TY  - JOUR
T1  - Foundation Models for Insect Molecular Biology Must Learn across Genomes, Transcriptomes, Proteomes, Phenotypes, Behaviour, and Chemical Space
A1  - Sarah Jensen
A1  - Pieter Vos
A1  - Lars Olsen
A1  - Klaus Weber
JF  - Entomology and Applied Science Letters
JO  - Entomol Appl Sci Lett
SN  - 2349-2864
Y1  - 2026
VL  - 13
IS  - 1
DO  - 10.51847/1GGg63ErEt
SP  - 61
EP  - 74
N2  - Insect biology increasingly depends on heterogeneous molecular, organismal, and environmental observations that cannot be interpreted reliably through isolated analytical models. Genome assemblies, single-cell transcriptomes, proteomic and metabolomic measurements, chemical structures, images, behavioural trajectories, and ecological records describe different biological objects at different spatial and temporal scales. Yet current artificial-intelligence applications in entomology are commonly developed around one modality, one species, one laboratory setting, or one narrowly defined prediction task. This architecture article addresses the resulting representational gap by proposing a non-validated multimodal foundation-model structure for insect molecular biology. The approach organizes evidence-supported computational capabilities into modality-specific encoders, biologically conditioned alignment layers, shared and modality-private representations, provenance and uncertainty controls, and task-specific outputs. The synthesis integrates genomic and transcriptomic representation, proteomic and metabolomic learning, chemical-space modelling, computer vision, behavioural time series, ecological context, pretraining, transfer, and biological alignment. The strongest defensible conclusion is that broadly pretrained models may support reusable insect representations only when modality-specific meaning, taxonomy, life stage, tissue, environment, measurement process, and uncertainty remain explicit. Multimodal scale is therefore not equivalent to biological understanding; pretraining performance is not downstream scientific validity; representational alignment is not causal mechanism; and benchmark performance is not evidence of transfer across species, modalities, or environments. Major limitations include taxonomic imbalance, incomplete molecular annotation, sparse cross-modal observations, laboratory–field divergence, leakage, reference-atlas bias, and insufficient biologically independent validation. The central implication is that insect foundation models should be developed as auditable, context-conditioned scientific infrastructures whose outputs can be rejected, qualified, or experimentally tested rather than treated as universal biological explanations or deployment-ready digital twins.
UR  - https://easletters.com/article/foundation-models-for-insect-molecular-biology-must-learn-across-genomes-transcriptomes-proteomes-fg9gaquujywy1ct
ER  -