Kristin Gleitsman

Leadership in Biotech

Tag: predictions

Illustration of a desk with a figure from a recent paper about the pLM

From Embeddings to Insight

By

On November 15, 2025

In AI ∩ Bio

Protein language models (pLMs) learn from raw amino-acid sequences and turn each protein into a numerical “embedding.” Unlike AlphaFold (which is trained to predict 3D structure), pLMs are trained only on sequence patterns, then reused for tasks like fast homology search, function hinting, or variant triage.

The appeal of pLMs is that they might:

  • Identify remote homology
  • Assist in function annotation, especially for uncharacterized proteins
  • Predict mutation effects
  • Serve as general-purpose inputs for fold prediction, domain classification, etc.

This study benchmarks 14 pLMs to evaluate how well they capture biological similarity along three axes:

  • Sequence
  • Structure
  • Function

For each axis, the authors compare distances between embeddings to a “ground truth” similarity metric (see below). They test whether models reflect these similarities in two ways:

  • Inherent information: Does the raw embedding distance between two proteins correlate with their similarity?
  • Extractable information: Can a small model trained on the embeddings predict the similarity score?

Scientific Insight

This paper provides a systematic, well-designed benchmark for protein language models, useful for those developing or deploying these models computationally.

But it defines “success” in terms of alignment with thorny labels, which is a limitation of the field at the moment. See PDF below.

With that caveat, the results are compelling:

Out-of-the-box, small models often perform as well as large ones.

  • “Size-performance paradox”: model size doesn’t guarantee better embedding quality for biological similarity.

Larger models encode more information, but it’s hidden.

  • You need to train a model on the embeddings to “extract” biological signal.

Task-specific models don’t generalize.

  • Fine-tuning a model for one problem (e.g., enzyme function) distorts the embedding space and reduces general usefulness.

Leadership Angle

Bigger isn’t automatically better (we’re seeing this with genomic models as well). But also be mindful of what you are benchmarking to, and its relevance to what you care about.

AI ∩ Bio

Illustration of scientific paper with the title "AI ∩ Bio: Reading the Revolution"

Mentorship Angle

Let pLMs speed the front end of discovery, then make the science real.

  • Start with a testable claim in plain language—e.g., “Nearest neighbors in embedding space enrich for shared catalytic residues better than a sequence-identity cutoff at the same recall.”
  • Pre-register thresholds; include hard negatives (same fold, different function; low identity, same mechanism) & use controls (shuffle labels, etc)
  • Report where it fails—disordered regions, multi-domain proteins, complexes, and say why you think it failed
  • Document assumptions (databases, structure confidence, training leakage)

The win isn’t a pretty plot; the win is turning speed into better experiments that change what you do at the bench.

View Fullscreen
Skip to PDF content
Illustration of a desk with a figure from a recent paper about AI interpretability in biology

What Are Genomic Transformers Actually Learning?

By

On August 16, 2025

In AI ∩ Bio

This week’s AI ∩ Bio: Reading the Revolution post explores a new paper tackling one of the biggest challenges in AI for biology: interpretability.

Specifically, how can we tell what genomic transformer models are learning and whether those internal patterns map to real biological concepts?

The study proposes a systematic framework to analyze attention heads in models like DNABERT and scGPT. It reveals that many attention heads naturally align with key biological features—such as transcription start sites, sequence motifs (e.g., TATAAA), GC content, conservation scores, and Gene Ontology categories like “neuron projection” or “mitochondrial membrane.” This interpretability emerges even during self-supervised pre-training and strengthens with task-specific fine-tuning.

Key Findings

  • Attention heads develop strong correlations with biologically meaningful features.
  • These heads tend to be more important for predictive performance, as shown by ablation experiments.
  • Interpretability is shaped by tokenization strategy (e.g., overlapping vs. non-overlapping k-mers).
  • GPT-4 was used to generate descriptive summaries of head activity (e.g., “TATA box detector”)—but these require validation, as they can miss label-specific nuance.

Limitations

  • Correlation ≠ causation: attention may highlight a feature without depending on it for prediction.
  • Current analyses treat attention heads in isolation—without clustering or organizing them into higher-level “concept modules.”
  • GPT-4 summaries, while useful, may overgeneralize or obscure context-dependent patterns.

Why This Matters

This work moves the field closer to a compelling vision: transformer models that don’t just output predictions, but expose interpretable biological patterns that scientists can name, reason about, and test. Much like ConcepPath in digital pathology that we covered earlier in this series, this paper hints at a future where AI supports concept-level ‘reasoning,’ not just classification.

AI ∩ Bio

Illustration of scientific paper with the title "AI ∩ Bio: Reading the Revolution"

For Early-Career Scientists

Interpretability isn’t just about trust, it’s also critical for scientific translation. These models may surface patterns that align with biology, but the real opportunity is in bridging those signals to concepts we recognize, can reason about, and critically, can test experimentally. That’s how interpretability becomes a tool for discovery, not because the model “knows” biology, but because we can now ask better questions about what it might be pointing to.

Powered by WordPress & Theme by Anders Norén