Artificial neural network×Computational biology

Artificial neural networks and computational biology have already shared one landmark collision in protein structure prediction, but the structural evidence shows the bulk of both communities remain siloed — meaning the deeper merger, spanning gene regulation, protein-protein interaction networks, and whole-cell simulation, is still ahead. The 81 dual-publishing authors and shared conceptual anchors in protein folding, transformation, and process computing signal that a new class of 'biological foundation model' — trained on sequence, structure, and interaction data simultaneously — is the imminent breakthrough zone. This will compress decades of wet-lab hypothesis cycles into weeks of generative in-silico design.

The bridge fields are unusually concrete: 'Protein folding' and 'Folding (DSP implementation)' are not metaphorical overlaps — they are the same mathematical object being worked on from two directions. AlphaFold appearing identically as the top-cited paper in both Field A and Field B paper lists is the clearest possible co-citation signal that a single paper already fused the epistemologies. Despite that, the graph shows no systematic direct co-publication community yet, which means the formalisation of the joint field — shared journals, shared benchmarks, shared training pipelines — is the next structural event. The 81 dual-affiliation authors are the translation layer; they will write the founding papers of the merged discipline over the next 24-36 months. Momentum in transformer architectures (bridge: Transformation/genetics) applied to genomic sequences (ESM, Evo, Nucleotide Transformer) is accelerating on a separate but convergent track from the phylogenetic/network biology side (IQ-TREE, STRING), setting up a collision in multi-omics graph-neural-network models.

Winners will come from three archetypes: (1) ML-native labs that have accumulated biological sequence datasets at scale and can treat genomes and proteomes as token streams for large language model pretraining; (2) computational biology groups inside large pharma with proprietary multi-omics phenotype data that can supervise fine-tuning of general biological foundation models on drug-relevant endpoints; (3) academic groups already sitting in the 81-author bridge set — specifically those publishing on protein interaction graphs and deep residual or attention architectures simultaneously, as they understand both the data topology and the architectural inductive biases required. Pure ML groups without biological domain depth and pure bioinformatics groups without large-scale training infrastructure will each be outpaced.

A protein-protein interaction foundation model trained jointly on structural embeddings (AlphaFold-derived) and graph topology from curated interaction databases (STRING-class), fine-tuned to predict network rewiring under genetic perturbation. The concrete experiment: benchmark whether a unified GNN-transformer architecture that ingests both sequence and known interaction edges can outperform STRING's co-expression and co-occurrence heuristics on held-out CRISPR screen phenotypes — if yes, it validates that structural deep learning can generalise from single-protein geometry to proteome-level network logic, unlocking a new layer of drug target prioritisation.