PaperlayerView exact revision on arXiv

Source-linked AI summary

When Do Biological Reasoning Models Use Their Biological Inputs?

Ada Fang, Nikitha Thoduguli, Lukas Fesser, Hanlin Zhang, Sham M. Kakade, Marinka Zitnik

arXiv:2610.00898v1cs.LG

TL;DR

Benchmark accuracy alone does not establish whether biological inputs drive biological reasoning models' predictions, particularly when inputs overlap or predictive text is available. The paper tests input dependence across six models with perturbations, evidence conflicts, probes, and trace analyses. It finds that input contributions vary across models and that higher post-training accuracy does not ensure greater contribution from foundation-model representations.

  • Problem

    Benchmark accuracy alone cannot establish which biological inputs drive predictions, especially when representations overlap and predictive text may support accurate answers.

    Source: §1 · p. 2

  • Method

    Across six models, the paper perturbs biological inputs, constructs evidence conflicts, probes received representations, and checks reasoning traces against supplied evidence.

    Source: §1 · p. 2

  • Results

    Foundation-model inputs contribute little to BioReason and BioReason-Pro performance but contribute to ChatNT, Prot2Text-V2, and CellWhisperer performance; higher BioReason and BioReason-Pro accuracy after post-training does not increase Evo2 and ESM3 contributions.

    Source: §1 · pp. 2–3

  • Takeaways & Limitations

    Access to a predictive representation or improved performance after SFT or RL does not ensure that the representation contributes to predictions.

    Source: §5 · p. 10

    The six-model evaluation uses specific tasks and biological inputs, so its findings may not generalize to other architectures or modalities.

    Source: §5 · p. 10

Abstract

from arXiv · show

Biological reasoning models use post-training to connect LLMs to biological foundation model representations and biological text. Their benchmark accuracy is taken as evidence that LLMs reason over these inputs. We test this assumption in six biological reasoning models across DNA, protein, and single-cell tasks. We perturb one biological input while holding the query and other inputs fixed, construct evidence conflicts that pair the foundation model representation of one genome, protein, or cell with the text of another, fit linear probes to the representations the language model receives, and analyze reasoning traces against the biological inputs. Evo2 and ESM3 contribute little to BioReason and BioReason-Pro performance on the evaluated tasks. Shuffling the DNA sequence barely changes BioReason disease prediction accuracy, and in evidence conflicts the two models follow the text in 97.9% and 99.7% of cases. Linear probes trained on the Evo2 and ESM3 representations predict the task targets, so these foundation models encode information relevant to the task, but provide limited overall performance improvement to BioReason and BioReason-Pro. In contrast, foundation model inputs contribute to ChatNT, Prot2Text-V2, and CellWhisperer performance, and differentially expressed genes in the gene sentence contribute to Cell2Sentence-Scale performance. Across SFT and RL checkpoints of BioReason-Pro and 42 BioReason checkpoints, increases in accuracy do not imply greater performance contributions from biological inputs. BioReason traces misstate nucleotide changes, while BioReason-Pro traces describe functions omitted from final predictions under evidence conflicts. We find that current post-training strategies do not ensure that foundation model representations contribute to task performance.

1 INTRODUCTION

Benchmark accuracy alone cannot establish which biological inputs drive predictions, especially when models receive overlapping representations and predictive text. The paper tests input dependence across biological reasoning models and finds that predictive foundation-model representations can contribute little to performance.

1 INTRODUCTION

  • Benchmark performance alone cannot identify which inputs drive predictions, because predictive annotations may support accurate answers while sequence or transcriptomic representations are underused.Inputs can encode overlapping biological properties, so a small contribution may reflect support from another input, limited predictive value, or limited model use.Source: §1 · p. 2
  • The paper tests biological-input dependence by perturbing one input, creating evidence conflicts, probing received representations, and checking reasoning traces against supplied evidence.The evaluation covers six models across post-training checkpoints and model scales.Source: §1 · p. 2
  • Evo2 and ESM3 representations contribute little to BioReason and BioReason-Pro performance, while foundation-model inputs contribute to ChatNT, Prot2Text-V2, and CellWhisperer performance and DEGs contribute to C2S-Scale performance.SFT and RL increase BioReason and BioReason-Pro accuracy without increasing the contribution from Evo2 and ESM3; the same pattern appears across 42 BioReason checkpoints.Source: §1 · pp. 2–3
  • BioReason traces misstate nucleotide changes, while BioReason-Pro traces describe functions omitted from final predictions under evidence conflicts.For C2S-Scale, auxiliary edited-base prediction improves sensitivity to sequence changes, but sequence shuffling lowers disease-prediction accuracy by only 0.021.Source: §1 · p. 3

2 RELATED WORK

Prior work studies multimodal input use, representation predictability, and generated rationales, but evaluations of biological reasoning models have largely emphasized task performance. This paper focuses on which biological inputs contribute to downstream predictions.

Biological reasoning models

  • Evaluations of biological reasoning models have largely emphasized task performance, scale, and post-training rather than whether predictions depend on biological input content.The models combine different biological representations and text, including genome, protein, and transcriptomic inputs.Source: §2 · p. 3

Multimodal representation learning and evaluation

  • Multimodal studies examine shared and modality-specific representations, while vision-language evaluations test for language-prior reliance and insensitivity to visual content.Controlled perturbations and counterfactual inputs have been used to assess whether predictions respond to a modality's content.Source: §2 · p. 3

Predictive representations and generated rationales

  • Probe performance and downstream model behavior should be evaluated separately, and plausible generated rationales need not identify features that affect predictions.The paper uses predictive baselines to interpret limited input contributions and compares prediction-affecting features with features cited in rationales.Source: §2 · p. 3

3 APPROACH

The paper defines input use through controlled changes to biological content and evaluates predictions, performance, source preference, and evidence in rationales. Its protocol also compares performance contributions across model variants and tests auxiliary supervision.

3.1 PROBLEM FORMULATION

  • A task query q and biological inputs Z form the prompt; the model predicts a target and may produce a rationale, with Z comprising continuous representations and text.Tasks share an objective and label space, such as disease prediction.Source: §3.1 · pp. 3–4
  • An input is underutilized when a predictive baseline can recover the target from it but changing its task-relevant content has little effect on model performance.The approach perturbs one input while holding the query and remaining inputs fixed, and uses evidence conflicts to measure which source predictions follow.Source: §3.1 · p. 4
  • The evaluation covers six models, their input sources, variants, tasks, and evaluation data.Table 1 identifies the benchmarked models and the checkpoints used unless otherwise specified.Source: Table 1 · p. 4

3.2 EVALUATION PROTOCOL

  • The protocol asks whether biological content affects predictions, whether higher performance corresponds to greater input contribution, whether rationales reflect evidence, and whether auxiliary supervision increases input use.When perturbations have little effect, the protocol tests target predictability and, where possible, controls other predictive inputs.Source: §3.2 · pp. 4–5

4 RESULTS: BIOLOGICAL INPUT USE IN REASONING MODELS

Across the evaluated tasks, biological inputs contribute unevenly to performance: some models rely little on foundation-model representations despite their predictive information, while others respond to biological inputs. Higher benchmark accuracy and generated rationales do not consistently indicate greater or faithful use of those inputs.

4.1 RQ1: FOUNDATION MODEL REPRESENTATIONS CAN CONTRIBUTE LITTLE TO PERFORMANCE EVEN WHEN THEY ENCODE TASK INFORMATION

  • ChatNT, Prot2Text-V2, C2S-Scale, and CellWhisperer performance responds to biological inputs, with C2S-Scale especially affected by removing differentially expressed genes.C2S-Scale accuracy falls after gene-order shuffling or resampling, and removing DEGs affects accuracy more than removing expression-matched non-DEGs.Source: §4.1 · pp. 5–6, Figure 2 · p. 5, Figure 5 · p. 7
  • Shuffling DNA barely changes BioReason accuracy, and evidence conflicts show BioReason and BioReason-Pro follow text, although probes predict targets from Evo2 and ESM3 representations.BioReason accuracy is 0.847 with either intact or shuffled DNA; in conflicts, BioReason and BioReason-Pro follow text in 97.9% and 99.7% of cases, respectively.Source: §4.1 · pp. 5–6, Figure 2 · p. 5, Figure 3 · p. 6
  • Linear probes predict disease and protein-function targets from Evo2 and ESM3 representations, despite their limited contribution to BioReason and BioReason-Pro performance.Evo2 output and projected-representation probes reach 0.736 and 0.735 disease accuracy; ESM3 probes reach mean per-family AUROCs of 0.766, 0.898, and 0.830 for the three tested properties.Source: §4.1 · p. 6, Figure 4 · p. 7

4.2 RQ2: POST-TRAINING RAISES OVERALL TASK ACCURACY WITHOUT INCREASING THE BIOLOGICAL INPUT’S PERFORMANCE CONTRIBUTION

  • Higher BioReason and BioReason-Pro accuracy after RL is not accompanied by greater performance contributions from Evo2 or ESM3 representations.Across 42 BioReason checkpoints, accuracy varies without a consistent increase in the advantage of original over shuffled-DNA-derived representations.Source: §4.2 · p. 7, Figure 6 · p. 8
  • Unlike BioReason and BioReason-Pro, C2S-Scale 27B shows a larger DEG-specific accuracy effect than 2B when DEGs are removed.At full DEG removal, the DEG-specific difference is 0.319 for 27B versus 0.223 for 2B.Source: §4.2 · pp. 7–8, Figure 5 · p. 7

4.3 RQ3: REASONING TRACES AND RATIONALES DO NOT REFLECT BIOLOGICAL INPUTS

  • BioReason traces often state incorrect substitution bases, while BioReason-Pro traces can describe a functional property omitted from the final GO prediction.C2S-Scale rationales show early DEG enrichment, but full rationales overrepresent ribosomal genes and dilute DEG enrichment.Source: §4.3 · pp. 8–9, Figure 7 · p. 8

4.4 RQ4: AUXILIARY SEQUENCE SUPERVISION INCREASES SENSITIVITY TO FOUNDATION MODEL REPRESENTATIONS WITH LITTLE TRANSFER TO DISEASE PREDICTION

  • Auxiliary sequence supervision increases BioReason’s edited-base sensitivity, but gains concentrate on previously seen genomes and disease accuracy changes little under DNA shuffling.On unseen genomes, edited-base accuracy is 0.148 and 0.106 for 257-bp and 2,048-bp models, respectively, versus 0.111 for always predicting the most frequent training base pair.Source: §4.4 · p. 9, Figure 8 · p. 9

5 DISCUSSION

Biological reasoning models can achieve strong benchmark performance while making limited use of predictive biological inputs, and neither predictive representations nor post-training gains ensure those inputs contribute. The evaluation is limited to six models, specific tasks, and inputs; the appendix records datasets, checkpoints, prompts, scoring rules, and uncertainty estimates.

5 DISCUSSION

  • Findings may not generalize beyond six evaluated models, specific tasks, and biological inputs; shuffled sequences can also fall outside the distribution of natural sequences.Evidence conflicts built from unmodified inputs support the same conclusions, while probes establish target predictability but not how the language model uses a representation.Source: §5 · p. 10
  • High benchmark performance can coexist with limited contributions from predictive biological inputs, and neither access to predictive representations nor SFT or RL gains ensures their use.The paper argues that correct use of task-relevant inputs should be considered alongside predictive accuracy.Source: §5 · p. 10

A DATASETS, MODELS, AND EVALUATION METHODS

A.1.1 GENOMIC DATASETS AND EVALUATION SPLITS

  • BioReason disease prediction uses 1,449 KEGG-derived queries across 708 genome pairs, 186 prompt texts, and 37 disease labels, with all queries evaluated in paired comparisons.The dataset includes reference and variant sequences encoded with Evo2; the disease labels form the closed scoring vocabulary.Source: §A.1.1 · p. 15
  • ChatNT evaluation samples up to 1,000 queries per class for each of three binary tasks, totaling 4,578 queries balanced within task.The tasks are splice donors, splice acceptors, and TATA promoters.Source: §A.1.1 · p. 15

A.1.2 PROTEIN DATASETS AND ANNOTATION PROCESSING

The appendices specify the datasets, model inputs, perturbations, evidence conflicts, and probing procedures used to evaluate whether biological inputs contribute to model performance.

A.1.2 PROTEIN DATASETS AND ANNOTATION PROCESSING

  • The BioReason-Pro proteome contains 14,102 reviewed human proteins with curated GO annotations; its 91-protein temporal holdout is interpreted only through intact-versus-shuffled contrasts.The temporal holdout contains proteins added on or after 2026-03-20; its Fmax is not comparable with the proteome.Source: §A.1.2 · p. 15

A.1.3 SINGLE-CELL ATLASES AND CELL SELECTION

  • Five single-cell atlases contribute 4,846 cells, including 2,000 sampled for rationale analysis; CellWhisperer uses the same 4,846 cells.The atlases contribute 900, 1,049, 957, 939, and 1,001 cells, respectively.Source: §A.1.3 · p. 16, Table S1 · p. 16

A.2 EVALUATED MODELS

  • The evaluated models receive different biological inputs, including projected Evo2 or ESM3 representations, gene sentences, transcriptome embeddings, and biological text.BioReason combines Evo2 with gene and pathway text; BioReason-Pro combines ESM3 with GO-GPT predictions and InterPro annotations.Source: §A.2 · pp. 16–17

B.1 BIOLOGICAL INPUT PERTURBATIONS

  • Input perturbations shuffle DNA or protein sequences, alter gene sentences, or permute transcriptome counts while retaining the task prompt and other inputs where specified.For C2S-Scale, resampling changes visible gene identities while preserving list length and rank structure; CellWhisperer recomputes its representation from permuted counts.Source: §B.1 · pp. 18–19

B.2 EVIDENCE CONFLICTS

  • Evidence conflicts pair one entity’s biological representation with another’s text, testing whether predictions follow the representation source or text source.BioReason-Pro conflicts vary GO-GPT and InterPro sources while holding the organism string with the representation source.Source: §B.2 · pp. 20–21

B.3.1 GENOME-DEPENDENT DISEASE PREDICTION SUBSET

  • The genome-dependent disease subset contains 165 queries in which repeated prompt text occurs with different genomes and disease labels.A text-only predictor assigns the most frequent training-fold label for each prompt group and scores 0.708 across the 165 queries.Source: §B.3.1 · p. 21

B.3.2 INTERPRO-CONTROLLED PROTEIN SUBSET AND FUNCTIONAL TARGETS

  • Linear probes test whether Evo2 and ESM3 representations encode targets, using genome-grouped or InterPro-family-grouped folds to limit related examples crossing folds.Protein probes use AUROC averaged equally across InterPro families for three functional properties; GO NOT pairs are excluded because their readout terms differ.Source: §B.3.1 · p. 21, §B.3.2 · p. 21

B.4 REMOVING DIFFERENTIALLY EXPRESSED GENES FROM CELL SENTENCES

  • Cell2Sentence-Scale DEG perturbations remove selected fractions of a cell’s differentially expressed genes and compare them with random or expression-matched non-DEGs.DEGs require log2 fold change of at least 1.0 and adjusted p-value below 0.05; removed genes are deleted rather than replaced.Source: §B.4 · pp. 21–22

B.5 DISCUSSION OF TRAINING-TIME AVAILABILITY OF PREDICTIVE TEXT

BioReason and BioReason-Pro have answer-predictive text available during training, unlike several comparison models; this pattern is observational, not a controlled test of training prompt content.

B.5 DISCUSSION OF TRAINING-TIME AVAILABILITY OF PREDICTIVE TEXT

  • A training-set majority-vote table using the queried gene predicts BioReason’s disease label with 0.932 test accuracy, and BioReason-Pro training uses InterPro and GO-GPT without dropout.ChatNT lacks sequence-specific text, Prot2Text-V2 independently masks protein name and taxon with probability 0.8, and C2S-Scale trains on the cell sentence alone.Source: §B.5 · p. 22
  • Models trained with always-present answer-predictive text show nearly unchanged sequence-shuffling performance and prefer text in conflicts, but architecture, task, data, and scale prevent attributing this pattern to prompt content.The authors describe this as an observation and do not test it with a controlled intervention.Source: §B.5 · p. 22

C RQ2: POST-TRAINING, MODEL SCALE, AND BIOLOGICAL INPUT

The section is headed “Contribution to performance.”

C RQ2: POST-TRAINING, MODEL SCALE, AND BIOLOGICAL INPUT

C.1 BIOREASON POST-TRAINING SWEEP CHECKPOINTS

Across BioReason checkpoints, post-training changes accuracy without consistently increasing the performance contribution of genomic representations; the supplied trace analyses also document mismatches between reasoning and final predictions.

C.1 BIOREASON POST-TRAINING SWEEP CHECKPOINTS

  • Across 42 BioReason checkpoints, higher accuracy does not consistently coincide with a larger advantage for intact ZEvo2 over shuffled-DNA representations.The sweeps vary model size, SFT and RL LoRA rank, and RL duration; SFT rank changes overall accuracy to approximately 0.75–0.88.Source: §C.1 · pp. 22–23, Figure S2 · p. 23

C.2 C2S-SCALE 2B AND 27B

  • The 2B and 27B C2S-Scale checkpoints are evaluated on the same 4,846 cells using the same questions and setup.Source: §C.2 · p. 23

D.1 BIOREASON REASONING TRACES

  • BioReason reasoning traces are assessed on held-out pathway-network queries under intact, repeated-identical, and shuffled-window inputs, with string matching extracting stated biological claims.Base-change accuracy is assessed on 111 single-substitution queries, while coordinates are compared for the 137 queries with available positions.Source: §D.1 · p. 23

D.2 BIOREASON-PRO REASONING TRACES

  • In all 53 analyzed BioReason-Pro traces, the final response omits the GO term implied by the referenced InterPro accession, even though 23 traces state its property.The cited accession matches protein B’s accession in all 53 traces.Source: §D.2 · p. 24

D.3.1 RATIONALE GENERATION

  • The 27B C2S-Scale checkpoint generates a Rationale field in 12,000 responses, whereas the 2B checkpoint generates none.The 12,000 responses cover 2,000 cells across six conditions; the rationale analysis therefore reports the 27B checkpoint only.Source: §D.3.1 · p. 24

D.3.2 GENE EXTRACTION AND ANALYSIS OF REFERENCED GENES

The section defines how referenced genes are extracted and evaluated against cell-type differentially expressed genes (DEGs). It also specifies random-draw baselines and analyses of gene composition and set stability under input perturbations.

D.3.2 GENE EXTRACTION AND ANALYSIS OF REFERENCED GENES

  • Referenced genes are extracted from generations by tokenizing, upper-casing, filtering to the vocabulary, and retaining each gene only at its first mention.First-mention order determines the genes used in first-k precision.Source: §D.3.2 · p. 24
  • DEG precision among the first k referenced genes counts genes that are in the input sentence and are DEGs for the annotated cell type.The random reference is the expected precision from drawing the same number of genes uniformly from the input sentence.Source: §D.3.2 · p. 24
  • Figure S3a categorizes referenced genes by whether they occur in the sentence, are DEGs, or are ribosomal; ribosomal counts include ribosomal genes absent from the sentence.Sentence composition is compared with a same-size uniform draw from sentence genes.Source: §D.3.2 · p. 25, Figure S3 · p. 25
  • Figure S3 compares referenced-gene composition across five atlases, first-k DEG precision, and Jaccard similarity after removing strongest DEGs or similarly expressed non-DEGs.The Jaccard comparison is between the unmodified and perturbed referenced-gene sets, computed per cell and averaged.Source: §D.3.2 · p. 24, Figure S3 · p. 25

D.3.3 INSTRUCTING THE RATIONALE TO NAME DEGS

The analyses test whether prompt changes can make C2S-Scale rationales name DEGs, whether rationale gene content tracks prediction correctness, and whether requesting a rationale changes predicted labels. DEG-focused instructions yield little change in pooled precision, while rationale requests alter some labels and lower pooled accuracy.

D.3.3 INSTRUCTING THE RATIONALE TO NAME DEGS

  • The DEG-instruction comparison uses seven query variants on 200 cells, including marker-gene requests, exclusions, supplied cell labels, and one-shot examples.The baseline on these cells has 11.2% DEG precision, 23.2% first-five precision, and a 77.9% ribosomal share.Source: §D.3.3 · pp. 25–26
  • Pooled DEG precision across seven queries is 10.4–13.3%, versus 11.2% for the original query; the largest paired increase is 2.1 points for a marker-gene request excluding ribosomal, mitochondrial, and housekeeping genes.That request lowers first-five precision by 1.4 points; one-shot learning lowers accuracy from 0.40 to 0.18.Source: §D.3.3 · p. 26, Table S5 · p. 26

D.3.4 RATIONALE CONTENT BY PREDICTION CORRECTNESS

  • For C2S-Scale 27B, first-five DEG precision is 35.2% for correct predictions and 13.4% for incorrect predictions; among incorrect predictions naming an atlas type, genes align more with the predicted type’s DEGs.The latter comparison is 24.2% against predicted-type DEGs versus 13.7% against annotated-type DEGs.Source: §D.3.4 · p. 26, Table S6 · p. 27

D.3.5 EFFECT OF THE RATIONALE REQUEST ON THE PREDICTED LABEL

  • Adding a rationale request changes the resolved cell type for 18.2% of cells and reduces pooled accuracy from 0.45 to 0.41, a 4.2-point decrease.Accuracy decreases in four atlases but rises from 0.26 to 0.34 in lung; the prompts differ only in the rationale request and accuracy uses Cell Ontology resolution.Source: §D.3.5 · pp. 26–27, Table S7 · p. 27

E.1 DATA

The DNA evaluation uses a pathway-network-held-out split of BioReason queries and tests an auxiliary target for identifying edited bases. Training follows the established BioReason setup, with a separate weighted loss for edit and disease tokens in the auxiliary condition.

E.1 DATA

  • The pathway-network split contains 1,159 training, 145 validation, and 145 held-out queries; 135 held-out genomes also occur in training under another prompt.The held-out set covers eight disease labels, with 60 amyotrophic lateral sclerosis queries; mutations include substitutions, deletions, insertions, and multi-base substitutions.Source: §E.1 · p. 27, Table S8 · p. 28

E.2 AUXILIARY TARGET AND TRAINING

  • The auxiliary target replaces the reasoning trace with one line, Edit at i: ref X -> var Y (class), while aligned 257- or 2,048-bp windows place the edit at a fixed offset; the control keeps the benchmark reasoning trace.Source: §E.2 · p. 28
  • The text model is Qwen3-4B after continued pretraining, adapted with LoRA; runs use AdamW, three seeds, and checkpoints selected by lowest validation loss.The setup trains 66.1M parameters for 12 epochs with learning rate 5 × 10^-5 and batch size 1 with eight gradient-accumulation steps.Source: §E.2 · p. 28

E.3 EVALUATION AND RESULTS

The evaluation compares intact DNA, shuffled DNA, and donor-query DNA conflicts while holding the prompt fixed. It reports edited-base-pair accuracy and conflict behavior, including a split by whether the genome appeared in training.

E.3 EVALUATION AND RESULTS

  • Each held-out query is evaluated with intact DNA, shuffled DNA, and a random held-out donor’s DNA; the written base pair determines edited-pair accuracy.The text-only disease reference reaches 0.917, while always writing the most frequent training variant pair reaches 0.103.Source: §E.3 · p. 28
  • Table S10 separates edited-base-pair accuracy by genome training overlap; only 72 of 290 held-out and validation queries have genomes absent from training.The absent-genome group comprises 10 held-out and 62 validation queries.Source: §E.3 · p. 28, Table S10 · p. 29
  • Table S9 reports mean edited-base-pair accuracy across three seeds for 145 held-out queries, intact-minus-shuffled accuracy differences, and the share of conflicts following the donor’s pair.Figure S4 visualizes normalized predicted reference-to-variant base pairs for the intact, shuffled, and conflict conditions.Source: Table S9 · p. 29, Figure S4 · p. 29

Related papers

Loading 2610.00898v1…