#3 · ML / AI methods

Pretrained biomolecular encoders reused across property prediction and generative design

The dominant thread is learned biomolecular representations coupled to a second modality or explicit constraint, spanning prediction and generation. MPMFMol pretrains a molecular graph encoder with fragment-based augmentations and fingerprint multitask learning before multimodal fine-tuning, while HKD-CPI aligns LLM-derived tokens with molecular graphs and applies knowledge distillation. Fusion architectures recur: DMPKformer combines Transformer, GAT, and MLP branches with latent OOD confidence, CM-MTL-DTI uses bidirectional cross-modal attention, and EpiFormer interleaves cross-attention within GNN layers. On the generative side, AlloGen separates backbone generation from a learned SE(3)-invariant state-selectivity scorer, and PepForge generates modified HELM peptides hierarchically. The copolymer MILP inverse-design work does not involve pretrained representations. What appears newer is coupling representations to structural or geometric constraints rather than task-specific training alone.