When Do Biological Reasoning Models Use Their Biological Inputs?

AI summary

Overview

Research area: Machine learning for biology, specifically multimodal LLMs that combine language models with biological foundation model representations and biological text (genomics, proteins, single-cell transcriptomics).

Technical level: Intermediate. The paper is an evaluation/benchmarking study; the concepts (perturbation, probing, evidence conflict) are explained in plain terms, but familiarity with biological foundation models (Evo2, ESM3, Nucleotide Transformer, ESM2) helps.

One-sentence scope: The paper tests, across six biological reasoning models and four kinds of controlled experiments, whether these models actually use the biological inputs they are given.

What This Paper Is About

Biological reasoning models are LLMs post-trained to accept biological inputs — foundation model representations of DNA, proteins, or cells, plus biological text such as gene descriptions, pathway definitions, and GO or InterPro annotations — and their benchmark accuracy is typically taken as proof that the model reasons over those inputs. This paper asks whether that inference is valid: does changing the biological content of an input actually change the model's answer? The authors build perturbation, evidence-conflict, linear-probe, and reasoning-trace tests to separate "this input is predictive of the target" from "this model uses this input."

Key Contributions

1. A four-test protocol for attributing predictions to biological inputs: (1) Perturbation — change one biological input while holding the query and other inputs fixed; (2) Evidence conflict — place inputs from two different genomes, proteins, or cells in one prompt and see which the model follows; (3) Linear probe — test whether the representation the model receives predicts the target; (4) Reasoning trace — check whether generated biological claims agree with the supplied inputs and with final predictions.
2. An evaluation of six biological reasoning models across three modalities: BioReason and ChatNT (DNA), BioReason-Pro and Prot2Text-V2 (protein), and Cell2Sentence-Scale (C2S-Scale) and CellWhisperer (single cell), evaluated across post-training checkpoints and model scales.
3. A separation of "informative" from "used": the paper shows that Evo2 and ESM3 representations encode task targets in linear probes yet contribute little to BioReason and BioReason-Pro task performance, while foundation model inputs do contribute to ChatNT, Prot2Text-V2, and CellWhisperer, and DEGs contribute to C2S-Scale.
4. Evidence that accuracy gains from post-training do not imply greater input use, demonstrated across SFT and RL checkpoints of BioReason-Pro and across 42 BioReason checkpoints, plus an auxiliary-supervision experiment (RQ4) that increases sequence sensitivity without improving disease prediction much.

Main Findings

• BioReason barely responds to shuffled DNA. Disease prediction accuracy is almost unchanged with shuffled DNA inputs (0.847 vs. 0.847), although shuffling changes 101 of 1,449 answers (7.0%). The largest accuracy decrease occurs when the gene name is removed.
• BioReason and BioReason-Pro follow the text in evidence conflicts. BioReason predictions are consistent with the gene and pathway text source in 97.9% of conflicts; BioReason-Pro is consistent with GO-GPT in 99.7% of conflicts with InterPro and ESM3 inputs. On pairs where one protein has an experimental GO NOT annotation, responses are consistent with the GO-GPT and InterPro sources in 98.1% of 106 conflicts.
• BioReason-Pro is nearly insensitive to protein sequence. Supplying the ESM3 embedding of a shuffled sequence with GO-GPT and InterPro unperturbed changes F_max by 0.855 vs. 0.856 across 14,102 human proteins (0.685 vs. 0.682 on a temporal holdout). The shuffled sequence does change the predicted GO term set for 46.8% of the 14,102 proteins, with mean Jaccard similarity 0.89.
• Some models do use their foundation model inputs. ChatNT accuracy drops from 0.988 to 0.502 for splice donors, 0.987 to 0.502 for splice acceptors, and 0.971 to 0.577 for TATA promoters under shuffled sequence. Prot2Text-V2 BioBERT BERTScore F1 decreases by 0.177 with name and taxon removed, and it favors the sequence source in 97.0% of conflicts with a conflicting protein name.
• Single-cell inputs matter to both single-cell models. C2S-Scale 27B mean cell type accuracy falls from 0.446 to 0.351 after gene order shuffling and to 0.406 after resampling genes. CellWhisperer's 1 − perplexity quantile falls in every atlas after permuting counts across genes, including 0.98 to 0.53 in immune1 and 0.84 to 0.25 in lung; it assigns lower perplexity to the representation-supported answer in 96.1% of conflicts.
• Underused representations are still predictive. On 165 genome-dependent disease queries, probes on the Evo2 output and on the projected Evo2 representation reach 0.736 and 0.735 accuracy versus 0.708 for a text-only predictor; the improvement is driven by queries with gene PRNP (98 of 165), where probe accuracy is 0.855 versus 0.708. For 540 proteins across 140 InterPro families, ESM3 probes reach mean per-family AUROCs of 0.766, 0.898, and 0.830 (0.713, 0.921, 0.898 on projected representations), compared with 0.729, 0.890, and 0.736 for BioReason-Pro.
• Higher accuracy after post-training does not mean more input use. For BioReason, SFT scores 0.823 with original Evo2 and 0.814 with shuffled-derived Evo2, while RL scores 0.847 and 0.847. For BioReason-Pro, RL raises F_max from 0.829 (SFT) to 0.855, but swapping in a shuffled-sequence ESM3 changes 0.829 to 0.833 for SFT and 0.855 to 0.856 for RL. Across 42 post-training checkpoints, accuracy varies without a consistent increase in the advantage of the original Evo2 representation.
• One model scales favorably. C2S-Scale mean cell type accuracy rises from 0.366 (2B) to 0.446 (27B), and the DEG-specific accuracy gap grows: at 100% DEG deletion, 2B scores 0.140 after removing strongest DEGs vs. 0.363 for expression-matched non-DEGs (difference 0.223), while 27B scores 0.127 vs. 0.446 (difference 0.319).
• Reasoning traces misstate inputs. On 145 held-out pathway-network queries, BioReason states the correct base for 7.2% (RL) and 3.6% (SFT) of substitutions; the RL checkpoint states the variant is T>C on all 145 responses even when the Evo2 input is shuffled. In 23 of 53 filtered BioReason-Pro evidence-conflict traces, the model names the tested property in words not in any printed InterPro entry name, yet emits a GO set that omits it.
• C2S-Scale rationales over-reference ribosomal genes. Rationales reference 63.4–67.8 genes, of which 93.9–98.7% occur in the gene sentence and 62.6–84.7% are ribosomal protein genes; mean DEG precision across all referenced genes is 10.63%, below the random-draw reference, though the first five references show enrichment. For 452 incorrect annotations, precision against predicted-type DEGs is 24.2% vs. 9.9%. Requesting a rationale lowers pooled accuracy from 0.45 to 0.41.
• Auxiliary supervision increases sequence sensitivity but barely transfers. Training BioReason to predict edited bases yields edited base pair accuracy of 0.671 with intact DNA and 0.122 after shuffling (257-bp windows); with 2,048-bp windows, 0.531 and 0.099, versus 0.080 without auxiliary supervision. Shuffling lowers edited base pair accuracy by 0.549 but disease prediction accuracy by only 0.021. On 72 queries with unseen genomes, accuracy falls to 0.148 and 0.106, compared with 0.111 for always predicting the most frequent training base pair.

Methodology in Plain English

The authors treat each model as a function that takes a query plus biological inputs and returns an answer (and often a rationale). Rather than only scoring answers against labels, they change one input at a time and watch what happens to the answer and the score.

• Perturbation: recompute a foundation model representation from a shuffled sequence, reorder or resample a gene sentence, permute counts across genes, or drop text fields such as gene name, pathway, name, taxon, GO-GPT, or InterPro — keeping everything else fixed.
• Evidence conflict: take a representation (Evo2 embedding, ESM3 embedding, or CellWhisperer transcriptome representation) from entity A and the text or gene list from entity B in a single prompt, so the two sources support different answers, then check which source the answer follows.
• Linear probe: fit a simple linear classifier on mean-pooled foundation model outputs and on the projected representations the language model actually receives, to test whether the target is decodable from the input at all.
• Reasoning trace analysis: compare the biological claims in generated traces (stated bases, recited annotations) against the supplied inputs and against the final predictions, including under evidence conflicts and sequence shuffling.
• Checkpoint and scale comparisons: repeat the same perturbations on SFT and RL checkpoints and across 42 BioReason post-training checkpoints (varying model size, SFT LoRA rank, RL LoRA rank, RL duration) and on C2S-Scale at 2B and 27B.
• Auxiliary supervision (RQ4): retrain BioReason with a short target specifying the reference and variant bases at the edited position while keeping the disease answer, then test whether Evo2 now affects the disease prediction.

Evaluation data include 1,449 KEGG-derived disease queries over 708 genomes and 37 disease labels (165 genome-dependent), 4,578 ChatNT test queries, 14,102 reviewed human proteins (plus a 91-protein temporal holdout and 540 proteins across 140 InterPro families), 3,917 proteins for Prot2Text-V2, and 4,846 cells across five atlases for the single-cell models.

Why This Matters

Benchmark numbers are routinely used as evidence that a model "understands" sequence or transcriptomic data. This paper shows that accuracy can be high while a predictive biological input contributes almost nothing, which means benchmark improvements and post-training gains can be misread as gains in biological grounding. The authors argue that correct use of task-relevant biological inputs should be an explicit objective alongside predictive accuracy, and release the BioMirage evaluation suite for this purpose.

Real-world applications:

• Clinical variant interpretation: a disease-prediction model that ignores the variant sequence (as shuffled DNA barely changes BioReason accuracy) is unsuited to prioritizing novel variants.
• Protein function annotation: GO predictions that omit functions the model itself described in its trace would mislead annotation pipelines relying on automated function assignment.
• Single-cell analysis: DEG-based cell type annotation depends on which genes the model actually weights, and rationale text that over-references ribosomal protein genes can mislead downstream biological interpretation.
• Model and benchmark development: builders of DNA, protein, and single-cell LLMs can adopt these tests to validate that their post-training objectives reward input use rather than text-side shortcuts.

Industry relevance: pharmaceutical and biotech groups integrating biological foundation models into target discovery, variant triage, and single-cell pipelines depend on knowing which inputs drive predictions; benchmark designers and model developers need evaluation signals that cannot be satisfied by annotation text alone.

Future Directions

• Evaluate additional biological reasoning models, architectures, and biological modalities beyond the six models and DNA/protein/single-cell tasks studied here.
• Develop post-training objectives, including reinforcement learning rewards for correct responses to changes in biological inputs, and test whether they increase the performance contribution of those inputs.
• Test training prompt variants within the same model while holding other training settings fixed, to isolate which design choices increase input use.
• Close the gap between predictability and use: probes show a target is decodable from Evo2 and ESM3 representations, but it remains open how the language model could use that representation, and how to make auxiliary supervision such as edited-base prediction generalize to unseen genomes.

Target Audience

Machine learning researchers working on multimodal and biological LLMs, benchmark designers, and computational biologists who consume these models. It is also useful for practitioners in biotech and pharma who select biological reasoning models based on reported benchmark accuracy, and for anyone studying whether generated reasoning traces can be trusted as explanations.

Authors’ abstract

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.