Reading a genomics model to find its dead weight

A screen that never saw the test found which parts of a DNA model carry almost nothing. Rebuilt without them, the model is 22% smaller and still ranks harmful mutations slightly better.

Genomics models are used to guess which changes in a person’s DNA are likely to cause disease. GPN-Star does this by learning to fill in blanks. Show it a stretch of genome with one position hidden and it predicts which letter belongs there; a real mutation that the model finds surprising, one it would not have predicted, tends to be a mutation that matters. The test used throughout this page asks how well that surprise separates known-harmful human variants from harmless ones, using ClinVar, a database of clinically interpreted variants.

A model like this is expensive to train, and every parameter it carries is capacity someone paid for. The ordinary way to find out whether some of that capacity is doing nothing is to remove a piece, retrain, and see whether the score drops. That is a slow way to learn anything, because each answer costs a full training run.

The result here is that a good deal of the waste can be found first, for free. A redundancy screen was shown only the trained model, never the mutation-scoring task it would eventually be graded on, and it identified parts of the network carrying almost no information. A model rebuilt without them, and trained again from random initialization, is 22% smaller than the published one and still scores slightly better on ClinVar. The screen is therefore useful as a design tool that sizes the next model, not only as an explanation of the current one.

The model, and what its parts do

GPN-Star reads DNA as an alignment rather than a single sequence: the human stretch, with the matching stretch from many other species stacked underneath it. That shape gives each of its sixteen layers three jobs. Row self-attention looks along the sequence, comparing positions in one genome. Column cross-attention looks down a single column, comparing species at one position, which is how the model learns that a position conserved across mammals is probably load-bearing. A feed-forward block then does per-position computation. Sixteen layers times three blocks is forty-eight places where capacity could be going to waste.

Four models, one control

Drag the slider to step from the published model down to the smallest rebuild. The track shows which of the forty-eight blocks each architecture keeps, with the green band drawn to scale so you can see column attention narrowing. The reading on the right is the ClinVar score, against the published model’s mark.

Every rebuild scores above the published model, and the smallest of them sits within 0.0018 of the largest. That gap is the honest limit on how much their ordering means, and it is discussed at the end.

What the screen looked at, and what it noticed

The screen examined two things about the trained model: the weights, and the activations, meaning the numbers actually flowing through each block when real DNA goes in. The weights say nothing useful. Column attention’s weight matrices are close to full rank, which is the mathematical way of saying that nothing about the stored parameters looks wasteful.

The activations say the opposite. Of the 64 dimensions available to each column-attention head, the realized activity in the query path occupies about 22 of them, and in the value path about 3. The model was given a wide channel and learned to use a narrow one. That gap between a full-rank weight matrix and a nearly flat activation is exactly the kind of thing a screen can see without running the downstream task.

How narrow is narrow? Keeping only 32 of the value path’s 256 dimensions preserves almost all of the variation in what it produces.

The cheapest possible test of whether that matters is to make the cut on the finished model, without retraining anything at all, and re-score it. Narrowing the value path to 32 dimensions on frozen weights costs 0.00005 of ClinVar AUROC. Squeezing the query path down to rank 8 moves it by 0.00006. Those are changes in the fifth decimal place, against a score that lives in the fourth.

A separate check asks the same question in the model’s own currency rather than the disease benchmark’s. On held-out DNA from chromosome 22, sequence the model was never trained on, truncating the query path to 16 dimensions and the value path to 32 changes perplexity by −0.0004, which is very slightly better than leaving it alone. Removing the same amount of capacity along randomly chosen directions instead does not come out that way, so the screen is finding specific structure and not merely discovering that the model tolerates being made smaller.

Rebuilding the model is what actually settles it

Everything above is a measurement of what is readable in a trained model. None of it proves the removed capacity was unnecessary, because a frozen model that tolerates a cut might still have needed that capacity to get where it is. The claim only becomes load bearing when the smaller design is built from scratch, trained, and scored, and that is where the evidence here comes from.

Retraining the 174.0M architecture from random initialization reaches 0.8441 on ClinVar at step 300000. That is −0.0002 from the 0.8443 of the original run at that size, and +0.0096 above the published 0.8345. The training curve crossed the published line at around step 90k and stayed above it for the rest of the run, peaking at 0.8443 at step 285k. So the result is not a fluke of one training run: an independent run of the same recipe lands in the same place.

A second benchmark points the same way. On a frequency-balanced set of variants drawn from gnomAD, a catalogue of natural human variation rather than clinical annotation, the retrained model scores 0.7254, against 0.7236 for the original run at that size and 0.7212 for the published model.

Pushing further, the value-path cut the screen predicted brings the model to 166.6M at 0.8438, and dropping the feed-forward block in layer 0 as well brings it to 158.2M at 0.8425. That last one is 22% smaller than the released model and +0.0080 above it.

Where the screen stops working

The screen did not only recommend narrowing column attention. It also nominated whole blocks to delete, and that half of its output does not hold up as a static read. On frozen held-out perplexity its layer-skip prescription is harmless, changing perplexity by −0.0009. But a matched set of randomly chosen layer skips is also harmless, and in fact came out slightly better at −0.0026. Measured this way the prescription is indistinguishable from picking blocks at random, so this comparison supports nothing about which blocks to drop. Only the column-attention cuts have been shown to be specific.

A search that is allowed to look at perplexity does better than a static read, but still only partway. Greedily removing whichever block hurts least picks row attention in layer 15 first, which is one of the three blocks the hand-derived architecture drops, and skipping it alone improves held-out perplexity by about 0.10. Its three-block set, row attention in layers 15 and 14 plus the feed-forward block in layer 14, gives −0.0612 on chromosome 22, better than 96% of matched random three-block draws, whose average is +0.0239. That is a real signal. But it overlaps the hand-derived set in only 1 of 3 blocks.

The reason for the mismatch is instructive. Removing the feed-forward block in layer 12, or row attention in layer 13, makes frozen perplexity worse when you do it to the finished model. No method that reads the finished model will ever nominate them. Their value only shows up when the smaller model is retrained from scratch, which means what they buy is a training-dynamics effect rather than redundant capacity, and no amount of staring at the trained weights can reveal it.

It is also worth naming a method that fails outright here. CKA, a widely used measure of how similar two layers’ representations are, is a natural candidate for spotting a redundant layer. Its top recommendation is row attention in layer 0, and removing that moves perplexity by +2.592, a catastrophic choice by the standards of every other number on this page.

One more detail cuts against a tidy story. When the task-agnostic screen’s ranking of all 48 blocks is compared with the earlier ranking that did use the downstream task, the two agree at a Spearman correlation of +0.55 (p = 5.8e-05). The agreement is strongest for feed-forward blocks, at +0.79, and weakest for column attention, at +0.12, which is precisely where the screen’s prescription turned out to be most trustworthy. Broad agreement between the two rankings is not what makes the column-attention cut credible; the random-subspace control and the retrained model are.

What this establishes, and what it does not

The narrow result is solid: in this model, column attention’s value path is real redundancy, it can be found by looking only at the trained network, and a model rebuilt without it and retrained from scratch is substantially smaller without giving up ClinVar accuracy. Four things bound how far that travels.

• One seed each. Every architecture here was trained once, and the three rebuilds fall within 0.0018 of each other, which is small enough that their ordering among themselves is not established. What is established is that all three sit above the published mark. Confirming any of it with independent seeds has not been done, and that is the outstanding gap.
• One benchmark, one metric. This is variant ranking on ClinVar, with the gnomAD comparison as corroboration. It is not a measure of general capability. A smaller model that scores as well at ranking mutations has not been shown to be as good at everything the original does, and clinical variant databases carry their own known biases in which genes and which variants they cover well.
• The screen found the wasted width, not the wasted depth. Low-rank column attention was rediscovered without any reference to the downstream task and survives a matched random control. The block removals largely do not: most only pay off after retraining, so a static read of the trained model cannot select them, and one popular similarity measure selects catastrophically.
• The variance and similarity analyses are correlational. They say what is readable in the weights and activations. The claim that the removed capacity was not needed rests on rebuilding the model and retraining it, and that retraining is the load-bearing evidence for everything asserted here.

The practical reading is narrower than “interpretability finds dead weight” and more useful than it sounds. For one specific and common kind of waste, a channel the model was given and never learned to fill, you can measure it on the finished model, in minutes, and act on it when sizing the next one. For the other kind, capacity that only matters while the model is learning, you still have to train something to find out.

All scores are per-variant AUROC over the model’s masked-language log-likelihood ratio at the variant position, at most 5000 variants, sequence length 256, computed in fp32 deterministically. The published figure is that same harness scoring the released GPN-Star checkpoint, not a number copied from a paper. Perplexity comparisons use held-out chromosome 22.