alphagenome-pytorch 0.0.1__py3-none-any.whl

alphagenome/__init__.py

@@ -0,0 +1,5 @@
+from alphagenome.alphagenome import (
+    AlphaGenome,
+    Attention,
+    TransformerTower
+)

alphagenome/alphagenome.py

@@ -0,0 +1,441 @@
+from __future__ import annotations
+from functools import partial
+
+import torch
+from torch import nn, cat, stack, arange
+import torch.nn.functional as F
+from torch.nn import Linear, Sequential, Module, ModuleList
+
+import einx
+from einops.layers.torch import Rearrange, Reduce
+from einops import rearrange, repeat, einsum
+
+# ein notation
+
+# b - batch
+# h - heads
+# n - sequence
+# d - feature dimension
+
+# constants
+
+LinearNoBias = partial(Linear, bias = False)
+
+# functions
+
+def exists(v):
+    return v is not None
+
+def divisible_by(num, den):
+    return (num % den) == 0
+
+def is_odd(num):
+    return not divisible_by(num, 2)
+
+def is_even(num):
+    return divisible_by(num, 2)
+
+def default(v, d):
+    return v if exists(v) else d
+
+def softclamp(t, value = 5.):
+    return (t / value).tanh() * value
+
+# rotary, but with attenuation of short relative distance frequencies
+
+class RotaryEmbedding(Module):
+    def __init__(
+        self,
+        dim,
+        max_positions = 8192
+    ):
+        super().__init__()
+        num_freqs = dim // 2
+        inv_freq = 1. / (arange(num_freqs).float() + torch.logspace(1, max_positions - num_freqs + 1, num_freqs))
+        self.register_buffer('inv_freq', inv_freq)
+
+    def forward(
+        self,
+        seq_len
+    ):
+        device = self.inv_freq.device
+        t = arange(seq_len, device = device).type_as(self.inv_freq)
+        freqs = einsum(t, self.inv_freq, 'i , j -> i j')
+        return cat((freqs, freqs), dim = -1)
+
+def rotate_half(x):
+    x1, x2 = x.chunk(2, dim = -1)
+    return torch.cat((-x2, x1), dim = -1)
+
+def apply_rotary_pos_emb(pos, t):
+    return t * pos.cos() + rotate_half(t) * pos.sin()
+
+# prenorm and sandwich norm - they use sandwich norm for single rep, prenorm for pairwise rep
+
+class NormWrapper(Module):
+    def __init__(
+        self,
+        dim,
+        block: Module,
+        dropout = 0.,
+        sandwich = False
+    ):
+        super().__init__()
+        self.block = block
+        self.pre_rmsnorm = nn.RMSNorm(dim) # they use an interesting variant of batchnorm, batch-rmsnorm. craft later and make sure it works distributed
+
+        self.post_block_dropout = nn.Dropout(dropout)
+        self.post_rmsnorm = nn.RMSNorm(dim) if sandwich else nn.Identity()
+
+    def forward(
+        self,
+        x,
+        **kwargs
+    ):
+        x = self.pre_rmsnorm(x)
+        out = self.block(x, **kwargs)
+        out = self.post_block_dropout(out)
+        return self.post_rmsnorm(out)
+
+# attention
+
+class Attention(Module):
+    def __init__(
+        self,
+        dim,
+        dim_head = 64,
+        heads = 8,
+        dim_head_qk = 128,
+        dim_head_v = 192,
+        dim_pairwise = None,
+        softclamp_value = 5. # they employ attention softclamping
+    ):
+        super().__init__()
+        dim_pairwise = default(dim_pairwise, dim)
+
+        self.scale = dim_head ** -0.5
+
+        qkv_proj_dim_out = (dim_head_qk * heads, dim_head_qk, dim_head_v)
+
+        # splitting and merging of attention heads
+
+        self.split_q_heads = Rearrange('b n (h d) -> b h n d', h = heads)
+        self.merge_heads = Rearrange('b h n d -> b n (h d)')
+
+        # projections
+
+        self.to_qkv = LinearNoBias(dim, sum(qkv_proj_dim_out))
+        self.to_out = LinearNoBias(dim_head_v * heads, dim)
+
+        # they add layernorms to queries, keys, and interestingly enough, values as well. first time i've seen this
+
+        self.q_norm = nn.LayerNorm(dim_head_qk, bias = False)
+        self.k_norm = nn.LayerNorm(dim_head_qk, bias = False)
+        self.v_norm = nn.LayerNorm(dim_head_v, bias = False)
+
+        # to attention bias
+
+        self.to_attn_bias = Sequential(
+            nn.RMSNorm(dim_pairwise), # replace with BatchRMSNorm once crafted
+            nn.GELU(),
+            LinearNoBias(dim_pairwise, heads),
+            Rearrange('b i j h -> b h i j')
+        )
+        # variables
+
+        self.qkv_dim_splits = qkv_proj_dim_out
+        self.softclamp_value = softclamp_value
+
+    def forward(
+        self,
+        x,
+        pairwise = None, # Float['b i j dp']
+        rotary_emb = None
+    ):
+
+        q, k, v = self.to_qkv(x).split(self.qkv_dim_splits, dim = -1)
+
+        # they use multi-query attention, with only 1 key / value head - pretty unconventional, but maybe enough for genomic modeling
+
+        q = self.split_q_heads(q)
+
+        q, k, v = self.q_norm(q), self.k_norm(k), self.v_norm(v)
+
+        q = q * self.scale
+
+        # maybe rotary
+
+        if exists(rotary_emb):
+            q, k = tuple(apply_rotary_pos_emb(rotary_emb, t) for t in (q, k))
+
+        # similarities
+
+        sim = einsum(q, k, 'b h i d, b j d -> b h i j')
+
+        # add attention bias + softclamping
+
+        if exists(pairwise):
+            attn_bias = self.to_attn_bias(pairwise)
+
+            assert divisible_by(sim.shape[-1], attn_bias.shape[-1])
+            expand_factor = sim.shape[-1] // attn_bias.shape[-1]
+
+            attn_bias = repeat(attn_bias, 'b h i j -> b h (i r1) (j r2)', r1 = expand_factor, r2 = expand_factor)
+
+            sim = softclamp(sim + attn_bias, value = self.softclamp_value)
+
+        # attention
+
+        attn = sim.softmax(dim = -1)
+
+        # aggregate
+
+        out = einsum(attn, v, 'b h i j, b j d -> b h i d')
+
+        out = self.merge_heads(out)
+        return self.to_out(out)
+
+# single to pairwise
+
+class SingleToPairwise(Module):
+    def __init__(
+        self,
+        dim,
+        pool_size = 16,
+        dim_pairwise = 128,
+        heads = 32
+    ):
+        super().__init__()
+        self.avg_pool = Reduce('b (n pool) d -> b n d', 'mean', pool = pool_size)
+
+        dim_inner = heads * dim_pairwise
+
+        self.split_heads = Rearrange('b n (h d) -> b n h d', h = heads)
+
+        self.to_outer_sum = Sequential(
+            LinearNoBias(dim, dim_pairwise * 2),
+            nn.GELU()
+        )
+
+        self.to_qk = LinearNoBias(dim, dim_inner * 2)
+        self.qk_to_pairwise = Linear(heads, dim_pairwise)
+
+    def forward(self, single):
+
+        single = self.avg_pool(single)
+
+        q, k = self.to_qk(single).chunk(2, dim = -1)
+        q, k = tuple(self.split_heads(t) for t in (q, k))
+
+        sim = einsum(q, k, 'b i h d, b j h d -> b i j h')
+
+        pairwise_from_sim = self.qk_to_pairwise(sim)
+
+        outer_q, outer_k = self.to_outer_sum(single).chunk(2, dim = -1)
+
+        outer_sum = einx.add('b i d, b j d -> b i j d', outer_q, outer_k)
+
+        return outer_sum
+
+# pairwise attention is a single headed attention across rows, they said columns did not help
+
+class PairwiseRowAttention(Module):
+    def __init__(
+        self,
+        dim
+    ):
+        super().__init__()
+        self.scale = dim ** -0.5
+
+        self.to_qk = LinearNoBias(dim, dim * 2)
+        self.to_v = Linear(dim, dim)
+
+    def forward(
+        self,
+        x
+    ):
+
+        q, k = self.to_qk(x).chunk(2, dim = -1)
+        v = self.to_v(x)
+
+        # similarity
+
+        sim = einsum(q, k, 'b n i d, b n j d -> b n i j')
+
+        # attention
+
+        attn = sim.softmax(dim = -1)
+
+        # aggregate
+
+        return einsum(attn, v, 'b n i j, b n j d -> b n i d')
+
+# feedforward for both single and pairwise
+
+def FeedForward(
+    dim,
+    *,
+    dropout = 0.,
+    expansion_factor = 2.,  # they only do expansion factor of 2, no glu
+):
+    dim_inner = int(dim * expansion_factor)
+
+    return Sequential(
+        Linear(dim, dim_inner),
+        nn.ReLU(),
+        nn.Dropout(dropout),
+        Linear(dim_inner, dim)
+    )
+
+# transformer
+
+class TransformerTower(Module):
+    def __init__(
+        self,
+        dim,
+        *,
+        depth = 8,
+        heads = 8,
+        dim_head_qk = 128,
+        dim_head_v = 192,
+        dropout = 0.,
+        ff_expansion_factor = 2.,
+        max_positions = 8192,
+        dim_pairwise = None,
+        pairwise_every_num_single_blocks = 2,   # how often to do a pairwise block
+        single_to_pairwise_heads = 32,          # they did 32
+        attn_kwargs: dict = dict(),
+        ff_kwargs: dict = dict()
+    ):
+        super().__init__()
+        dim_pairwise = default(dim_pairwise, dim)
+
+        layers = []
+
+        self.pairwise_every = pairwise_every_num_single_blocks
+
+        self.rotary_emb = RotaryEmbedding(dim_head_qk, max_positions = max_positions)
+
+        for layer_index in range(depth):
+
+            attn = Attention(dim = dim, dim_head_qk = dim_head_qk, dim_head_v = dim_head_v, heads = heads, dim_pairwise = dim_pairwise)
+
+            ff = FeedForward(dim = dim, expansion_factor = ff_expansion_factor)
+
+            attn = NormWrapper(dim = dim, block = attn, dropout = dropout, sandwich = True)
+            ff = NormWrapper(dim = dim, block = ff, dropout = dropout, sandwich = True)
+
+            # maybe pairwise
+
+            single_to_pairwise, pairwise_attn, pairwise_ff = None, None, None
+
+            if divisible_by(layer_index, self.pairwise_every):
+                single_to_pairwise = SingleToPairwise(dim = dim, dim_pairwise = dim_pairwise, heads = single_to_pairwise_heads)
+                pairwise_attn = PairwiseRowAttention(dim_pairwise)
+                pairwise_ff = FeedForward(dim = dim_pairwise, expansion_factor = ff_expansion_factor)
+
+                single_to_pairwise = NormWrapper(dim = dim, block = single_to_pairwise, dropout = dropout)
+                pairwise_attn = NormWrapper(dim = dim_pairwise, block = pairwise_attn, dropout = dropout)
+                pairwise_ff = NormWrapper(dim = dim_pairwise, block = pairwise_ff, dropout = dropout)
+
+            # add to layers
+
+            layers.append(ModuleList([
+                attn,
+                ff,
+                single_to_pairwise,
+                pairwise_attn,
+                pairwise_ff
+            ]))
+
+
+        self.layers = ModuleList(layers)
+
+    def forward(
+        self,
+        single
+    ):
+
+        seq_len = single.shape[1]
+
+        pairwise = None
+
+        rotary_emb = self.rotary_emb(seq_len)
+
+        for (
+            attn,
+            ff,
+            maybe_single_to_pair,
+            maybe_pairwise_attn,
+            maybe_pairwise_ff
+        ) in self.layers:
+
+            single = attn(single, rotary_emb = rotary_emb, pairwise = None) + single
+            single = ff(single) + single
+
+            if exists(maybe_single_to_pair):
+                pairwise = maybe_single_to_pair(single) + default(pairwise, 0.)
+                pairwise = maybe_pairwise_attn(pairwise) + pairwise
+                pairwise = maybe_pairwise_ff(pairwise) + pairwise
+
+        return single, pairwise
+
+# embedding
+
+class DNAEmbed(Module):
+    def __init__(
+        self,
+        dim,
+        dim_input = 5, # 5 basepairs
+        width = 15
+    ):
+        super().__init__()
+        assert is_odd(width)
+        self.dim_input = dim_input
+        self.conv = nn.Conv1d(dim_input, dim, width, padding = width // 2)
+        self.pointwise = nn.Conv1d(dim, dim, 1)
+
+    def forward(
+        self,
+        seq # Int['b n']
+    ):
+        onehot = F.one_hot(seq, num_classes = self.dim_input).float()
+        x = rearrange(onehot, 'b n d -> b d n')
+
+        out = self.conv(x)
+        out = out + self.pointwise(out)
+        return rearrange(out, 'b d n -> b n d')
+
+# classes
+
+class AlphaGenome(Module):
+    def __init__(
+        self,
+        dim = 768,
+        basepairs = 5,
+        dna_embed_width = 15,
+        dim_pairwise = None,
+        transformer_kwargs: dict = dict()
+    ):
+        super().__init__()
+        assert is_odd(dna_embed_width)
+
+        self.to_dna_embed = DNAEmbed(dim, dim_input = basepairs, width = dna_embed_width)
+
+        self.transformer = Transformer(
+            dim = dim,
+            dim_pairwise = dim_pairwise,
+            **transformer_kwargs
+        )
+
+    def forward(
+        self,
+        seq,
+        pairwise
+    ):
+
+        dna_embed = self.to_dna_embed(seq)
+
+        attended = self.transformer(dna_embed)
+
+        return attended

alphagenome_pytorch-0.0.1.dist-info/METADATA

@@ -0,0 +1,78 @@
+Metadata-Version: 2.4
+Name: alphagenome-pytorch
+Version: 0.0.1
+Summary: AlphaGenome
+Project-URL: Homepage, https://pypi.org/project/alphagenome-pytorch/
+Project-URL: Repository, https://github.com/lucidrains/alphagenome
+Author-email: Phil Wang <lucidrains@gmail.com>
+License: MIT License
+        
+        Copyright (c) 2025 Phil Wang
+        
+        Permission is hereby granted, free of charge, to any person obtaining a copy
+        of this software and associated documentation files (the "Software"), to deal
+        in the Software without restriction, including without limitation the rights
+        to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+        copies of the Software, and to permit persons to whom the Software is
+        furnished to do so, subject to the following conditions:
+        
+        The above copyright notice and this permission notice shall be included in all
+        copies or substantial portions of the Software.
+        
+        THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+        IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+        FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+        AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+        LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+        OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+        SOFTWARE.
+License-File: LICENSE
+Keywords: artificial intelligence,attention mechanism,deep learning,genomics,splicing,transformers
+Classifier: Development Status :: 4 - Beta
+Classifier: Intended Audience :: Developers
+Classifier: License :: OSI Approved :: MIT License
+Classifier: Programming Language :: Python :: 3.9
+Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
+Requires-Python: >=3.9
+Requires-Dist: einops>=0.8.0
+Requires-Dist: einx>=0.3.0
+Requires-Dist: torch>=2.4
+Provides-Extra: examples
+Provides-Extra: test
+Requires-Dist: pytest; extra == 'test'
+Description-Content-Type: text/markdown
+
+<img src="./extended-figure-1.png" width="450px"></img>
+
+## AlphaGenome (wip)
+
+Implementation of [AlphaGenome](https://deepmind.google/discover/blog/alphagenome-ai-for-better-understanding-the-genome/), Deepmind's updated genomic attention model
+
+## Install
+
+```bash
+$ pip install alphagenome-pytorch
+```
+
+## Usage
+
+```python
+import torch
+from alphagenome import TransformerTower
+
+transformer = TransformerTower(dim = 768, dim_pairwise = 128)
+
+single = torch.randn(2, 512, 768)
+
+attended_single, attended_pairwise = transformer(single)
+```
+
+## Citations
+
+```bibtex
+@article{avsec2025alphagenome,
+  title   = {AlphaGenome: advancing regulatory variant effect prediction with a unified DNA sequence model},
+  author  = {Avsec, {\v{Z}}iga and Latysheva, Natasha and Cheng, Jun and Novati, Guido and Taylor, Kyle R and Ward, Tom and Bycroft, Clare and Nicolaisen, Lauren and Arvaniti, Eirini and Pan, Joshua and Thomas, Raina and Dutordoir, Vincent and Perino, Matteo and De, Soham and Karollus, Alexander and Gayoso, Adam and Sargeant, Toby and Mottram, Anne and Wong, Lai Hong and Drot{\'a}r, Pavol and Kosiorek, Adam and Senior, Andrew and Tanburn, Richard and Applebaum, Taylor and Basu, Souradeep and Hassabis, Demis and Kohli, Pushmeet},
+  year    = {2025}
+}
+```

alphagenome_pytorch-0.0.1.dist-info/RECORD

@@ -0,0 +1,6 @@
+alphagenome/__init__.py,sha256=FjaT3_la9IG9w-PJ0Tk7ZK550O4Zq8SM5jej3nlzE6U,93
+alphagenome/alphagenome.py,sha256=xIsgMa23nANZyrIlsY4Og7fpibpDED9XVxkeQQXZmYg,11817
+alphagenome_pytorch-0.0.1.dist-info/METADATA,sha256=19TlrcXyHn1-YbLK8V_VZ3nPima2NRBflOFveDsO9Bg,3378
+alphagenome_pytorch-0.0.1.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
+alphagenome_pytorch-0.0.1.dist-info/licenses/LICENSE,sha256=1yCiA9b5nhslTavxPjsQAO-wpOnwJR9-l8LTVi7GJuk,1066
+alphagenome_pytorch-0.0.1.dist-info/RECORD,,

alphagenome_pytorch-0.0.1.dist-info/WHEEL

@@ -0,0 +1,4 @@
+Wheel-Version: 1.0
+Generator: hatchling 1.27.0
+Root-Is-Purelib: true
+Tag: py3-none-any

alphagenome_pytorch-0.0.1.dist-info/licenses/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2025 Phil Wang
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.