasticity analysis reveals that the amplification term is crucial; when a BIN's reads double, its predicted amplification rises with an elasticity of 0.29, indicating that part of the increased reads is attributed to stronger amplification rather than just higher specimen counts. p. 1

Where the paper says it

Individual barcoding (manual sorting and per-specimen sequencing) gives accurate species counts but is slow and costly; metabarcoding (bulk amplification of the sample’s DNA) is fast and cheaper, but its read counts are not linearly related to specimen abundances. Rather than abundance, the output of metabarcoding is a vector of sequence read counts, one count per taxon.

Ablation studies confirm the importance of several components: removing Cross-BIN attention increases JSD by +0. p. 3
Where the paper says it

(B) Insect Size Estimator predicts a BIN’s average specimen size from a representative barcode sequence, trained on estimates found from paired image data. (C) BIOTRAM’s architecture integrates barcode, taxonomic, and sample metadata to jointly predict each BIN’s body-mass term and presence; an attention module corrects raw read counts for amplification bias across replicates before the corrected estimates are merged into a final abundance prediction, trained with negative binomial, Jensen–Shannon divergence (JSD), binary cross-entropy (BCE) and Huber losses (Section 4.

0127, and removing the Taxonomy embedding results in a significant JSD increase (+34). p. 5
Where the paper says it

Each computes a frozen BarcodeMamba embedding of the barcode (Gao & Taylor, 2024; Gao et al. , 2025) through its own attention layer, adds a learned taxonomy embedding and the bottle’s location and collection date (through a pretrained geographic encoder (Klemmer et al.

The input features Bottle’s total specimen count nu(b) and BIN’s own read count in the amplification module (read-blind variant) are also identified as significant contributors to model performance.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Per-taxon specimen count estimation: The improved BIOTRAM system can estimate per-taxon specimen counts by dividing observed read counts by a learned read yield per specimen, which factors into the DNA each specimen contributes and how efficiently that DNA is amplified. This allows for estimates of taxa absent from training data.

  2. Latent quantity interpretation: The model predicts latent quantities such as body mass and primer binding efficiency, which are physically interpretable. This allows the system to produce physically interpretable latent estimates of both properties, which directly shape the relationship between specimen counts and read counts per BIN and bottle.

  3. Robustness to data shift: The improved BIOTRAM can generalize to out-of-domain data better by being trained on novel collection locations, later collection dates, and unseen sequencing batches. This generalization is demonstrated by beating all baselines on bottles from unseen regions, later collection dates and unseen sequencing batches.

  4. Ground truth correction for excess specimens: The system incorporates a mechanism to correct for excess specimens in the training data by using a model trained only on no-excess bottles to predict relative abundance in bottles with excess, assigning counts based on model prediction factor (MPF).

Abstract

Arthropod monitoring at scale begins with metabarcoding of bulk samples: trap collections of hundreds to thousands of specimens processed as one unit, extraction of DNA from the bulked specimens, then sequencing of the DNA from all specimens together. Metabarcoding reveals which taxa such a sample contains, but due to factors such as differing DNA amplification rates and body sizes, its read counts do not neatly correspond to how abundant the different taxa are relative to one another. This lack of relative abundance limits the data's utility for the multitude of ecological analyses that depend on it. We propose BIOTRAM: a neural network that estimates per-taxon specimen counts by dividing the observed read counts by a learned read yield per specimen, which it factors into the DNA each specimen contributes and how efficiently that DNA is amplified. On a paired dataset of 1,188 bottles (800 k specimens) processed both by metabarcoding and by comprehensive individual barcoding for ground-truthing, BIOTRAM reduces the Jensen-Shannon divergence between the estimated and the true composition from 0.3586 for raw read proportions (the usual proxy for abundance) to 0.1200. Trained on the same objective and inputs as four generic deep architectures, all with tuned loss weights, it reaches a count error 3.2% below that of the strongest of them and the lowest composition divergence. It also generalizes to out-of-domain data better, beating all baselines on bottles from unseen regions, later collection dates and unseen sequencing batches, and transferring to data from an unseen laboratory. Structuring the mapping from reads to counts around the steps of the experimental pipeline thus yields more accurate abundance estimates than leaving that mapping unconstrained.

Sources

Related papers