#4 · ML / AI methods
Learned latents steer protein conformations and encode evolutionary structure
The dominant thread is the use of learned latent or generative representations as handles on biological structure. ConforNets applies learned channel-wise affine transforms to AF3 pre-Pairformer pair latents to globally steer conformational sampling, and AlphaInterp probes AF3 internals, including Pairformer latent geometry and causal activation patching of confidence. What is new is the use of pretrained model internals as controllable sampling handles. HyperEvoGen learns hyperbolic Poincare variational autoencoder protein sequence embeddings whose distances track evolutionary divergence. Quotient-space diffusion, GeomMotif, and Surleton address generation, motif benchmarking, and inverse folding without the same latent-steering mechanism.
Published papers
Latent Style-Based Quantum Wasserstein GAN for Drug Design.
Proposes a latent style-based quantum Wasserstein GAN with a VAE latent space for de novo molecule generation, benchmarked against classical latent GANs on…
PDI-Spot: A Robust Representation Learning Framework for Identifying Energetic Hotspots in Protein-DNA Complexes.
PDI-Spot, a representation-learning model using structure- and sequence-derived residue embeddings, predicts energetic hotspots at protein-DNA interfaces…
PRIDA: Cross-domain protein-RNA interaction prediction based on biological language models and conditional domain-adversarial learning.
PRIDA, a bilinear-attention framework with conditional domain-adversarial learning and pretrained protein/RNA language models, predicts protein-RNA…
Preprints
Building from the ground up: augmentation strategies to improve protein language model performance on protein sets
Probes how protein language models learn from small protein datasets and tests augmentation strategies, including ESM-3 generated complexes, to improve…
Neopep: Reinforcement Learning From Machine-Learned Force Field Energy for Non-Canonical Peptide Structure Prediction
Neopep is a sequence-conditioned diffusion model for non-canonical amino acid peptide structure prediction, pretrained on PDB and fine-tuned with RL using…
Topic history
| Date | Rank |
|---|---|
| #4 | |
| #5 | |
| #3 | |
| #2 | |
| #3 | |
| #4 | |
| #4 | |
| #4 | |
| #2 | |
| #1 | |
| #4 | |
| #5 | |
| #3 | |
| #3 | |
| #2 | |
| #2 | |
| #4 | |
| #2 | |
| #4 | |
| #4 | |
| #5 | |
| #3 | |
| #3 | |
| #6 | |
| #10 | |
| #12 | |
| #8 | |
| #4 | |
| #15 | |
| #16 |