Modalities

• Sequence, graph, structure and text encoders and what each captures
• Transcriptomic and clinical data as additional evidence about a target
• Fusion strategies — early, late and cross-attention — and their failure modes

Generation

• SMILES, graph and 3D diffusion generators compared on validity and novelty
• Structure-conditioned generation and the tendency to produce unsynthesisable output
• Synthetic accessibility scoring and retrosynthesis checks as a required gate

Proteins

• RFdiffusion and ProteinMPNN in the design-then-sequence workflow
• In silico filtering of designs before any wet-lab commitment
• Reported success rates and why most designed binders still fail

Judgement

• Benchmark leakage and why held-out targets matter more than headline metrics
• Where multimodal models genuinely beat single-modality baselines, and where they do not
• Translating a computational hit into a testable experimental plan